Springer Tracts in Electrical and Electronics Engineering

Debanjan Bhowmik

Spintronics-Based
Neuromorphic
Computing
Springer Tracts in Electrical and Electronics
Engineering

Series Editors
Brajesh Kumar Kaushik, Department of Electronics and Communication
Engineering, Indian Institute of Technology Roorkee, Roorkee, Uttarakhand, India
Mohan Lal Kolhe, Faculty of Engineering and Sciences, University of Agder,
Kristiansand, Norway
Springer Tracts in Electrical and Electronics Engineering (STEEE) publishes the
latest developments in Electrical and Electronics Engineering - quickly, informally
and with high quality. The intent is to cover all the main branches of electrical and
electronics engineering, both theoretical and applied, including:
• Signal, Speech and Image Processing
• Speech and Audio Processing
• Image Processing
• Human-Machine Interfaces
• Digital and Analog Signal Processing
• Microwaves, RF Engineering and Optical Communications
• Electronics and Microelectronics, Instrumentation
• Electronic Circuits and Systems
• Embedded Systems
• Electronics Design and Verification
• Cyber-Physical Systems
• Electrical Power Engineering
• Power Electronics
• Photovoltaics
• Energy Grids and Networks
• Electrical Machines
• Control, Robotics, Automation
• Robotic Engineering
• Mechatronics
• Control and Systems Theory
• Automation
• Communications Engineering, Networks
• Wireless and Mobile Communication
• Internet of Things
• Computer Networks
Within the scope of the series are monographs, professional books or graduate text-
books, edited volumes as well as outstanding PhD theses and books purposely
devoted to support education in electrical and electronics engineering at graduate
and post-graduate levels.
Review Process
The proposal for each volume is reviewed by the main editor and/or the advisory
board. The books of this series are reviewed in a single blind peer review process.
Ethics Statement for this series can be found in the Springer standard guidelines
here https://www.springer.com/us/authors-editors/journal-author/journal-author-hel
pdesk/before-you-start/before-you-start/1330#c14214.
Debanjan Bhowmik

Spintronics-Based
Neuromorphic Computing
Debanjan Bhowmik
Department of Electrical Engineering
Indian Institute of Technology Bombay
Mumbai, Maharashtra, India

ISSN 2731-4200 ISSN 2731-4219 (electronic)

Springer Tracts in Electrical and Electronics Engineering
ISBN 978-981-97-4444-2 ISBN 978-981-97-4445-9 (eBook)
https://doi.org/10.1007/978-981-97-4445-9

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Singapore Pte Ltd. 2024

This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore

If disposing of this product, please recycle the paper.

Preface

It won’t be an exaggeration to refer to the current age as the age of artificial intelli-
gence (AI), with AI tools and applications, including search engines, large language
models, virtual assistants, generative AI, and the like, entering and even some-
times controlling different aspects of our lives currently, and applications like self-
driving cars and humanoid robots on the verge of being an everyday reality. Most of
these AI applications are typically implemented on conventional computing hard-
ware like CPUs and GPUs. Neuromorphic computing takes inspiration from not
just the functioning but the underlying hardware of the original source of intelli-
gence, the human brain, and aims to build much more energy-efficient hardware
compared to conventional AI hardware with the final applications being robotics,
edge healthcare, etc.
Along with innovative algorithms, neuromorphic computing relies on different
kinds of emerging devices and circuits for efficient implementation. Spintronic
devices are considered interesting for neuromorphic computing because of various
properties they offer: non-volatility, electrical control and read-out of the non-volatile
states, self-sustaining oscillations, stochastic switching, etc. While books have been
published in the past on various other kinds of neuromorphic devices (like resistive
random access memory (RRAM), phase change memory (PCM), etc.), to the best of
our knowledge, this is the first book that focusses on spintronics-based neuromor-
phic computing. But though our book mainly focusses on spintronic-devices-based
implementation of neuromorphic computing, the discussion on algorithms, circuits,
and even devices presented in the book can be useful for research on other kinds of
neuromorphic devices and circuits as well.
Almost for the past one and a half decade now, I have been carrying out research
in the field of nanomagnetism and spintronics, with the last 7 years completely
dedicated to spintronics-based neuromorphic computing. This book is a product of
that research. As a result, my group’s research results (both simulation-based and
experimental) get a lot of focus in this book. Nonetheless, I have also laid down a
general framework here, which is relevant to any researcher working on spintronics-
based neuromorphic computing or neuromorphic computing in general, and I have
explained my group’s research results in the context of that framework only. Also in

v
vi Preface

that context, I have discussed the recent results from various other research groups
across the world working in this area. I hope the readers find all these discussions
helpful.
Given that this is a heavily interdisciplinary topic lying at the intersection of
physics, materials science, electrical engineering, and computer science, I have tried
to provide sufficient introduction (and also numerical exercises) for each funda-
mental topic in the book (neural network algorithms, neuromorphic crossbar array
architectures, nanomagnetism, and spintronics) to make the book accessible to a
wider audience. Anybody with decent understanding of high-school- and college-
level mathematics and physics should be able to follow the ideas in the book. Thus,
this book is suitable for advanced undergraduate students, graduate and doctoral
students, postdoctoral researchers, and faculty members interested in this area.
I present below a chapter-by-chapter layout of the book:

Part I: Motivation

Chapter 1 (Why Spintronics-Based Neuromorphic Computing?): In this chapter, the

advantages of neuromorphic computing, along with its potential applications, have
been discussed first. Next, several reasons have been stated, with short explanations
and mentions of recent experimental and simulation-based demonstrations, to justify
that spintronic devices are indeed highly relevant for neuromorphic computing.

Part II: Background Material

Chapter 2 (Introduction to Nanomagnetism and Spintronics): Most of the spin-

tronic devices proposed for neuromorphic computing are based on the heavy-metal/
ferromagnetic-metal hetero-structure introduced in Chap. 1. In Chap. 2, we first
discuss the physics of this hetero-structure at a qualitative level. Then we intro-
duce the reader to the single-domain modelling and micromagnetic frameworks,
essential for modelling this kind of nanomagnetic and spintronic devices. Within
the single-domain framework, we discuss the physics of magnetic-field-driven and
spin-orbit-torque-driven magnetic switching of nanomagnets in this chapter, and we
show relevant numerical results. We postpone the discussion of spin-orbit-torque-
driven domain-wall motion (within the micromagnetic framework) and spin-orbit-
torque-driven magnetic oscillation to later chapters, where we also discuss their
neuromorphic applications.
Chapter 3 (Introduction to Artificial Neural Networks (ANN) and Spiking Neural
Networks (SNN)): This chapter deals with various algorithms related to neuromor-
phic computing. We first introduce the reader to various popular non-spiking neural
networks, or ANNs (including how the networks are trained): fully connected neural
network/multi-layer perceptron (MLP), convolutional neural network (CNN), etc. In
Preface vii

the second part of Chap. 3, we specifically discuss the spiking neural network (SNN),
which is considered the third generation of neural networks. And in that context, we
discuss various biologically possible models of neurons and synapses in the brain
and connect them through SNN algorithms.

Part III: Neuromorphic Computing Using Spintronic

Devices

Chapter 4 (The Ferromagnetic Domain-Wall Synapse Device): In Chap. 4 of the

book, we discuss the physics of spin-orbit-torque-driven domain-wall motion in
heavy-metal-ferromagnet-based hetero-structures, connecting it with the micromag-
netic framework discussed in Chap. 2 before. We discuss some recently reported
results based on modelling and micromagnetic simulation, as well as fabrication and
experimental characterization, of such domain-wall-based synapse devices. Other
alternative devices for similar purpose, e.g. skyrmionic devices and ferrimagnetic
domain-wall devices, are also discussed briefly.
Chapter 5 (Design of Artificial Neural Networks (ANN) with Domain-Wall
Synapse Devices): In Chap. 5 of the book, we show the design and circuit simu-
lation of crossbar-array-based ANNs using domain-wall synapse devices discussed
in Chap. 4. We discuss results both for inference and on-chip learning using these
crossbar arrays.
Chapter 6 (Design of Spiking Neural Networks (SNN) with Domain-Wall
Devices): In Chap. 6 of the book, we show the design and circuit simulation of
biologically inspired SNNs with domain-wall devices, acting both as synapses and
neurons.
Chapter 7 (Spintronic Oscillators, Their Synchronization Properties, and Appli-
cations in Oscillatory Neural Networks (ONNs): In heavy-metal-ferromagnet-based
hetero-structures, spin-orbit torque at the interface of the heavy-metal layer and the
ferromagnet layer not only can cause magnetic switching and domain-wall motion
but also can trigger sustained oscillation of the magnetic moments of the ferro-
magnetic layer around a particular axis. Such spintronic oscillators, based on the
nanopillar or nano-constriction geometry, are known as Spin Hall Nano-Oscillators
(SHNOs). They can synchronize with each other as well as to external RF magnetic
fields. In Chap. 7 of the book, we discuss recent simulation and experimental studies
on such spin oscillators and their synchronization properties. Finally, in the chapter,
we discuss data-classification schemes, proposed and experimentally implemented
with spin oscillators using their synchronization properties, with the target applica-
tion being neuromorphic computing. Such neuromorphic implementations are also
known as oscillatory neural networks (ONNs).

Mumbai, India Debanjan Bhowmik

February 2024
Acknowledgments

As mentioned in the Preface, and as evident from the references throughout the
book, recent simulation-based and experimental results from my own research group
feature heavily here. I am incredibly thankful to the talented, motivated, and hard-
working students who worked under my supervision in the last 7 years, first while
I was a faculty member at Indian Institute of Technology Delhi and then a faculty
member at Indian Institute of Technology Bombay. Without their research efforts, it
would not have been possible for me to write a book entirely dedicated to spintronics-
based neuromorphic computing.
So in that context, I would first like to thank the doctoral students who have
completed or are about to complete their doctoral theses on this topic under my
supervision: Divya Kaushik, Upasana Sahu, Ram Singh Yadav, and Neha Garg. Many
sections of this book overlap in content with their doctoral dissertations and the papers
I have published with them. I would also like to thank the undergraduate students
who worked under my supervision and co-authored several papers with my doctoral
students: Utkarsh Saxena, Utkarsh Singh, Janak Sharda, Varun Desai, Kushaagra
Goyal, Sri Vasudha Hemadri Bhotla, Tanmay Aggarwal, Aniket Sadashiva, Amod
Holla, Sanyam Singhal, Mudit Bansal, Anand Verma, Apoorv Dankar, Saurabh
Kumar, and Aadit Pandey. I am also thankful to my collaborator, Prof. Pranaba
Kishor Muduli, at Indian Institute of Technology Delhi, for co-supervising some of
these students with me.
I would also like to thank Prof. Sayeef Salahuddin, who was the supervisor for my
doctoral studies conducted at University of California Berkeley, USA, and Prof. Long
You, who was a postdoctoral researcher at the same institute and in the same research
group at the time of my doctoral study, for teaching me the basics of theoretical
and experimental spintronics. I am also indebted to Prof. Kaushik Roy of Purdue
University, USA; his seminal papers on spintronics-based neuromorphic computing,
which I have cited several times in this book, have helped me connect my device-level
spintronics work from doctoral study with system-level neuromorphic applications
when I started my own independent research career as a faculty member. Since
then, I have actively worked on the intersection of spin physics and neuromorphic
computing, which has finally resulted in this book.

ix
x Acknowledgments

At a personal level, I would like to thank my wife Poulomi for providing me with
constant support throughout the time I prepared the manuscript. I would also like to
thank my own parents and my parents-in-law for their support.
Contents

Part I Motivation
1 Why Spintronics-Based Neuromorphic Computing? . . . . . . . . . . . . . . . 3
1.1 Introduction: Why Neuromorphic Computing? . . . . . . . . . . . . . . . . . . 3
1.2 Why Spintronics-Based Neuromorphic Computing? . . . . . . . . . . . . . 6
1.2.1 Non-volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.2 Electrical Access and Control . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.3 Near-Analog Weight Storage (High Bit Resolution) . . . . . . . 9
1.2.4 Integration Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2.5 Auto-oscillation and Synchronization Properties . . . . . . . . . . 12
1.2.6 Stochastic Switching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

Part II Background Material

2 Introduction to Nanomagnetism and Spintronics . . . . . . . . . . . . . . . . . . 23
2.1 Relevant Materials and Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 Single-Domain Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3 Micromagnetic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4 Magnetic-Field-Driven Switching: Landau–Lifschitz–Gilbert
(LLG) Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.5 Spin-Orbit-Torque-Driven Switching:
Landau–Lifschitz–Gilbert–Slonczewski (LLGS)
Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

xi
xii Contents

3 Introduction to Artificial Neural Networks (ANN) and Spiking

Neural Networks (SNN) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.1 Introduction to Supervised Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 ANN: Fully Connected Neural Network (FCNN) Without
a Hidden Layer/Single Layer Perceptron (SLP) . . . . . . . . . . . . . . . . . 46
3.3 ANN: Fully Connected Neural Network (FCNN) with Hidden
Layers/Multi-layer Perceptron (MLP) . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.4 ANN: Convolutional Neural Network (CNN) . . . . . . . . . . . . . . . . . . . 50
3.5 Implementation of ANNs Through Crossbar Arrays
of Non-volatile Memory (NVM) Synapses . . . . . . . . . . . . . . . . . . . . . 51
3.6 Spiking Neural Network (SNN) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.6.1 Leaky Integrate Fire (LIF) Neurons . . . . . . . . . . . . . . . . . . . . . 54
3.6.2 Spike Time-Dependent Plasticity (STDP) of Synapses . . . . . 55
3.6.3 SNN Composed of LIF Neurons and STDP-Following
Synapses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

Part III Neuromorphic Computing Using Spintronic Devices

4 The Ferromagnetic Domain-Wall Synapse Device . . . . . . . . . . . . . . . . . 63
4.1 Working Principle of a Domain-Wall Synapse Device . . . . . . . . . . . . 63
4.2 Micromagnetic Simulation of the Domain-Wall Synapse
Device . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.3 Experimental Demonstrations of the Domain-Wall Synapse
Device . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.4 Micromagnetic Simulation to Explain Experimentally
Obtained Non-linearity of LTP and LTD . . . . . . . . . . . . . . . . . . . . . . . 72
4.5 Improving Linearity and Symmetry of LTP and LTD . . . . . . . . . . . . 75
4.6 Alternative Spintronic Synapse Devices . . . . . . . . . . . . . . . . . . . . . . . . 76
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5 Design of Artificial Neural Networks (ANN) with Domain-Wall
Synapse Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.1 Thresholding-Based Modification to the Neural Network
Training Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.2 Design of the Synapse Cell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.3 Design of Crossbar Arrays to Implement the Back-Propagation
Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Contents xiii

6 Design of Spiking Neural Networks (SNN) with Domain-Wall

Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.1 Design and Working of the STDP-Enabling Circuit . . . . . . . . . . . . . . 93
6.2 Design and Working of the Domain-Wall-Device-Based LIF
Neuron . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.3 Performance of SNN with Domain-Wall-Device-Based
Neurons and Synapses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7 Spintronic Oscillators, Their Synchronization Properties,
and Applications in Oscillatory Neural Networks (ONNs) . . . . . . . . . . 107
7.1 Auto-oscillations in SHNOs of Nano-pillar Geometry . . . . . . . . . . . . 107
7.2 Auto-oscillations in SHNOs of Nano-constriction Geometry . . . . . . 111
7.3 Synchronization of Dipole-Coupled Nano-pillar SHNOs . . . . . . . . . 111
7.4 Synchronization of Spin-Wave-Coupled Nano-constriction
SHNOs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.5 Neuromorphic Computing Using Synchronized Spintronic
Oscillators: Oscillatory Neural Networks (ONNs) . . . . . . . . . . . . . . . 119
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
About the Author

Debanjan Bhowmik is an Associate Professor in the Department of Electrical Engi-

neering at the Indian Institute of Technology Bombay. Earlier, he was an Assistant
Professor in the Indian Institute of Technology Delhi from 2017 to 2021, and an Assis-
tant Professor in the Indian Institute of Technology Bombay from 2022 to 2023. He
completed his Ph.D. from the Department of Electrical Engineering and Computer
Sciences, University of California, Berkeley, CA, USA, in 2015. He completed his
B.Tech., Department of Electrical Engineering from the Indian Institute of Tech-
nology, Kharagpur, India, in 2010. His research areas are specialized energy-efficient
hardware for artificial intelligence (neuromorphic computing), nanomagnetism and
spintronics, computational neuroscience, and quantum computing for artificial intel-
ligence/machine learning. He has published several research papers in international
journals of repute in these areas.

xv
 Part I
Motivation
Chapter 1
Why Spintronics-Based Neuromorphic
Computing?

1.1 Introduction: Why Neuromorphic Computing?

Neural network algorithms are routinely and widely used now by the Artificial
Intelligence (AI) and Machine Learning (ML) community for various tasks like
data classification, recognition of objects in images, transcribing speech into text,
voice assistance, showing news items and product advertisements on the user’s feed
based on the user’s interests, providing answers to the questions asked through
large language models, etc. All neural network algorithms, from simple multi-layer
perceptrons (MLP) to state-of-the-art convolutional neural networks (CNN) and
long short-term memories (LSTM), involve the use of a large number of weight
parameters which need to be continuously used and updated in the computing unit
of the computer and subsequently stored in the memory unit (LeCun et al. 2015). In
a traditional computer which follows the von Neumann architecture, e.g., the central
processing unit (CPU), the memory and computing units are physically separate,
with a large part of the memory unit off-chip (on a different chip compared to the
computing/processing unit) (Patterson and Hennessey 2017; Sarangi 2017). So, a
large amount of time and energy is spent in shuffling data like the weight parameters
between the computing and off-chip memory units while running neural network
algorithms. This is known as the von Neumann bottleneck in literature (Wulf and
McKee 1996). Though the graphics processing unit (GPU) has a lot more computing
cores than the CPU and has way more parallelism and more frequent memory access,
the von Neumann bottleneck still exists for the GPU since the cores are physically
separated from the memory (Zidan et al. 2018).
The matrix-vector multiplication (MVM) operation and outer product calculation
(during training) are operations fundamental in all neural network algorithms which
use frequent interaction between the memory and computing units and are subject to
the von Neumann bottleneck. In order to overcome this bottleneck, a new comput-
ing paradigm has been proposed and implemented through crossbar arrays of analog
memory devices known as synaptic devices (inspired by synapses in the brain) where

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 3
D. Bhowmik, Spintronics-Based Neuromorphic Computing, Springer Tracts in Electrical
and Electronics Engineering, https://doi.org/10.1007/978-981-97-4445-9_1
4 1 Why Spintronics-Based Neuromorphic Computing?

Fig. 1.1 Crossbar-array architecture: Schematic of the crossbar-array architecture, popularly used
in neuromorphic computing, is shown here. The memory elements (synapses) are intertwined here
with the computation elements: adding of currents flowing through the synapses connected to
a particular vertical bar of the crossbar array, and operation of a non-linear activation element, or
neuron, on it. a shows the vector-matrix multiplication (VMM) operation, needed both for inference
and on-chip learning, and the outer product calculation needed only for on-chip learning

memory and computing are essentially intertwined (Sadashiva et al. 2024; Tsai et al.
2018; Sebastian et al. 2018; Chakroborty et al. 2020; Burr et al. 2015). Such an ana-
log crossbar array compute at the site of memory itself using a grid of non-volatile
synaptic devices and can implement the MVM operation and outer product calcu-
lation operations much faster than traditional computing units (Chakroborty et al.
2020; Burr et al. 2015; Luo and Yu 2020), as shown in Fig. 1.1a and b, respectively.
In order to implement simple neural networks like MLP, such crossbar arrays
along with peripheral analog circuit for the non-linear activation function and weight-
update calculation has been shown to be sufficient (through circuit-level simulation)
(Bhowmik et al. 2019; Kaushik et al. 2020a, b). So, in the case of MLP imple-
mentation, the full system is analog. But for CNN implementation, given the many
complex operations involved, all the MVM and outer product calculation opera-
tions can be implemented in analog crossbar arrays, and all other operations can
be implemented in traditional digital computing units, with analog-to-digital (ADC)
and digital-to-analog (DAC) converters converting between the analog and digital
units (Chakroborty et al. 2020; Luo and Yu 2020; Ankit et al. 2019, 2020).
The crossbar array is loosely inspired by the brain because, in the brain, neurons
are connected to each other through synapses. Signal propagating through one neuron
gets modulated by the synapse and reaches the next-stage neuron. Signals propagating
from other neurons and similarly modulated by synapses get added to it. Similarly,
in a crossbar array, as shown in Fig. 1.1a, electrical voltage applied at any horizontal
bar results in a current flow in the vertical bar, where the current is the product of the
voltage and the conductance of the synaptic device connecting the horizontal bar with
1.1 Introduction: Why Neuromorphic Computing? 5

the vertical bar. The conductance is proportional to the weight stored in the synapse
(an element in the weight matrix of the MVM operation to be executed by the crossbar
array). Such currents from all the horizontal bars in the crossbar array add up at each
vertical bar. Thus, using Ohm’s law and Kirchoff’s current law, the MVM operation is
completed. A peripheral analog circuit element can be designed to implement a non-
linear activation function like “tanh” or “Relu” on the signal at the vertical bar. Thus,
this crossbar array mimics the flow of signals inside the brain to some degree and
hence can be called “neuromorphic” (Chakroborty et al. 2020; Bhowmik et al. 2019).
However, there are some other designs within this non-von Neumann or memory-
computing-intertwining paradigm which are even closer to the working of the brain
and the term “neuromorphic” better suits them. Much like the brain, these designs
use spikes (in the time domain) for computation and information transfer (Diehl and
Cook 2015; Bi and Poo 1998; Zhang and Linden 2003). So, they are called spiking
neural networks (SNN) (the neural networks discussed previously, which do not use
spikes, are often called artificial neural networks (ANN) in neuromorphic literature
Christensen et al. 2022). It has been shown that spike-based computation consumes
much less power than computation without spikes (Christensen et al. 2022; Maass
1996, 1997, 2015; Davies 2019; Roy et al. 2019). Several spike-based digital chips
and systems like True North, SpiNNAker, and Loihi have also been designed and
shown to be competitive with respect to traditional computers for various AI/ML tasks
(Merolla et al. 2014; Furber et al. 2014; Davies et al. 2018). Analog architectures
which use various emerging non-volatile devices and mimic spike-based computing
in the brain have also been proposed and experimentally demonstrated (Bouvier et al.
2019; Thakur et al. 2018; Indiveri et al. 2011).
In addition, there are also various oscillator-based computing schemes and archi-
tectures which are inspired by evidence of oscillatory behaviour of neurons in the
brain (Fell and Axmacher 2011; Mizrahi et al. 2018a; Mirollo and Strogatz 1990).
Computing schemes have been proposed which use synchronization among oscilla-
tors and carry out data classification/machine learning tasks, while sometimes need-
ing less number of adjustable parameters compared to standard neural networks
(Grollier et al. 2016, 2020). A computing scheme, known as oscillator-based Ising
computing, has also been proposed where sub-harmonic injection locking is applied
on the oscillators, leading to their phases acquiring binary values. Subsequently,
these injection-locked oscillators can be used to solve combinatorial optimization
problems like Max-Cut, Maximum Independent Set, and Traveling Salesman prob-
lems, all very relevant for machine learning (Mohseni et al. 2022; Wang et al. 2021;
Houshang et al. 2022).
Given that most of the aforementioned neuromorphic systems focus on imple-
menting AI/ML tasks while consuming a low amount of energy and time, they are
considered attractive for edge-computing applications (a resource-constrained envi-
ronment) (Nwakanma et al. 2021; Nicholas et al. 2018). In this Internet of Things
(IoT) era, neuromorphic chips can be embedded in the edge devices and can process
information at the edge itself. For example, autonomous drones can include simple
neural network circuits which actively learn and adapt to subtle variations in their
environment. In a wireless sensor network, neural network circuits associated with
6 1 Why Spintronics-Based Neuromorphic Computing?

the sensor nodes can do some basic signal processing on the received information
and then send the most useful information to the hub.
Two approaches can be followed while integrating edge devices with neuromor-
phic technology: inference only and on-chip learning. In the former approach, the
neural network implemented on a neuromorphic chip is pre-trained, i.e., the weights
stored in the synaptic devices are already of desired values corresponding to that of a
pre-trained neural network. The pre-training is carried out on a traditional computer
which follows the von Neumann architecture. The pre-trained neuromorphic chip,
deployed at an edge device, only carries out signal processing/data classification on
the information received at the edge device (testing/inference). In the latter approach,
both the training and inference are carried out on a neuromorphic chip at the edge
device itself.
The inference-only approach is easier to implement because the training process,
which is the harder part of the computation here, doesn’t need to be implemented on
the novel architectures followed in neuromorphic chips. But at the same time, it has
several disadvantages. A neural network for edge devices makes use of information
collected at the edge; sending such information to a central computer for train-
ing involves data privacy concerns. Besides, the central computer follows the von
Neumann architecture, and training a large neural network on such a computer can
consume a lot of time and energy for the aforementioned reasons (Luo and Yu 2020;
Saxena et al. 2018). Also, training a neural network with multiple layers involves
frequent use of VMM and outer product calculation operations, following the back-
propagation algorithm (LeCun et al. 2015). Since crossbar arrays used in a neuromor-
phic architecture not only enable fast VMM but also fast outer product calculation
(Chakroborty et al. 2020), if training/learning is not carried out at the edge, the capa-
bilities of neuromorphic architecture are not fully utilized. All these factors make
on-chip learning in neuromorphic systems highly attractive for practical applications.
This book focuses on how spintronics, an emerging device technology that uses
the physical phenomenon of manipulation of magnetic moments through electri-
cal transport in nanomagnetic systems for information storage and processing, can
be used for neuromorphic computing. We discuss spintronics-based neuromorphic
computing both in the context of inference and on-chip learning in this book while
laying specific emphasis on on-chip learning when discussing spintronic-synapse-
based crossbar arrays. In the next part of this chapter, we provide an overview of
this subject by briefly discussing what salient properties of spintronic devices make
them attractive for neuromorphic computing.

1.2 Why Spintronics-Based Neuromorphic Computing?

In this sub-section, we outline the various advantages spintronics offers that make
different spintronic devices attractive for neuromorphic computing (including Ising
computing). We discuss the operating physics and design of these spintronic devices,
and how neuromorphic circuits are designed with them, in detail in the subsequent
chapters.
1.2 Why Spintronics-Based Neuromorphic Computing? 7

1.2.1 Non-volatility

As mentioned earlier, crossbar arrays implementing neural networks use synaptic

devices to store the elements of the weight matrix in the MVM operation in the
neural network. For inference, only the final weight values need to be stored for
a long time duration; for on-chip learning, intermediate weight values during the
training process need to be stored for shorter duration (Tsai et al. 2018; Sebastian
et al. 2018; Chakroborty et al. 2020; Burr et al. 2015).
Transistor-based SRAM cells have been proposed for synapses in neural network
circuits, but around six transistors are needed per synapse cell leading to large area
consumption. Also, the transistor, being a charge-based semiconductor device, is
volatile, and consumes power continuously for weight storage (Sengupta and Roy
2016; Sengupta et al. 2016). Hence, non-volatility is an important property that a
synaptic device in a crossbar array must have. Spintronic devices use the magnetic
moments in the ferromagnetic (or ferrimagnetic) layer for information storage and
manipulation. In the absence of an external magnetic field or current flow (of sig-
nificant magnitude), a ferromagnetic layer retains its magnetic configuration. This
makes spintronic devices non-volatile. Because of their non-volatility, spintronic
devices are used in hard disks for data storage and in main memory and cache mem-
ory of computers (magnetic random access memory (MRAM), magnetic racetrack
devices) (Fullerton and Childress 2016; Apalkov et al. 2016; Lee and Lee 2016).
This same property of non-volatility is the primary motivation for using spintronic
devices as synaptic elements in a crossbar array (Chakroborty et al. 2020; Sengupta
and Roy 2016; Sengupta et al. 2016).

1.2.2 Electrical Access and Control

In both magnetic hard drives and MRAM devices, the value stored in each memory
unit through the magnetic configuration of the ferromagnetic layer is accessed elec-
trically using the tunneling magneto-resistance (TMR) effect. Because of this effect,
the conductance of a magnetic tunnel junction (MTJ), as shown in Fig. 1.2, is higher
when the magnetic moments of the two ferromagnetic layers (fixed layer, . M2 and
the free layer, . M1 ), separated by the insulating layer (through which electrons tun-
nel), are aligned parallel to each other (.G P ). The conductance is lower when the two
moments . M1 and . M2 are anti-parallel to each other (.G A P ). The ratio of .G P to .G A P
is known as the TMR ratio; higher TMR ratio makes it easier for electrical access of
the information stored in the device (Apalkov et al. 2016; Lee and Lee 2016).
In the traditional MRAM device, information stored in the ferromagnetic layer is
manipulated using an external magnetic field, generated through the Oersted effect
by a current-carrying line close to the device (Apalkov et al. 2016). However, in
the spin-transfer-torque MRAM device (STTMRAM), which is a modification to
the MRAM device (shown in the schematic of Fig. 1.2a), electrical current flowing
through the “fixed” layer in the MTJ (the ferromagnetic layer where it is more difficult
8 1 Why Spintronics-Based Neuromorphic Computing?

Fig. 1.2 STTMRAM and SOTMRAM: a Schematic of the spin-transfer-torque MRAM device
(STTMRAM) is shown here. The paths for the “read” current and the “write” current are the same
here: the vertical magnetic tunnel junction (MTJ) structure. b Schematic of the spin-orbit-torque
MRAM device (SOTMRAM) is shown here. While the path for the “read” current is still the vertical
MTJ structure here, the path for “write” current is horizontal, through the underlying heavy-metal
layer. Solid green arrows indicate magnetic moments for both the schematics: . M1 for the free layer
and . M2 for the fixed layer

to switch the magnetization due to higher thickness of the layer or its moment being
pinned to an anti-ferromagnetic layer) becomes spin-polarized. When that current
flows into the “free” layer (the magnetic moments inside this layer are indicative of
the information stored in the device), magnetic moment in the “free” layer . M1 may
get switched by the current due to spin-transfer torque (Apalkov et al. 2016; Lee and
Lee 2016). Thus information is not just accessed electrically (the “read” process)
but also manipulated electrically (the “write” process). The electrical control here
is much more energy-efficient than in a traditional MRAM devices because for a
given magnitude of current, the strength and efficiency (for magnetic switching) of
spin-transfer torque is much higher than that Oersted field generated by the current.
Based on our description of the working principle of an STTMRAM device, it is
clear that it is a two-terminal device with a shared “read” and “write” path, which
is the MTJ structure. Higher-magnitude “write” current pulses flowing through the
MTJ structure can damage the tunneling layer (typically an oxide layer) and hence
the “read” path. In order to physically separate the “read” and “write” paths, a three-
terminal spintronic device has been proposed and experimentally demonstrated for
memory applications. This is the spin-orbit-torque MRAM device (SOTMRAM)
device, shown in the schematic of Fig. 1.2b (Lee and Lee 2016; Liu et al. 2012a, b;
Miron et al. 2010). Here the “read” process is based on the TMR effect in MTJ,
just like in STTMRAM, but the “write” process involves an in-plane current flow-
ing through the heavy-metal layer underneath the ferromagnetic-metal layer. Spin-
orbit torque generated at the heavy metal-ferromagnet interface causes the magnetic
moment in the ferromagnetic free layer (. M1 ) to switch. Thus spintronic devices
like STTMRAM and SOTMRAM offer non-volatility as well as energy-efficient
electrical read-out and control.
If the aforementioned crossbar array is used only for inference, then the conduc-
tances of the synaptic devices need to be set to appropriate values only a few times.
1.2 Why Spintronics-Based Neuromorphic Computing? 9

But forward computation (MVM operation) needs to be carried out on the crossbar
array often; so accessing the weight stored in the synaptic device frequently (with-
out the need for changing it) through its conductance is imperative for the crossbar
array. Hence, the TMR effect makes any MTJ-based spintronic device attractive for
synaptic application in a crossbar array for inference purpose (Apalkov et al. 2016;
Lee and Lee 2016).
For on-chip learning, along with frequently accessing the weight stored in the
synapse, the weight value needs to be updated frequently. Phenomena like spin-
transfer torque and spin-orbit torque enable manipulation of magnetic configuration
and hence updating of weight values through current pulses in an energy-efficient
way. This makes spintronic devices attractive as synapses in a crossbar array even
for on-chip learning.

1.2.3 Near-Analog Weight Storage (High Bit Resolution)

Magnetic hard drive, MRAM, STTMRAM, and SOTMRAM are available as com-
mercial products and can be used for conventional digital computing with von Neu-
mann architecture (Fullerton and Childress 2016; Apalkov et al. 2016; Lee and Lee
2016). Hence, information is stored in them as bits: in each device, all the magnetic
moments in the ferromagnetic layer are aligned parallel to each other and can be
thought of as a single, effective magnetization vector (shown as a green arrow in
Fig. 1.2a and b: . M1 and . M2 ). This magnetization vector pointing in one direction
corresponds to bit 0, while it pointing to the opposite direction corresponds to bit
1. Spin-transfer torque or spin-orbit torque from a “write” current pulse is used to
switch the magnetization vector of the “free” layer from one direction to the other
(shown in Fig. 1.2a and b), thereby flipping the bit from 0 to 1 or 1 to 0.
But crossbar arrays that implement neural networks are essentially analog in
nature, with each synapse storing any weight value between a minimum possible
value (.wmin ) and a maximum possible value (.wmax ) in a near-analog fashion (preci-
sion of 5-10 bits) (Chakroborty et al. 2020; Aabrar et al. 2019). So an STTMRAM
or an SOTMRAM cell which stores one bit of information is not suitable for this
purpose, unless a binary neural network is implemented, or several such cells are
used together as a single synapse (Jung et al. 2022; Zhang et al. 2022).
A spin-orbit-torque-driven domain-wall synapse, shown in Fig. 1.3, serves the
purpose of a near-analog/multi-bit synapse really well (Chakroborty et al. 2020;
Sengupta and Roy 2016; Sengupta et al. 2016). Unlike the STTMRAM or the SOTM-
RAM device, all the magnetic moments in the “free” layer are not oriented parallel to
each other here. Instead, in the ferromagnetic “free” layer of the domain-wall synapse,
a domain wall separates a region with all moments pointed vertically upwards from
a region with all moments pointed vertically downward (shown by white arrows in
Fig. 1.3). Given that all the magnetic moments in the “fixed” layer are always pinned
to a specific direction (say upward), a domain wall moving from the left end of the
device to the right end increases the conductance of the MTJ (the “read” path) from
10 1 Why Spintronics-Based Neuromorphic Computing?

Fig. 1.3 Domain-wall Synapse: Schematic of the heavy-metal/ferromagnet hetero-structure-based

domain-wall synapse is shown here. “Write” current flowing through the heavy-metal layer moves
the domain wall (DW) in the ferromagnetic layer via spin-orbit torque. The conductance of the
vertical MTJ structure (path for the “read” current) depends on the position of the domain wall and
determines the weight stored in this domain-wall synapse. White arrows correspond to the magnetic
moments of the ferromagnetic layers here: the lower green layer in which the domain wall (DW)
is present is the “free” layer and the upper green layer is the “fixed” layer of the MTJ structure.
(Reprinted with permission from: Kaushik D, Singh U, Sahu U, Sreedevi I, Bhowmik D (2020)
Comparing domain-wall synapse with other non-volatile memory devices for on-chip learning in
analog hardware neural network. AIP Adv 10(2):025111 Kaushik et al. 2020a)

Fig. 1.4 Conductance Response of a Domain-wall Synapse: The conductance response of the
heavy-metal/ferromagnet hetero-structure-based, spin-orbit-torque-driven domain-wall synapse (of
Fig. 1.3) is shown here, as obtained from micromagnetic simulation of the device. “Write” current
pulses of constant magnitude (25.µA here) flowing through the heavy-metal layer move the domain
wall in the ferromagnetic layer by fixed amounts. So the conductance goes up from .G A P , when
the domain wall is at the left end of the device, to .G P , when the domain wall is at the right end
of the device, in 50 discrete steps. Thus this simulated device exhibits 50 conductance states, and
so the corresponding weight stored has 5–6 bits of resolution, much higher than that stored in an
STTMRAM or an SOTMRAM cell (which stores 1 bit only). (Reprinted with permission from:
Kaushik D, Singh U, Sahu U, Sreedevi I, Bhowmik D (2020) Comparing domain-wall synapse with
other non-volatile memory devices for on-chip learning in analog hardware neural network. AIP
Adv 10(2):025111 Kaushik et al. 2020a)

G A P to .G P (as shown in Fig. 1.4, obtained from the micromagnetic simulation of

.
a domain-wall device shown in Fig. 1.3). The corresponding weight of the synapse
changes from .wmin to .wmax . Any intermediate position of the domain wall corre-
sponds to an intermediate weight value, thereby imparting a near-analog nature to
the device. Motion of the domain wall in the opposite direction results in a gradual
decrease of the conductance and the corresponding weight value, as shown in Fig. 1.4.
1.2 Why Spintronics-Based Neuromorphic Computing? 11

It has been demonstrated through experiments and micromagnetic simulations that

in the heavy metal/ferromagnetic metal heterostructure which exhibits perpendicular
magnetic anisotropy (PMA) and Dzyaloshinskii Moriya Interaction (DMI), a domain
wall of Neel chirality is formed. When in-plane current flows through the heavy-metal
layer (Pt or Ta), spin-orbit torque at the heavy metal-ferromagnet interface moves the
domain wall in a particular direction even in the absence of an external magnetic field.
If the direction of the current flow is reversed, the direction of domain-wall motion
reverses (Emori et al. 2013; Ryu et al. 2013; Bhowmik et al. 2015). This phenomenon
is utilized in designing the domain-wall synapse for crossbar arrays, with the synapse
storing nearly analog values of weights (8–16 bit precision) (Chakroborty et al.
2020; Sengupta and Roy 2016; Sengupta et al. 2016). The bit precision depends
logarithmically on the number of conductance states possible in the device, which is
given by the ratio of the length of the device to the distance moved by the domain wall
due to a single pulse of the smallest possible magnitude which overcomes the pinning
of a domain wall due to defects (Kaushik et al. 2020a, b). We discuss this spintronic
device in details in Chaps. 4 and 5 and benchmark it with respect to other synaptic
devices, like resistive random-access memory (RRAM) and phase change memory
(PCM) devices, more commonly used for neuromorphic computing and are not based
on spintronics (Tsai et al. 2018; Sebastian et al. 2018; Chakroborty et al. 2020).
In Chap. 4, we also discuss an alternative spintronic device which exhibits similar
current-controlled weight-update characteristic: the skyrmionic device. Skyrmions
are smaller in size compared to domain walls and are known for their topological
stability. Also much like domain walls, skyrmions can be moved by current pulses,
leading to synaptic behaviour (Kang et al. 2016; Sampaio et al. 2013; Maccariello
et al. 2018; Bhattacharya et al. 2019).

1.2.4 Integration Property

In this chapter, thus far, we have highlighted what properties make spintronic devices
interesting for use as synapses in hardware implementations of ANNs. As mentioned
earlier, implementing another class of neural networks known as spiking neural net-
works (SNN) is also of great interest in neuromorphic computing. Here, we explain
briefly what aspect of the aforementioned spin-orbit-torque-driven domain-wall
device makes it interesting for application as a neuron device in an SNN.
In an SNN, neurons and synapses follow biologically plausible models as opposed
to purely mathematical models used in the conventional non-spiking neural net-
work algorithms (Diehl and Cook 2015; Bi and Poo 1998; Zhang and Linden 2003).
The integrate-and-fire (IF) and leaky-integrate-and-fire (LIF) models of the neuron
are popular for modelling spiking neural networks because of their computational
simplicity. The state of the neuron is determined by its potential. As shown in the
schematic for the LIF model of a neuron in Fig. 1.5, the time-derivative of the neuron
potential is proportional to the input current, which gets injected into it from the other
neurons after being modulated by the synapses. So the neuron potential is an integral
12 1 Why Spintronics-Based Neuromorphic Computing?

Fig. 1.5 Leaky-integrate and fire (LIF) model of a neuron: Schematic of the biologically plausible
LIF model of a neuron is shown here. The presence of the capacitive element in the model (.C)
shows that the neuron voltage, given by .V (t), is an integral of the input current . I (t) over time .t.
The resistive element (conductance: .G L ) corresponds to the leaky part of the model

of the input current over time. The plots obtained by solving the LIF model for dif-
ferent values of input current to the neuron show that when the potential reaches
a threshold value, the neuron fires and the potential goes back to its lowest value
(resting potential) (Dayan and Abbott 2005). The firing/spiking rate of the neuron is
proportional to the magnitude of the input current.
The physics of domain-wall motion can be utilized to implement this integral
property in a neuron device (Sengupta et al. 2016; Sengupta and Roy 2017; Hassan
et al. 2018; Akinola et al. 2019; Bennett et al. 2019; Yue et al. 2019). Following
the physics of in-plane-current-driven domain-wall motion discussed above, for a
large range of current magnitude, the domain-wall velocity is proportional to the
current (Sahu et al. 2019). The domain-wall position, being an integral of the velocity
over time, is equivalent to the potential of a neuron here (Sahu et al. 2019; Dayan
and Abbott 2005). As shown in the schematic for a domain-wall neuron device in
Fig. 1.6, once the domain wall reaches the other end of the device, i.e., the domain-
wall position exceeds a threshold value, the associated circuit is designed such that a
spike is generated and the domain wall goes back to its original position (equivalent
to the potential of a neuron going back to its resting potential) (Sahu et al. 2019). In
Chap. 6, we discuss in details the design of such a domain-wall based neuron device
and also show design of an SNN using domain-wall synapses and domain-wall
neurons.

1.2.5 Auto-oscillation and Synchronization Properties

Along with magnetic switching and domain-wall motion, spin transfer torque and
spin-orbit torque are known to trigger magnetic oscillation in MTJs (Chen et al. 2016).
1.2 Why Spintronics-Based Neuromorphic Computing? 13

Fig. 1.6 Domain-wall synapse acting as a LIF neuron: a Frequency of neuron spiking versus
magnitude of current pulse plot, obtained from solving the LIF model of Fig. 1.5, is consistent with
that of the frequency of spiking of a domain-wall-based neuron device as a function of magnitude
of the in-plane current pulse which moves the domain wall b Schematic of the domain-wall-based
neuron device is shown here (Sahu et al. 2019). Once the domain wall reaches the right end of
the device, because of the presence of the MTJ structure, the transistor circuit associated with it
generates a spike. This is equivalent to the neuron voltage reaching a threshold in the LIF model.
Then the circuit also drives a reverse current which drives the domain wall to its original position
at the left end of the device. This is equivalent to the neuron voltage going back to the resting
potential in the LIF model. (Reprinted with permission from: Sahu U, Pandey A, Goyal K, Bhowmik
D (2019) Spike time dependent plasticity (STDP) enabled learning in spiking neural networks using
domain-wall-based synapses and neurons. AIP Adv 9(12) Sahu et al. 2019)

Magnetic moments in such an MTJ are known to undergo sustained auto-oscillations

in the vortex mode or the uniform mode (based on the size of the MTJ and other
parameters) under the application of a DC current, as shown in Fig. 1.7. Such an
oscillator is known to synchronize to an external RF magnetic field of frequency
close to the oscillator’s natural frequency (Romera et al. 2018; Singh et al. 2021).
Similarly, multiple spin oscillators with natural-frequency values close to each other
have been shown to synchronize with each other, using dipole coupling or spin-wave
coupling (Zeng et al. 2021; Flovik et al. 2016; Garg et al. 2021).
Oscillatory behaviour and synchronization behaviour have been observed in
the neurons of the brain (Fell and Axmacher 2011; Mizrahi et al. 2018a; Mirollo
and Strogatz 1990). Inspired by that, various data-classification schemes have been
proposed utilizing the synchronization behaviour of oscillators for neuromorphic
computing (Vodenicarevic et al. 2018). Since spin oscillators also exhibit both auto-
oscillation and synchronization behaviour, various inference and learning tasks for
neuromorphic computing have been demonstrated through simulations and experi-
ments on an array of spin oscillators (Grollier et al. 2016, 2020; Romera et al. 2018;
14 1 Why Spintronics-Based Neuromorphic Computing?

Fig. 1.7 Schematic of spin Hall nano-oscillator (SHNO): In a heavy-metal/ferromagnet-based

hetero-structure, in-plane current flowing through the heavy-metal layer can induce auto-oscillation
in the ferromagnetic layer. Magnetic moments of the ferromagnetic layer precess around the vertical
axis, along which the external DC magnetic field (. Happ ) is applied.(Reprinted with permission from:
Singh U, Garg N, Kumar S, Muduli PK, Bhowmik D (2021) Learning of classification tasks with
tasks with an array of uniform-mode spin Hall nano-oscillators. AIP Adv 11:045117 Singh et al.
2021)

Singh et al. 2021; Zahedinejad et al. 2020; Garg et al. 2021). We discuss this in
details in Chap. 7.
The same model that describes the synchronization among oscillators and helps
develop neuromorphic computing schemes with them also enables them to behave
like an Ising machine. In the Ising computing scheme, all oscillators have roughly the
same natural frequency and they are all subjected to a perturbation of frequency twice
of that natural frequency. All the oscillators are found to synchronize, with the phase
of some being 0 and the rest being .π. This kind of locking to the external perturbation
is known as sub-harmonic injection locking. When such locking happens, if the initial
conditions of the system correspond to the inputs for a particular NP-complete prob-
lem (Max-Cut, Maximum Independent Set, Traveling Salesman etc.), then the final
conditions of the system often correspond to the ground state of the corresponding
Ising Hamiltonian and hence yield the correct solution to the problem (Mohseni et al.
2022; Wang et al. 2021). Given the NP-complete nature of the problem, reaching the
correct solution to the problem may take much longer on a conventional computer.
It has been shown recently experimentally that such sub-harmonic injection locking
can indeed be implemented on spin oscillators, leading to phase binarization, making
them potential candidates for Ising machines (Houshang et al. 2022).
References 15

1.2.6 Stochastic Switching

In an MTJ, if the energy barrier between the two magnetization states in the “free”
layer is low, the magnetization can randomly switch from one direction to the other
due to thermal fluctuations. When such an MTJ is used for memory technology like
STTMRAM or SOTMRAM, this is not a desired property. So, in order to prevent such
random switching events and maximize the data retention time, the MTJ is engineered
to have a high energy barrier in STTMRAM and SOTMRAM devices. This in turn
increases the energy consumption in the writing/information manipulation process
due to high-magnitude current pulses.
In probabilistic computing, however, such random switching events are utilized
for new functionalities. The energy barrier in a super-paramagnetic tunnel junction
is deliberately kept low and comparable to the thermal energy (Grollier et al. 2020).
So the configuration of the moments in the “fixed” and the “free” layer can fluctuate
randomly between parallel and anti-parallel configurations, with the conductance
fluctuating between that for the two configurations (.G P and .G A P ).
Such stochastic switching in a super-paramagnetic tunnel junction can be used to
model the Poisson spiking behaviour found among neurons in the brain and hence
makes such a tunnel junction attractive for neuromorphic computing (Grollier et al.
2020; Mizrahi et al. 2018a, b). Stochastic magnetic switching has also been used to
propose and implement probabilistic bits, or p-bits, and Ising computing has been
proposed and also experimentally demonstrated with p-bits (Grollier et al. 2020;
Camsari et al. 2017, 2019; Borders et al. 2019).

References

Aabrar KA et al (2019) BEOL compatible superlattice FerroFET-based high precision analog

weight cell with superior linearity and symmetry. In: Proceedings of international electron devices
meeting (IEDM), pp 19.6.1–19.6.4
Akinola O, Hu X, Bennett CH, Marinella M, Friedman JS, Incorvia JAC (2019) Three-terminal
magnetic tunnel junction synapse circuits showing spike-timing-dependent plasticity. J Phys D:
Appl Phys 52(49):49LT01
Ankit A, Hajj IE, Chalamalasetti SR, Aggarwal S, Marinella M, Foltin M, Strachan JP, Milojicic DS,
Roy K (2020) PANTHER: a programmable architecture for neural network training harnessing
energy-efficient ReRAM. IEEE Trans Comput 69(8):1128–1142
Ankit A, Hajj IE, Chalamalasetti SR, Ndu G, Foltin M, Williams RS, Faraboschi P, Hwu WW,
Strachan JP, Roy K, Milojicic DS (2019) PUMA: a programmable ultra-efficient memristor-
based accelerator for machine learning inference. In: ASPLOS ’19: proceedings of the twenty-
fourth international conference on architectural support for programming languages and operating
systems, pp 715–731
Apalkov D, Dieny B, Slaughter JM (2016) Magnetoresistive random access memory. Proc IEEE
104(10):1796–1830
16 1 Why Spintronics-Based Neuromorphic Computing?

Bennett CH, Hassan N, Hu X, Incornvia JAC, Friedman JS, Marinella MJ (2019) Semi-supervised
learning and inference in domain-wall magnetic tunnel junction (DW-MTJ) neural networks. In:
Spintronics XII, 2019 international society for optics and photonics, vol 11090, p 110903I
Bhattacharya T, Li S, Huang Y, Kang W, Zhao W, Suri M (2019) Low-power (1t1n) skyrmionic
synapses for spiking neuromorphic systems. IEEE Access 7:5034–5044
Bhowmik D, Nowakowski ME, You L, Lee O, Keating D, Wong M, Jeffrey B, Salahuddin S (2015)
Deterministic domain wall motion orthogonal to current flow due to spin orbit torque. Sci Rep
5(1):1–10
Bhowmik D, Saxena U, Dankar A, Verma A, Kaushik D, Chatterjee S, Singh U (2019) On-chip
learning for domain wall synapse based fully connected neural network. J Magn Magn Mater
489:165434
Bi G, Poo M (1998) Synaptic modifications in cultured hippocampal neurons: dependence on spike
timing, synaptic strength, and postsynaptic cell type. J Neurosci 18(24):10464–10472
Borders WA, Pervaiz AZ, Fukami S, Camsari KY, Ohno H, Datta S (2019) Integer factorization
using stochastic magnetic tunnel junctions. Nature 573:390–393
Bouvier M, Valentian A, Mesquida T, Rummens F, Reyboz M, Vianello E, Beigne E (2019) Spiking
neural networks hardware implementations and challenges: a survey. ACM J Emerg Technol
Comput Syst (JETC) 15(2):1–35
Burr GW et al (2015) Large-scale neural networks implemented with non-volatile memory as the
synaptic weight element: comparative performance analysis (accuracy, speed, and power). IEDM
Tech Dig 4.4.1–4.4.4
Camsari KY, Faria R, Sutton BM, Datta S (2017) Stochastic p-bits for invertible logic. Phys Rev X
7:031014
Camsari KY, Chowdhury S, Datta S (2019) Scalable emulation of signproblem- free Hamiltonians
with room-temperature p-bits. Phys Rev Appl 12:034061
Chakroborty I et al (2020) Resistive crossbars as approximate hardware building blocks for machine
learning: opportunities and challenges. Proc IEEE 1–35
Chen T, Dumas RK, Eklund A, Muduli PK, Houshang A, Awad AA, Dürrenfeld P, Malm BG, Rusu
A, Åkerman J (2016) Spin-torque and spin-Hall nano-oscillators. Proc IEEE 104(10):1919–1945
Christensen DV et al (2022) 2022 roadmap on neuromorphic computing and engineering.
Neuromorph Comput Eng 2:022501
Davies M et al (2018) Loihi: a neuromorphic manycore processor with on-chip learning. IEEE
Micro 82–89
Davies M (2019) Benchmarks for progress in neuromorphic computing. Nat Mach Intell 1(9):386–
388
Dayan P, Abbott LF (2005) Chapter 5. The MIT Press
Diehl PU, Cook M (2015) Unsupervised learning of digit recognition using spike-timing-dependent
plasticity. Front Comput Neurosci 9:99
Emori S, Bauer U, Ahn SM, Martinez E, Beach GSD (2013) Current-driven dynamics of chiral
ferromagnetic domain walls. Nat Mat 12(7):611–616
Fell J, Axmacher N (2011) The role of phase synchronization in memory processes. Nat Rev
Neurosci 12:105–118
Flovik V, Macia F, Wahlström E (2016) Describing synchronization and topological excitations in
arrays of magnetic spin torque oscillators through the Kuramoto model. Sci Rep 6:1–10
Fullerton ER, Childress JR (2016) Spintronics, magnetoresistive heads, and the emergence of the
digital world. Proc IEEE 104(10):1787–1795
Furber SB et al (2014) The SpiNNaker project. Proc IEEE 102(5):652–665
Garg N, Bhotla SVH, Muduli PK, Bhowmik D (2021) Kuramoto-model-based data classification
using the synchronization dynamics of uniform-mode spin Hall nano-oscillators. Neuromorphic
Comput Eng 1(2)
Grollier J, Querlioz D, Stiles MD (2016) Spintronic nanodevices for bioinspired computing. Proc
IEEE 104:2024–2039
References 17

Grollier J, Querlioz D, Camsari KY, Everschor-Sitte K, Fukami S, Stiles MD (2020) Neuromorphic

spintronics. Nat Electr 3(7):360–370
Hassan N, Hu X, Jiang-Wei L, Brigner WH, Akinola OG, Garcia-Sanchez F, Pasquale M, Bennett
CH, Incorvia JAC, Friedman JS (2018) Magnetic domain wall neuron with lateral inhibition. J
Appl Phys 124(15):152127
Houshang A et al (2022) Phase-binarized spin Hall nano-oscillator arrays: towards spin Hall Ising
machines. Phys Rev Appl 17:014003
Indiveri G, Linares-Baranco B́, Hamilton TJ, van Schaik A, Etienne-Cummings R et al (2011)
Neuromorphic silicon neuron circuits. Front Neurosci 5:73
Jung S et al (2022) A crossbar array of magnetoresistive memory devices for in-memory computing.
Nature 601:211–216
Kang W, Huang Y, Zhang X, Zhou Y, Zhao W (2016) Skyrmion-electronics: an overview and
outlook. Proc IEEE 104(10):2040–2061
Kaushik D, Singh U, Sahu U, Sreedevi I, Bhowmik D (2020a) Comparing domain wall synapse
with other non volatile memory devices for on-chip learning in analog hardware neural network.
AIP Adv 10(2):025111
Kaushik D, Sharda J, Bhowmik D (2020b) Synapse cell optimization and back-propagation algo-
rithm implementation in a domain wall synapse based crossbar neural network for scalable on-chip
learning. Nanotechnology 31(36):364004
LeCun Y, Bengio Y, Hinton G (2015) Deep learning. Nature 521:436–444
Lee S-W, Lee K-J (2016) Emerging three-terminal magnetic memory devices. Proc IEEE
104(10):1831–1843
Liu L, Pai CF, Li Y, Tseng HW, Ralph DC, Buhrman RA (2012a) Spin-torque switching with the
giant spin Hall effect of tantalum. Science 336(6081):555–558
Liu L, Lee OJ, Gudmundsen TJ, Ralph DC, Buhrman RA (2012b) Current-induced switching of
perpendicularly magnetized magnetic layers using spin torque from the spin Hall effect. Phys
Rev Lett 109(9):096602
Luo Y, Yu S (2020) Accelerating deep neural network in-situ training with non-volatile and volatile
memory based hybrid precision synapses. IEEE Trans Comput 69(8):1113–1127
Maass W (1997) Networks of spiking neurons: the third generation of neural network models.
Neural Netw 10(9)
Maass W (2015) To spike or not to spike: that is the question. Proc IEEE 103(12)
Maass W (1996) Lower bounds for the computational power of networks of spiking neurons. Neural
Comput 8:1–40
Maccariello D, Legrand W, Reyren N, Garcia K, Bouzehouane K, Collin S et al (2018) Electri-
cal detection of single magnetic skyrmions in metallic multilayers at room temperature. Nat
Nanotechnol 13:233–237
Merolla PA et al (2014) A million spiking-neuron integrated circuit with a scalable communication
network and interface. Science 345(6197):668–697
Mirollo RE, Strogatz SH (1990) Synchronization of pulse-coupled biological oscillators. SIAM J
Appl Math 50:1645–1662
Miron IM et al (2010) Current-driven spin torque induced by the Rashba effect in a ferromagnetic
metal layer. Nat Mater 9:230–234
Mizrahi A et al (2018a) Neural-like computing with populations of superparamagnetic basis
functions. Nat Commun 9:1533
Mizrahi A, Grollier J, Querlioz D, Stiles MD (2018b) Overcoming device unreliability with
continuous learning in a population coding based computing system. J Appl Phys 124:152111
Mohseni N, McMahon PL, Byrnes T (2022) Ising machines: hardware solvers for combinatorial
optimization problems. Nat Rev Phys 4
Nicholas GC, Skuda D, Schuman CD, Plank JS, Dean Garrett ME, Rose S (2018) Energy and area
efficiency in neuromorphic computing for resource constrained devices. In: Proceedings of ACM
great lakes symposium on VLSI (GLSVLSI)
18 1 Why Spintronics-Based Neuromorphic Computing?

Nwakanma CI, Kim J, Lee J, Kim DS (2021) Edge AI prospect using the NeuroEdge computing
system: introducing a novel neuromorphic technology. ICT Express 7:152–157
Patterson DA, Hennessey JL (2017) Computer architecture: a quantitative approach. Morgan
Kaufmann
Romera M, Talatchian P, Tsunegi S, Araujo FA, Cros V, Bortolotti P, Trastoy J, Yakushiji K,
Fukushima A, Kubota H, Yuasa S (2018) Vowel recognition with four coupled spin-torque nano-
oscillators. Nature 563:230–34
Roy K, Jaiswal A, Panda P (2019) Nature 475:607–617
Ryu KS, Thomas L, Yang SH, Parkin S (2013) Chiral spin torque at magnetic domain walls. Nat
Nanotechnol 8(7):527–533
Sadashiva A, Holla A, Bhowmik D (2024) Impact of non-idealities of synapse devices on on-
chip inference and learning performance of crossbar arrays, as studied through “Crossbar sim”
simulator. Under review
Sahu U, Pandey A, Goyal K, Bhowmik D (2019) Spike time dependent plasticity (STDP) enabled
learning in spiking neural networks using domain wall based synapses and neurons. AIP Adv
9(12)
Sampaio J, Cros V, Rohart S, Thiaville A, Fert A (2013) Nucleation, stability and current-induced
motion of isolated magnetic skyrmions in nanostructures. Nat Nanotechnol 8:839–844
Sarangi SR (2017) Computer organisation and architecture. Tata McGraw Hill
Saxena V et al (2018) Towards neuromorphic learning machines using emerging memory devices
with brain-like energy efficiency. J Low Power Electr Appl 8:34
Sebastian A et al (2018) Tutorial: brain-inspired computing using phase-change memory devices.
J Appl Phys 124:111101
Sengupta A, Banerjee A, Roy K (2016) Hybrid spintronic-CMOS spiking neural network with
on-chip learning: devices, circuits, and systems. Phys Rev Appl 6(6):064003
Sengupta A, Roy K (2016) A vision for all-spin neural networks: a device to system perspective.
IEEE Trans Circuits Syst- I 63(12)
Sengupta A, Roy K (2017) Appl Phys Rev 4:041105
Sengupta A, Shim Y, Roy K (2016) Proposal for an all-spin artificial neural network: emulating
neural and synaptic functionalities through domain wall motion in ferromagnets. IEEE Trans
Biomed Circuits Syst 10(6)
Singh U, Garg N, Kumar S, Muduli PK, Bhowmik D (2021) Learning of classification tasks with
tasks with an array of uniform-mode spin Hall nano-oscillators. AIP Adv 11:045117
Thakur CS, Molin JL, Cauwenberghs G, Indiveri G, Kumar K, Qiao N, Schemmel J et al (2018)
Large-scale neuromorphic spiking array processors: a quest to mimic the brain. Front Neurosci
12:891
Tsai H et al (2018) Recent progress in analog memory-based accelerators for deep learning. J Phys
D: Appl Phys 51:283001
Vodenicarevic D, Locatelli N, Grollier J, Querlioz D (2018) Nano-oscillator-based classification
with a machine learning-compatible architecture. J Appl Phys 124:152117
Wang T, Wu L, Nobel P, Roychowdhury J (2021) Solving combinatorial optimization problems
using oscillator based Ising machines. Nat Comput 20:287–306
Wulf W, McKee SA (1996) Hitting the memory wall: implicatons of the obvious. In: ACM
SIGARCH computer architecture
Yue K, Liu Y, Lake RK, Parker AC (2019) A brain-plausible neuromorphic on-the-fly learning
system implemented with magnetic domain wall analog memristors. Sci Adv 5(4):eaau8170
Zahedinejad M, Awad AA, Muralidhar S, Khymyn R, Fulara H, Mazraati H, Dvornik M, Åker-
man J (2020) Two-dimensional mutually synchronized spin Hall nano-oscillator arrays for
neuromorphic computing. Nat Nanotechnol 15:47–52
Zeng Z, Luo Z, Heyderman LJ, Kim JV, Hrabec A (2021) Synchronization of chiral vortex nano-
oscillators. Appl Phys Lett 118:222405
Zhang K et al (2022) High on/off ratio spintronic multi-level memory unit for deep neural network.
Adv Sci 2103357
References 19

Zhang W, Linden DJ (2003) The other side of the engram: experience-driven changes in neuronal
intrinsic excitability. Nat Rev Neurosci 4(11):885–900
Zidan MA, Strachan JP, Lu WD (2018) The future of electronics based on memristive systems. Nat
Electr 1:22–29
 Part II
Background Material
Chapter 2
Introduction to Nanomagnetism and
Spintronics

2.1 Relevant Materials and Devices

In the previous chapter, we have briefly discussed the working principle of the
spintronic devices essential for spin-based neuromorphic computing: Spin-Orbit
Torque Magnetic Random Access Memory (SOTMRAM), Spin Orbit-Torque-
Driven Domain-Wall device, and Spin Hall Nano-Oscillator (SHNO). As per the
schematics provided for these devices in that chapter, one can infer that they are all
based on the heavy-metal-ferromagnetic-metal-oxide hetero-structure. As a result,
most of these nanomagnetic and spintronic devices have the following properties in
common which are essential for their operation:

1. The ferromagnetic metal used is typically a 3d metal like Fe, Ni, or Co or an

alloy between 3d metals (CoFeB, permalloy, etc.). So the magnetism originates
here from the uncompensated spin angular momentum of the 3d electrons in each
atom and quantum mechanical exchange interaction between these atomic spins
(Coey 2010; Aharoni 2000; Blundell 2001).
2. The ferromagnetic-metal layer exhibits perpendicular magnetic anisotropy
(PMA) when its thickness is below .≈1.5 nm (Ikeda et al. 2010). The PMA is
considered to originate from the interface, though its exact microscopic origin
has been debated. The oxide layer above the ferromagnetic layer as well as the
heavy-metal layer below the ferromagnetic-metal layer has been considered to be
contributors to the PMA.
3. When in-plane current flows through this structure, the magnetic moments of the
ferromagnetic layer exhibit a torque known as spin-orbit torque (SOT), which
can lead to magnetic switching, domain-wall motion, magnetic auto-oscillations,
etc. The exact origin of SOT is again debatable. According to one theory, when
in-plane current flows through the device, the moving conduction electrons expe-
rience an effective magnetic field in their moving frame of reference due to spin-
orbit coupling (Fig. 2.1a), and due to s-d exchange coupling between these con-
duction electrons and the d-shell core electrons of the ferromagnetic layer, the

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 23
D. Bhowmik, Spintronics-Based Neuromorphic Computing, Springer Tracts in Electrical
and Electronics Engineering, https://doi.org/10.1007/978-981-97-4445-9_2
24 2 Introduction to Nanomagnetism and Spintronics

magnetic moments of the ferromagnetic layer experience a torque (Miron et al.

2010, 2011; Bhowmik et al. 2016a). This effect is called Rashba effect. According
to another theory, flow of in-plane current through the heavy-metal layer leads
to separation of electrons with opposite spin polarization inside the heavy metal
due to Spin Hall effect which leads to accumulation of spin-polarized electrons at
the ferromagnet-heavy-metal interface (Fig. 2.1b). These spin-polarized electrons
flowing into the ferromagnetic layer generate SOT on the magnetic moments of
the ferromagnetic layer (Liu et al. 2012a, b; Bhowmik et al. 2016a). But in case of
either effect, spin Hall effect, or Rashba effect, the ferromagnetic layer above the
heavy-metal layer experiences a torque, which can lead to magnetic switching,
domain-wall motion, and magnetic auto-oscillation (as we will see throughout
this book).

Fig. 2.1 Microscopic effects that contribute to spin-orbit torque (SOT) at a heavy-magnet-
ferromagnetic-metal interface are shown here: a Rashba effect: due to current (. j) flowing in-plane
in .−x direction, and due to symmetry breaking in the out-of-plane (.z) direction, the magnetic
moments inside the ferromagnetic layer experience a Rashba field. H R in the direction:.ẑ × →j b Spin
Hall effect: due to current (. j) flowing in-plane in .−x direction, electrons with spin polarization in
.−y direction (obtained from (.ẑ × → j), can also be .−(.ẑ × →j) depending on the choice of the heavy
metal) accumulate at the heavy-metal-ferromagnetic-metal interface, leading to the application of
SOT on the magnetic moments inside the ferromagnetic layer. (Reprinted with permission from:
Bhowmik D, Lee OJ, You L, Salahuddin S (2016) Magnetization switching and domain-wall motion
due to spin-orbit torque. In: Atulasimha J, Bandyopadhyay S (eds) Nanomagnetic and spintronic
devices for energy-efficient memory and computing. Wiley Bhowmik et al. 2016a)
2.2 Single-Domain Model 25

In order to innovate in terms of spintronic materials for non-volatile memory

(NVM) technology or neuromorphic computing, one needs to understand the micro-
scopic physics governing the aforementioned phenomena in detail, both through
theory and experiments. But if the material choice has been roughly made (after
verification through experiments), and innovations need to be carried out in terms of
device, circuit, and algorithms’ design, then phenomenological models (as opposed
to microscopic models) of the spintronic devices should be sufficient. This is the
approach we take in this book. We primarily use the Landau–Lifschitz–Gilbert–
Slonczewski (LLGS) equations to model the magnetization dynamics of the spin-
tronic devices, both at single-domain level and micromagnetic level. In such models,
the aforementioned phenomena like exchange, anisotropy, and spin-orbit torque (due
to spin Hall effect or Rashba effect) are captured through some simulation parameters,
as we discuss in this chapter.
We first discuss the single-domain and micromagnetic model from statics per-
spective (energy minimization), introduce the concept of a ferromagnetic domain
wall, and then discuss magnetization dynamics, both in the absence and presence
of spin-orbit torque. If the reader is interested in going beyond our phenomeno-
logical treatment and wants to delve into the microscopic physics, the readers are
recommended to go through textbooks (Coey 2010; Aharoni 2000; Blundell 2001;
Maekawa et al. 2017) and other references we have cited throughout this chapter.

2.2 Single-Domain Model

As mentioned earlier, the spontaneous magnetic moment in each atom of the

ferromagnetic-metal layer arises from the uncompensated spin angular momentum
in the d-shell of the atom. Let such spontaneous magnetic moment in each atom be
a . S→i , where i .=1, 2, 3, …. Then, according to the Heisenberg model, the energy of
the system is given by the Heisenberg Hamiltonian:

. H = −Σi, j Jex ( S→i . S→j ), (2.1)

where i and j correspond to neighbouring atoms (Coey 2010; Blundell 2001). The
motivation for this model as well as calculation of the exchange integral. Jex originates
from the microscopic picture of the system which is again outside the scope of this
book. Also, the band theory of ferromagnetism is a much more valid model for
ferromagnetic metals than Heisenberg model since the electrons are delocalized
in metals and the Heisenberg model is meant for localized electrons (Coey 2010;
Blundell 2001). Nevertheless, the Heisenberg model serves as a good toy model for
understanding the single-domain model and the micromagnetic model, and so we
start from the Heisenberg model here while discussing ferromagnetism in metallic
layers of our devices.
26 2 Introduction to Nanomagnetism and Spintronics

If the exchange-integral term in the Heisenberg Hamiltonian of Eq. 2.1 is con-

sidered infinite, then the exchange energy dominates over all other energy terms
and all the atomic magnetic moments . S→i are in the same direction, notwithstand-
→
ing what the direction is. So they can be treated as one giant macro-spin vector . M
with its norm conserved (saturation magnetization . Ms remains unchanged). This is
the single-domain model, also known as the Stoner-Wohlfarth model (Coey 2010;
Aharoni 2000; Blundell 2001; Arrott 2005).
It is valid under the following two conditions mainly:

1. The temperature of the system doesn’t change significantly, i.e., the saturation
magnetization of the magnetic layer doesn’t change. Most of the devices discussed
in this book operate at room temperature. Joule heating due to the current (which
causes switching, domain-wall motion, etc.) is also not high enough to cause a
significant change in the temperature.
2. The lateral dimensions of the device are low enough for dipole energy not to
dominate over exchange energy and break the nanomagnet into domains. Once
domains are formed, moments in different domains point in different directions,
and in the boundary region between two domains (a domain wall), the moment
gradually rotates from one direction to the other. So, for domain-wall device
modelling, which forms a significant component of the book (as explained in
Chap. 1), the single-domain model cannot be used; the micromagnetic model
needs to be used instead.

So, in the single-domain model, all moments point in one direction only, given by
→ But what the direction is depends on other energy terms,
the macro-spin vector . M.
which we mention next (all these energy terms need to be added and the net energy
→
needs to minimized to obtain steady-state value of . M):

1. Crystalline anisotropy: Electric fields due to the atoms in the crystal get trans-
formed to magnetic fields in the electrons’ reference frame due to spin-orbit
coupling. Since the atoms have specific arrangements in space, i.e., the crystal
is not isotropic in space, effective magnetic fields seen by the electrons, which
contribute to the magnetic moment, are only in certain directions. So the spins
of the electrons only want to align in specific directions. This is the origin of
crystalline anisotropy in simple terms. The corresponding energy term is given
as . K 1 Mx2 + K 2 M y2 + K 3 Mz2 , where . K 1 , . K 2 , and . K 3 are the anisotropy constants
in x, y, and z directions, and . M → = {Mx , M y , Mz } (we have normalized . M → here,
since the norm of macro-spin vector is always conserved in the single-domain
model). The magnitude and polarities of . K 1 , . K 2 , and . K 3 determine the preferred
direction of the magnetization due to crystalline anisotropy (Arrott 2005).
2. Interfacial anisotropy: In the case of thin films, often the symmetry breaking
of the crystal structure at the interface leads to an extra term in the energy. Let
that interface anisotropy constant be given by . K i , the bulk anisotropy be . K b , and
,
the thickness of the thin film be .th. Then the average anisotropy . K int is given
,
{ th
by . K int = th 0 (K b + K int δ(x))d x = K b + th . This term shows that for thin
1 K int
2.2 Single-Domain Model 27

films (.th value very low), interfacial anisotropy can dominate over bulk anisotropy.
The corresponding energy term is . K int Mz2 , with z being the out-of-plane direction
(along which the symmetry of the system is broken) (Ikeda et al. 2010).
3. Demagnetizing field: In a thin film, purely from a magneto-statics standpoint,
the energy of the system is much lower when the moments are all in the plane
compared to when moments are out of the plane. This is because in the latter case,
the dipole-coupling field from a much higher number of neighbouring moments
opposes a particular moment than in the former case. This leads to an additional
term .μ0 Ms2 Mz2 where . Ms is saturation magnetization of the ferromagnetic layer
and .μ0 is the magnetic permeability of free space.

For the ferromagnetic-metal-heavy-metal-hetero-structure we have considered in

this book, interfacial perpendicular magnetic anisotropy (PMA) dominates over the
other anisotropy terms, and the net anisotropy energy can be written as .−K Mz2 ,
where . K is the effective PMA constant and it takes a large positive value so that
. M z .= +/−1 (magnetization along the out-of-plane axis) minimizes the anisotropy
energy. Hence, in the absence of any magnetic field, the magnetization stays along.+z
or .−z, but where it points towards (at equilibrium) when external magnetic fields are
applied in different directions is what’s interesting and can be studied by minimizing
the net energy given as follows (based on this single-domain model):

. → H→
E = −K Mz2 − μ0 Ms M. (2.2)

where . K > 0, . H→ is applied externally, and . M

→ has been normalized like before.
If an external magnetic field . H (.μ0 H if the unit is tesla (T)) is applied in the z
direction, energy E can be written as

. E = −K Mz2 − μ0 Ms H Mz (2.3)

→ make an angle .θ with the z-axis. Then, . Mz = cos(θ), and

Let magnetization . M
net energy:
. E = −K cos (θ) − μ0 Ms H cos(θ)
2
(2.4)

dE
. = 0 ⇒ sinθ(2K cosθ + μ0 Ms H ) = 0 (2.5)
dθ
has three roots:
2
θ = 0 for which . ddθE2 > 0 (minima) when .μ0 H > − 2K
.
Ms
2
θ = π for which . ddθE2 > 0 (minima) when .μ0 H <
.
2K
Ms
θ = cos −1 (− μ02K
2
.
Ms H
) for which . ddθE2 > 0 (minima) when .|μ0 H | > | 2KMs
|, but .|cos(θ)|
cannot be greater than 1. So this root doesn’t yield a valid solution.
Hence the two possible solutions are .θ = 0 and .θ = π. When .μ0 H > 2K Ms
, .θ = 0,
i.e., . Mz = 1 (Fig. 2.2a). When .μ0 H < − Ms .θ = π, i.e., . Mz = −1 (Fig. 2.2b). But in
2K

the range .− 2K
Ms
< μ0 H < 2K Ms
, there are two solutions for .θ: 0 and .π (Fig. 2.2a). The
28 2 Introduction to Nanomagnetism and Spintronics

Fig. 2.2 a For a single-domain magnet/macro-spin exhibiting perpendicular magnetic anisotropy

(given by .−K Mz2 , . K > 0), roots for .θ (where . Mz = Ms cos(θ)) are shown for different values
of magnetic field applied in the out-of-plane (.z) direction (.μ0 Hz ). b Energy due to PMA shown
as a function of .θ. The two minima (.θ = 0, .π) correspond to the two equilibrium positions of
the magnet between which the magnet switches sharply, by crossing the energy barrier, upon the
application of sufficiently strong magnetic field in the out-of-plane (.z) direction (magnitude greater
than . 2K
Ms ). c The corresponding plot of out-of-plane component of magnetization (. Mz ) versus out-
of-plane magnetic field (.μ0 Hz ) shows this sharp switching at .μ0 Hz = 2K
Ms and .μ0 Hz = − 2K
Ms . The
out-of-plane direction is the easy axis for the magnet

solution that will physically occur depends on the history of the system, i.e., if the
magnetization . M → is in z direction initially due to applied field . H→ (.θ = 0) then it will
continue to be so until.mu 0 H→ is in negative direction and its magnitude is greater than
.
2K
Ms
. Once the magnetic field crosses .− 2K Ms
(the switching field), the magnetization
switches to .−z direction (.θ = π) (Fig. 2.2c). The same thing happens we start from
magnetization in .−z direction (.θ = π) and magnetic field is swept from .−z direction
to .+z direction: the magnetization switches to .+z direction once the magnetic field
crosses .− 2K
Ms
(Fig. 2.2c). The out-of-plane axis (z) is known as the easy axis of the
magnet.
This history-dependent switching behaviour in ferromagnets when the magnetic
field is applied along the easy axis results in retention of the magnetic state and
leads to its non-volatility. This hysteresis behaviour can also be explained intuitively
from the energy landscape of the magnetization, shown in Fig. 2.2b. When . Mz = 1
(.θ = 0 in Fig. 2.2b), the magnet is at an energy minimum, and unless a sufficiently
high magnetic field is applied in the .−z direction, it will stay along .+z. The same is
true when . Mz = 1, or .θ = π (Fig. 2.2b): unless a sufficiently high magnetic field is
applied in the .+z direction, the magnet will stay along .−z.
2.2 Single-Domain Model 29

Fig. 2.3 a For a single-domain magnet/macro-spin exhibiting perpendicular magnetic anisotropy

(given by .−K Mz2 , . K > 0), roots for .θ (where . Mz = Ms cos(θ)) are shown for different values
of magnetic field applied in the in-plane (.x) direction (.μ0 Hx ). b The corresponding plot of in-
plane component of magnetization (. Mx ) versus in-plane magnetic field (.μ0 Hx ) shows saturation at
.μ0 Hx =
Ms and .μ0 Hx = − Ms . The in-plane direction is the hard axis for the magnet
2K 2K

Let us now consider what happens when the magnetic field is applied in-plane
(say x direction). Energy of the system (E) is given by

. E = −K Mz2 − μ0 Ms H Mx (2.6)

→ make an angle .θ with the z-axis (out-of-plane axis). Then,

Let magnetization . M
Mz = Ms cos(θ), . Mx = sin(θ), and net energy:
.

. E = −K cos 2 (θ) − μ0 Ms H sin(θ) (2.7)

dE
. = 0 ⇒ cosθ(2K sinθ − μ0 Ms H ) = 0 (2.8)
dθ

Following the same procedure as before we get that when .μ0 H > 2K Ms
, .θ = π2 is the
π
solution (. Mx = 1) (Fig. 2.3a). When .μ0 H < − Ms , .θ = − 2 is the solution (. Mx =
2K

−1) (Fig. 2.3a). When .− 2KMs

< μ0 H < 2K Ms
, . Mx = sin(θ) = μ02K
Ms H
is the solution
(Fig. 2.3a). Sharp switching is absent here unlike in the easy-axis case, the field at
which the magnetization saturates (. 2K
Ms
) is known as the anisotropy field (. Hk ), and
the x-axis (and any other direction in the plane of the device) is the hard axis for the
magnet (Fig. 2.3b).
30 2 Introduction to Nanomagnetism and Spintronics

Ikeda et al. have reported a seminal study on PMA in heavy-metal-ferromagnetic-

metal-oxide-hetero-structure-based thin films which later paved the way for the use of
PMA in MRAM, STTMRAM, SOTMRAM, domain-wall devices, etc. In Ta (heavy
metal)/CoFeB (ferromagnetic metal)/MgO (oxide) thin films, they showed that when
the thickness of the CoFeB layer is 1.3 nm or below, it exhibits PMA, i.e., the out-
of-plane direction is the easy axis. Above 1.3 nm, magneto-static energy dominates
over interfacial PMA and the out-of-plane direction is the hard axis.
For the thin film with CoFeB layer of 1.3 nm thickness, along the hard axis
(any in-plane direction), the saturation field (. Hk = 2K
Ms
) is as high as 0.4 T, which
shows the presence of very strong PMA. However, along the easy axis (out-of-plane
direction), the switching field is much lower than 0.4 T (. 2K Ms
) (Ikeda et al. 2010),
which is inconsistent with the prediction from the single-domain model above, where
switching field along easy axis = saturation field along hard axis .= . 2K
Ms
(Fig. 2.2c, b).
This is because in thin films, magnetization switching happens through the nucleation
of reverse domains and subsequent domain-wall motion (Coey 2010), which the
single-domain model cannot capture (since it assumes all moments are aligned along
the same direction, which isn’t the case when domains are formed). But when devices
are fabricated from these PMA-exhibiting thin films, and the device dimensions are
reduced, the switching field increases and the experimentally obtained results get
more aligned with the predictions from the single-domain model. In the next sub-
section, we discuss how to model domains and domain walls in ferromagnetic layers
with the help of micromagnetics.

2.3 Micromagnetic Model

Through micromagnetics, we develop a formalism to model a large number of mag-

netic moments interacting with each other through dipole coupling and exchange
(without assuming that all the moments are in the same direction). We account for
all possible energy terms: exchange, dipole interaction, anisotropy, energy due to
external magnetic field (Zeeman energy), etc.
Going back to the Heisenberg model (Eq. 2.1) and normalizing the spin of each
individual atom by the saturation magnetization, we obtain reduced magnetization
vectors .m→i and .m→ j corresponding to two atoms i and j, separated by a displacement
vector .ri,→ j . .(m→i .m→ j ) .= .cos(φi, j ), which can be approximated as .(1 − φi,2 j ), given
.φi, j is very small (atoms .i and . j are physically very close to each other). Thus, the
Hamiltonian and the corresponding energy now depend on .φi,2 j . Now,

|φi, j | ≈ |m→i − m→ j |
. (2.9)

Next, a very important assumption is made which forms the very core of the
micromagnetics formalism. Instead of considering that the magnetization arises out
of individual atoms (which is the actual physical case), the magnetization is assumed
2.3 Micromagnetic Model 31

in micromagnetics to be a continuous field (Aharoni 2000; Blundell 2001). Hence

m→i − m→ j can be approximated as
.

→i
.m− m→ j = (m i,x x̂ + m i,y ŷ + m i,z ẑ) − (m j,x x̂ + m j,y ŷ + m j,z ẑ)
( ) ( )
∂m x ∂m x ∂m x ∂m y ∂m y ∂m y
≈ Δx + Δy + Δz x̂ + Δx + Δy + Δz ŷ
∂x ∂y ∂z ∂x ∂y ∂z
( )
∂m z ∂m z ∂m z
+ Δx + Δy + Δz ẑ
∂x ∂y ∂z
( )( )
∂ ∂ ∂
= Δx + Δy + Δz m x x̂ + m y ŷ + m z ẑ
∂x ∂y ∂z
( ( ))
∂ ∂ ∂ → m
= (Δx x̂ + Δy ŷ + Δz ẑ). x̂ + ŷ + ẑ (m x x̂ + m y ŷ + m z ẑ) = (ri,→ j .∇) →
∂x ∂y ∂z
(2.10)

The exchange energy of the system can be written as (Blundell 2001)

. E exchange = Jex S 2 Σi, j |ri,→ j .∇ m|

→ 2=
{ (( ) ( )
J S2 Z ∂m x ∂m x ∂m x 2 ∂m y ∂m y ∂m y 2
Δx + Δy + Δz + Δx + Δy + Δz +
a3 V ∂x ∂y ∂z ∂x ∂y ∂z
( ) )
∂m z ∂m z ∂m z 2
Δx + Δy + Δz dV
∂x ∂y ∂z
{ (( ) ( ) (
Jex S 2 Z ∂m x ∂m x ∂m x 2 ∂m y ∂m y ∂m y 2 ∂m z ∂m z
= + + + + + +v + +
a V ∂x ∂y ∂z ∂x ∂y ∂z ∂x ∂y
) ) {
∂m z 2
d V = A ((∇m x )2 + (∇m y )2 + (∇m z )2 )d V (2.11)
∂z V

where .Δx = Δy = Δz = a (the distance between nearest neighbouring atoms), Z

is the nearest number of neighbouring atoms for a given atom, and the exchange-
2
correlation constant . A = Jex aS Z . The value of . A can be calculated from first prin-
ciples (microscopic picture) or obtained experimentally through the value of . J ,
and then it’s used in this phenomenological micromagnetic model. For the heavy-
metal-ferromagnetic-metal-hetero-structure-based devices we explore in this book,
the value of . A used is in the range of .10−11 to .10−10 J/m (Ikeda et al. 2010; Bhowmik
et al. 2019; Kaushik et al. 2020).
The anisotropy energy is given by
{
. E ani = (K 1 m 2x ) + (K 2 m 2y ) + (K 3 m 2z )d V (2.12)
V

where . K 1 , K 2 , K 3 are the anisotropy constants. For our heavy-metal-ferromagnetic-

metal-hetero-structure-based devices exhibiting PMA, . K 1 = 0, . K 2 = 0, and . K 3 =
K P M A (.0.5 × 106 J/m.3 to.106 J/m.3 ). For saturation magnetization. Ms .=.2 × 106 A/m
(an approximate value for this kind of ferromagnetic layers) and . K P M A = .0.5 × 106
32 2 Introduction to Nanomagnetism and Spintronics

Fig. 2.4 Transition from model of atomic spins. S→1 ,. S→2 ,…(Heisenberg Hamiltonian) to the analytical
micromagnetic model (magnetization .m → treated as a continuous field, to be solved analytically),
and then to the numerical micromagnetic model (the continuous field is now discretized to apply
numerical methods), is shown here

J/m.3 , this leads to PMA field . HK (. 2K

Ms
) of 0.5 T (Ikeda et al. 2010; Bhowmik et al.
2019; Kaushik et al. 2020).
The Zeeman energy (energy due to interaction of the magnetic moments with
external magnetic field) is given by
{
. E Z eeman = − μ0 (m x Hx + m y Hy + m z Hz )d V (2.13)
V

The demagnetization/magneto-static/dipole energy, which accounts for interac-

tion between magnetic moments through dipole fields, is given by
{
μ0
. E di pole = − → H→di pole d V
m. (2.14)
2 V

The net energy of the system is given by

E = E exchange + E ani + E Z eeman + E di pole , where the individual energy terms
.
are given by Eqs. 2.11, 2.12, 2.13, and 2.14. The equilibrium value of magnetization
vector .m→ (.m x , .m y , .m z ) as a function of position .(x, y, z) is obtained by minimizing
the total energy.
It is to be noted that in this micromagnetic approach, we have moved from magne-
tization arising from discrete atoms, separated by the lattice constant, to a continuous
magnetization distribution (Fig. 2.4). In order to solve for this continuous distribution
numerically through micromagnetic simulation packages (Vansteenkiste et al. 2014),
2.3 Micromagnetic Model 33

again we have to go discrete, and a numerical grid needs to be chosen (Fig. 2.4). But
the mesh size (distance between two adjacent grid points) can be much larger than
the lattice constant. For example, to model domain walls in ferromagnetic films with
perpendicular anisotropy, a mesh size of about 1–2 nm is good enough, which is
one order higher than atomistic distances which are in Angstroms (Fig. 2.4). Thus
moving from the atomistic picture to a continuous picture and subsequently dis-
cretizing it helps in modelling the system at a scale which is fine enough but is not
as computationally resource intense as the atomic scale.
Formation of a Domain Wall: Domains are formed in the ferromagnetic layer
to minimize the magneto-static energy . E di pole . But they come at a cost: two domains
with moments anti-parallel to each other cannot be right next to each other without
a transition region. If there’s no transition region, two adjacent moments will have
opposite orientations, and then their contribution to exchange energy . E exchange will
be infinite, following Eq. 2.11. So minimizing . E exchange requires the formation of a
wide transition region between two adjacent domains, with the magnetization vector
rotating smoothly from one orientation (. Mz = 1 for PMA) to the opposite orientation
(. Mz = −1 for PMA) (Fig. 2.5). Higher the value of the exchange correlation (. A),
wider is the associated domain wall.
At the same time, minimizing the anisotropy energy . E ani requires reduction in the
size of the domain wall since inside the domain wall, the moments are not oriented
along the easy axis (out-of-plane or .z-axis for PMA) leading to higher anisotropy
energy, following Eq. 2.12 (Fig. 2.5). Thus the width of the domain wall (.δ) depends
on a trade-off between the . E exchange and . E ani and is given as follows (for thin films
exhibiting PMA) (Aharoni 2000; Blundell 2001):
/
A
.δ = π (2.15)
KPMA

Based on the aforementioned values of. A (.10−10 J/m) and. K P M A (.0.5 × 106 J/m.3 ),
the domain-wall width turns out to be .≈45 nm. Thus, when the lateral dimensions of
a nanomagnet exhibiting PMA are below .≈45 nm each, the single-domain/macro-
spin model can be applied to the nanomagnet; else, a micromagnetic model needs to
be used.

Fig. 2.5 The directions of

the magnetic moments inside
a Neel domain wall, which
extends from position .x = 0
to position .x = W (.W is the
domain-wall width), are
shown here
34 2 Introduction to Nanomagnetism and Spintronics

In ferromagnetic layers exhibiting PMA, a domain wall can either be a Neel wall or
a Bloch wall (Emori et al. 2013; Ryu et al. 2013; Bhowmik et al. 2015). The directions
of magnetic moments inside a Neel wall are shown in Fig. 2.5. The magnetization
→ as a function of position (.x, y, z) (.x − y-plane being
inside the Neel domain wall .m,
the film plane), can also be written as follows:
( ) ( )
πx πx
.m z = cos , m x = sin , my = 0 (2.16)
W W

If the domain wall is of Bloch type instead, the magnetization inside the wall is
given as follows:
( ) ( )
πx πx
.m z = cos , m y = sin , mx = 0 (2.17)
W W

For both the above magnetization profiles, the domain wall extends from position
. x = 0 to position.x = W (.W is the domain-wall width), with the moments for position
. x < 0 being vertically up (.+z) uniformly and moments for position . x > 0 being
vertically down (.−z) uniformly, since the ferromagnetic layer exhibits PMA. But in
the Neel wall, the magnetization rotates in the .x − z-plane (.m y = 0), as shown in
Fig. 2.5. In the Bloch wall, the magnetization rotates in the . y − z-plane (.m z = 0).
Whether a Neel wall is formed or a Bloch wall depends on various factors including
the width of the ferromagnetic nanowire/nanobar in which the domain wall is formed,
whether the domain wall is longitudinal or transverse, and also on the magnitude
of an asymmetric exchange interaction, called Dzyaloshinskii-Moriya Interaction
(DMI), at the heavy-metal-ferromagnetic-metal interface (Emori et al. 2013; Ryu
et al. 2013; Bhowmik et al. 2015). Both Neel walls and Bloch walls have been reported
experimentally in heavy-metal-ferromagnetic-metal hetero-structures, which are of
our interest in this book (Emori et al. 2013; Ryu et al. 2013; Bhowmik et al. 2015).
Micromagnetic simulations have shown that a high value of DMI supports Neel wall
formation (Emori et al. 2013, 2014; Ryu et al. 2013), while a low value of DMI
supports Bloch wall formation (Bhowmik et al. 2015).

2.4 Magnetic-Field-Driven Switching:

Landau–Lifschitz–Gilbert (LLG) Equations

Earlier in the chapter, while discussing the single-domain/macro-spin model, we

only discussed how its energy term can be expressed (Eq. 2.2) and how the final
equilibrium state of the macro-spin . M → can be found through energy minimization
→
(Eqs. 2.5 and 2.8). Here, we describe the dynamics of the macro-spin vector . M(t) as
it starts from its initial state and reaches that final equilibrium state.
2.4 Magnetic-Field-Driven Switching: Landau–Lifschitz–Gilbert (LLG) Equations 35

The effective magnetic field . H→e f f (t) can be obtained from the net energy term in
Eq. 2.2 (now a function of time .t since the magnetization is also a function of time)
as follows:
∂E
. H→e f f (t) = − (2.18)
→
∂ M(t)

→
The dynamics of . M(t) as a function of time .t, in the presence of . H→e f f (t), is given
by Arrott (2005):

→ ( ( → ))
d M(t) → α d M(t)
. = −γ M(t) × H→e f f (t) − (2.19)
dt γ Ms dt

where .γ is the gyromagnetic ratio and .α is the damping factor. Here, . M(t) → × H→e f f (t)
corresponds to the magnetic precession term, but only having that term will mean that
the magnetization keeps rotating/precessing with the axis being along the direction
of the effective magnetic field . H→e f f . But we know from the single-domain model
discussed before that in the end, at equilibrium, the magnetization is such that the
net energy term in Eq. 2.2 is minimized and . M → aligns along . H→e f f . To enable this
→
in the dynamics equation, the .α dt term has been added to the effective magnetic
dM

→
field. Considering that the norm of the macro-spin vector (.| M(t)|) is conserved in
the single-domain model, Eq. 2.19 reduces to the Landau–Lifschitz–Gilbert (LLG)
equation given below Arrott (2005):

→
d M(t) γ αγ
. =− →
( M(t) × H→e f f ((t)) − →
( M(t) →
× ( M(t) × H→e f f (t)))
dt 1 + α2 Ms (1 + α2 )
(2.20)
For the single-domain magnet exhibiting PMA, for which we have solved the
energy minimization equation earlier, we now solve the LLG equation numerically
over time as the applied magnetic field is varied from a high positive value to a high
negative value and back to the high positive value, while being applied first in the
out-of-plane direction (.z). The hysteresis plot we thus obtain (shown in Fig. 2.6) is
similar to what we obtained earlier from energy minimization (Fig. 2.2c), but in this
case, we do see that as we increase the temporal rate at which the field is varied (ramp
rate), the switching field goes down (Fig. 2.6). This is an effect which can only be
captured if the dynamics of the system is modelled as we have done here. Writing
the numerical code to solve LLG equations and obtain these hysteresis/easy-axis-
switching plots has been left as an exercise at the end of this chapter. The parameters
in the same code can be changed to obtain hard-axis plots like in Fig. 2.3b.
As another example of magnetization dynamics, obtained by solving the LLG
equations, we show in Fig. 2.7 how fast the magnetization aligns with a constant
applied magnetic field (0.1 T applied along .z-axis from time .t = 0, no anisotropy
field has been considered here), as a function of the damping constant .α (in Eq. 2.20).
36 2 Introduction to Nanomagnetism and Spintronics

Fig. 2.6 Out-of-plane magnetization (. Mz ) versus out-of-plane magnetic field plot, obtained by
solving the LLG equation numerically at every value of applied magnetic field. The magnetic field
has been swept uniformly from 8 T to .−8 T, and back to 8 T, at a ramp rate of 7 T/ns (in (a)) and
0.07 T/ns (in (b))

At .t = 0, the magnetization is along the .x-axis. We observe that for a low value of
α, the magnetization precesses for a long time around the axis of the applied field
.

(sinusoidal behaviour of . Mx and . M y with time) and slowly aligns along the z-axis
(. Mz increases to 1) (Fig. 2.7a) (Arrott 2005). But when .α is high, the precession
decays very fast and the magnetization aligns with the applied field (its equilibrium
direction) (Fig. 2.7c). The period of the oscillation is.≈0.37 ns, and the corresponding
frequency is .≈2.7 GHz. With the gyromagnetic ratio (in Hz/T) being 28 GHz/T and
0.1 T being applied, the frequency value obtained makes intuitive sense. Writing the
numerical code to solve LLG equations and obtain these magnetic precession plots
has been left as an exercise at the end of this chapter.

2.5 Spin-Orbit-Torque-Driven Switching:

Landau–Lifschitz–Gilbert–Slonczewski (LLGS)
Equations

In nanomagnetic memory devices, information is stored as the state of magneti-

zation in ferromagnets. Using these devices for memory, logic, and neuromorphic
applications involves switching the magnetization. If an external magnetic field is
used to switch the magnetization as shown above, the power dissipation is very high
because a lot of current needs to flow through a conducting material to generate a
high magnetic field, thereby resulting in a large amount of Joule heating. Control-
ling ferromagnetic switching with spin-polarized electrons through current injection
has proven to be a lower power consuming and more scalable option compared to
2.5 Spin-Orbit-Torque-Driven Switching … 37

Fig. 2.7 An external magnetic field is applied in the .z direction starting from time .t = 0. The
magnet has no anisotropy. Three different values of damping constant (.α) are considered: a 0.01,
b 0.05, c 0.1

magnetic-field-driven switching (Apalkov et al. 2016; Lee and Lee 2016). This use
of the flow of spin-polarized electrons (spin current) to switch the magnetization
marks the transition from nanomagnetism to spintronics.
Spin transfer torque (STT) was the first popular mechanism used in magnetic
memory devices to generate spin current and switch the magnetization. In a spin valve
or a magnetic tunnel junction, as electrons pass through the ferromagnetic pinned
layer, they become spin polarized. When these electrons enter the ferromagnetic free
layer, they transfer spin angular momentum to the free layer and thus manipulate
the magnetization of the free layer. This mechanism is the working principle of Spin
Transfer Torque Magnetic Random Access Memory (STTMRAM), the schematic
of which has been shown in Chap. 1 (Apalkov et al. 2016; Lee and Lee 2016).
38 2 Introduction to Nanomagnetism and Spintronics

STT has also been used to drive ferromagnetic-domain-wall motion using very
low current pulses and has been used for the operation of magnetic racetrack mem-
ories (Parkin et al. 2008). When spin-polarized electrons coming from a uniformly
magnetized region impinge on a domain wall, they try to align the moments of the
domain wall in the direction of their own spin, resulting in domain-wall motion.
In heavy-metal/ferromagnetic-metal hetero-structures, an alternative to STT for
magnetic switching and domain-wall motion is spin-orbit torque (SOT), the origin of
which can be Rashba effect or spin Hall effect as mentioned before. We first present
here a brief overview of seminal experimental reports on SOT-driven magnetic
switching. We then discuss how the LLG equation presented above (Eq. 2.20) can
be modified to incorporate SOT and model this experimentally observed SOT-driven
switching.
In Miron et al. (2010), reported the measurement of SOT in a heavy
metal/ferromagnet/oxide tri-layer structure: a 0.6-nm-thick Co layer was used as
the ferromagnetic layer, sandwiched between a 3-nm-thick Pt layer (heavy metal)
and a 1.6 nm AlOx (oxide) layer. Due to the Rashba effect, Co magnetic moments
experience an effective magnetic field when an in-plane current flows through the
Co layer (Fig. 2.1a). 0.5 .µm wide and 5 .µm long nanowires are fabricated from this
Pt/Co/AlOx stack, which also exhibits PMA. Starting from a saturated state of the
magnet in the out-of-plane (.+z) direction, application of a current pulse at a zero
magnetic field was found experimentally to nucleate reverse polarized domains in
the magnet.
In Miron et al. (2011), deterministic switching between vertically up (.+z) and
vertically down (.−z) states of the magnet in a 500 nm by 500 nm square Co dot made
from the same stack by applying current pulses parallel to an externally applied in-
plane magnetic field was demonstrated. When the applied in-plane magnetic field
is positive (along .+x), a negative current pulse (red colour) of magnitude 2.6 mA
switches the magnet up (. Mz = 1), while a positive current pulse (black colour) of the
same magnitude switches the magnet down (. Mz = −1). When the applied field is
negative (along .−x), a negative current pulse switches the magnet down (. Mz = −1)
and a positive pulse switches the magnet up (. Mz = 1) (Table 2.1).
Liu et al. also reported the same deterministic switching on micron-wide Hall bars
fabricated from the Pt (2 nm)/Co (0.6 nm)/AlO.x stack, exhibiting PMA (Table 2.1), in
Liu et al. (2012b). But they ascribed the origin of SOT to spin Hall effect (SHE) in the
heavy metal (Pt) and resultant accumulation of spin-polarized electrons at the heavy-
metal-ferromagnetic-metal interface (Fig. 2.1b), as opposed to Rashba effect at the
same interface. Similar to the experiment by Miron et al., here also, in the presence of
an externally applied in-plane magnetic field, current pulses flowing along the axis of
the field generate hysteric magnetic switching between up (. Mz > 0) and down states
(. Mz < 0). Switching the direction of the applied field changes the sense/handedness
of the switching curve (Table 2.1).
Liu et al. carried out similar switching experiments on micron-wide Hall bars
fabricated from Ta (4 nm)/CoFeB (1 nm)/MgO (1.6 nm)/Ta (1 nm) stack, which
also exhibits PMA, and reported the results in Liu et al. (2012a). Here, Ta is the
heavy metal instead of Pt, CoFeB is the ferromagnet instead of Co, and MgO is the
2.5 Spin-Orbit-Torque-Driven Switching … 39

Table 2.1 Experimentally observed spin-orbit-torque-driven magnetic switching in the Co/Pt/AlO.x

stack and the Ta/CoFeB/MgO stack, both of which exhibit PMA. The final magnetization state,
hence, is along .+z (. Mz = 1) or .−z (. Mz = −1). The current and the magnetic field are both applied
in-plane, along the same axis. Only their signs are reversed to form the different cases shown here
Material stack Magnetic field Charge current (mA) Stable magnetization
state
Pt/Co/AlO.x Positive 2.58 . Mz = −1
Pt/Co/AlO.x Positive .−2.58 . Mz = +1
Pt/Co/AlO.x Negative 2.58 . Mz = +1
Pt/Co/AlO.x Negative .−2.58 . Mz = −1
Pt/Co/AlO.x 10 mT .>15 . Mz = −1
Pt/Co/AlO.x 10 mT .<−15 . Mz = +1
Pt/Co/AlO.x .−10 mT .>15 . Mz = +1
Pt/Co/AlO.x .−10 mT .<−15 . Mz = −1
Ta/CoFeB/MgO 10 mT .>1 . Mz = +1
Ta/CoFeB/MgO 10 mT .<−1 . Mz = −1
Ta/CoFeB/MgO .−10 mT .>1 . Mz = −1
Ta/CoFeB/MgO .−10 mT .<−1 . Mz = +1

oxide layer instead of AlO.x . Similar deterministic switching is also observed for
Ta/CoFeB/MgO, but the handedness of the current-induced hysteresis is opposite
to that of Pt/Co/AlO.x (Table 2.1), and the switching current density is also lower
compared to that for Co/Pt/AlO.x . Since Liu et al. ascribed the origin of SOT to spin
Hall effect (SHE), they concluded that the spin Hall angle (.θ S H E ), or the ratio of the
spin current to the charge current, for Ta is larger in magnitude and opposite in sign
compared to that for Pt.
We next show how the LLG equation discussed in the previous section, which
models the dynamic of a single-domain magnet under the application of magnetic
fields, can be modified to include SOT-driven magnetization dynamics. The modified
LLG equation, known as Landau–Lifschitz–Gilbert–Slonczewski (LLGS) equation,
which represents the magnetization dynamics inside the ferromagnetic layer, is as
follows:

→
d M(t) γ α
. =− →
( M(t) × H→e f f (t)) − γ →
( M(t) →
× ( M(t) × H→e f f (t))) − γ τ→S L
dt 1 + α2 1 + α2
(2.21)
where .τ→SL accounts for SOT due to the in-plane current flowing through the heavy-
metal layer below the ferromagnetic layer. .τ→SL is given as follows:
β → →
τ→
. SL = ( M(t) × ( M(t) × σ→ )) (2.22)
1 + α2
40 2 Introduction to Nanomagnetism and Spintronics

This SOT term is known as Slonczewski torque, or also as anti-damping torque

(since it acts against the damping torque due to a magnetic field). There’s another term
known as field-like torque (since it resembles the precession torque term due to a mag-
netic field), but we ignore it in our simulation next because it cannot explain the kind
of switching obtained experimentally, which we have discussed above (Table 2.1).
It is to be noted here that the same equations (Eqs. 2.21 and 2.22) are also valid for
STT, only the direction of spin accumulation (.σ) changes (more details in suggested
exercises at the end of the chapter).
Following the spin-Hall-effect-based explanation for spin-orbit torque in Liu
et al. (2012a, b), .σ→ corresponds to the direction of spin polarization of the electrons
that accumulate at the heavy-metal-ferromagnetic-metal interface, and
Jc θ S H E h
.β= (2.23)
Msat ed

where . Jc is the charge current density flowing through the heavy metal, .θ S H E = spin
Hall angle of the heavy metal, .d = thickness of the ferromagnetic layer, and .e =
charge of an electron.
Next, we numerically simulate the LLGS equation (Eq. 2.21) and obtain magne-
tization trajectories, as shown in Fig. 2.8, for different polarities of applied charge
current and magnetic field (writing the numeric code and obtaining these results is
left as an exercise at the end of this chapter). In all cases, charge current. Jc flows along
the same axis in which the external magnetic field is applied. This makes the spin
polarization (due to the current) .σ→ orthogonal to the direction of the magnetic field.
The magnitude of the current density (. Jc ) is always 10.6 A/cm.2 , while the magnitude
of the magnetic field is always 50 mT. Only their polarities are reversed, one at a
time, to obtain four possible cases as shown in Fig. 2.8. We observe from our numer-
ical LLGS solution that when both the current and the field have the same polarity,
starting from . Mz = 1, the magnet switches to .−z (. Mz = −1), i.e., . Mz = −1 is the
stable state of the magnet. But when the current and the field have opposite polarities,
starting from . Mz = 1, the magnet stays at that state, i.e., . Mz = 1 is now the stable
state of the magnet. This switching is consistent with that experimentally obtained
for Pt/Co/AlOx, as shown in Table 2.1.
For the case of Ta/CoFeB/MgO, the sign of .θ S H E in Eq. 2.23 changes. So now, for
a given polarity of current density, the direction of spin accumulation (.σ→ ) is opposite
of that in Fig. 2.8. So now, when current and magnetic field have the same polarity,
. M z = 1 is the stable state, and when current and magnetic field have the opposite

polarity, . Mz = −1 is the stable state. This explains the experimentally obtained

switching in Ta/CoFeB/MgO devices, shown in Table 2.1. More details on numerical
simulation of SOT-driven magnetic switching can be found in Bhowmik et al. (2012).
It is to be noted here that the LLGS-equation-based modelling of SOT-driven
switching discussed here is applicable for single-domain magnets (very small size
magnets effectively) in the absence of thermal field (this is known as the preces-
sional regime for switching). As a result, the prediction from this model matches
2.5 Spin-Orbit-Torque-Driven Switching … 41

Fig. 2.8 Time dynamics of the magnetization, as obtained from solving the LLGS equation numer-
ically for a nanomagnet exhibiting PMA, under the application of a constant external magnetic field
in .+/−y direction and charge current in .+/−y direction, which leads to accumulation of electrons
with spin polarization in .+/−x direction. Different cases are shown here: a current in .+y, field in
.+y b current in .−y, field in .+y c current in .+y, field in .−y d current in .−y, field in .−y

with experimental results qualitatively. For example, as mentioned above, the differ-
ent switching directions predicted for different polarities of in-plane field and in-plane
SOT-causing current, which this single-domain model predicts, match with the exper-
imental results shown in Table 2.1. But for a quantitative match of required switch-
ing current, required in-plane magnetic field, etc., SOT-driven nucleation of reverse
domains and SOT-driven domain-wall motion (where the single-domain assumption
doesn’t hold) need to be modelled. This has been discussed in detail in Lee et al.
(2014) and Bhowmik et al. (2016b). Also, thermal field needs to be incorporated in
the LLGS equation which will make the switching stochastic. In fact, because of the
thermal field, the magnet can switch below the critical current density predicted by
LLGS-equation-based modelling which doesn’t include the thermal field. This has
been discussed in detail in Apalkov et al. (2016) and Lee and Lee (2016).
42 2 Introduction to Nanomagnetism and Spintronics

SOT-driven domain-wall motion has also been considered for synaptic devices
in neuromorphic computing, and we discuss that in detail later in the book. There
we show that we can use a micromagnetic package called “mumax3” (Vansteenkiste
et al. 2014) and numerically solve the LLGS equation (Eq. 2.21) for several magnetic
moments in a micromagnetic framework, as opposed to just for one macro-spin that
we have shown here. This will help us model the SOT-driven domain-wall dynamics
and calibrate it with experiments (Emori et al. 2013, 2014; Ryu et al. 2013; Bhowmik
et al. 2015). Also later in the book, we discuss SOT-driven magnetic oscillation in spin
Hall nano-oscillators (SHNO) and the application of SHNOs in neuromorphic and
Ising computing. To model SHNOs, we again numerically solve the LLGS equation
(Eq. 2.21), both in the single-domain and the micromagnetic frameworks.

Suggested Exercises

Q1. Simulate the hysteresis of a single-domain magnet with perpendicular anisotropy

(PMA) using LLG equations described above. Use the following parameters to
describe the PMA magnet: saturation magnetization . Ms = 1.3 × 106 A/m and PMA
constant . K = 9 × 105 J/m.3 . Apply magnetic field in the out-of-plane direction,
ranging from a negative value to a positive value comparable to or larger than the
anisotropy field. Ramp the field up and down in small steps, and obtain the magneti-
zation values some time interval after your field change (this determines your ramp
rate). Plot out-of-plane component of magnetization (.z component) as a function
of the magnetic field using LLG equations. You shall get a square-shaped hystere-
sis curve with width/switching field proportional to the anisotropy field (easy axis).
Repeat the numerical experiment for different ramp rates and different anisotropy
constants. Show all your plots and perform a comparative study. Some example plots
are provided in Fig. 2.6 of the chapter.
Q2. Simulate the magnetic precession of a single-domain magnet (neglecting any
anisotropy) under an applied magnetic field of 0.1 T, using LLG equations described
above. Obtain precession plots for different damping factors 0.01, 0.05, 0.1, etc. (like
that shown in Fig. 2.7) and compare the results.
Q3. (a) Take a single-domain magnet with PMA (to be modelled by LLG equations
like in Q1 above). Starting from magnet in vertically upward direction (.+z), apply
magnetic field in vertically downward direction (.−z) in the model, and solve the
LLG equation in the absence of any spin-transfer torque term. First start with small
field and then increase the field to the anisotropy field and more. Note down the
field at which magnet starts switching. Also record the trajectory of the magnet
during this magnetic-field-driven switching. Vary the damping parameter and study
the influence of damping parameter on switching field. Use simulation parameters
from Q1 above if needed.
(b) Repeat the same simulation in the presence of spin-transfer torque (STT)
instead of magnetic field. The magnetization of the pinned layer should be opposite
to the initial direction of magnetization of free layer (which is modelled by LLGS
References 43

equations here), i.e., if the magnetization of the free layer is towards .+z initially,
that of the pinned layer is towards .−z. Vary the magnitude of current and note
down the magnitude at which switching starts happening. Also record the trajectory
of the magnet during this magnetic-field-driven switching. Now, vary the damping
parameter, repeat the same exercise, and plot the influence of damping parameter on
switching current. Compare your results with that you obtained for magnetic-field-
driven switching above.
(c) In the case of spin-orbit torque (SOT), electrons with spin polarized in-plane get
injected into the ferromagnet. Its influence on magnetization can be easily simulated
by changing the direction of magnetization of pinned layer, in the model you used for
part (b) above, from vertically up/down (.+/−z) to some direction in-plane. Please
do that and repeat the exercise pf part (b) above. You will see that the magnet does
not switch from .+z to .−z anymore; it goes in-plane and stays there as long as the
spin torque is present. What happens when the spin torque is removed?
(d) In order to make the magnet switch fully from .+z to .−z, an in-plane exter-
nal magnetic field needs to be applied in addition, orthogonal to direction of spin
polarization (direction of magnetization of the pinned layer). Please add that in the
simulation and see how the switching occurs. Change the polarity of the current and
the polarity of the external field and study how the nature of the switching changes.
You will obtain signature SOT-driven switching plots, like that shown in Fig. 2.8.

References

Aharoni A (2000) Introduction to the theory of ferromagnetism. Oxford University Press

Apalkov D, Dieny B, Slaughter JM (2016) Magnetoresistive random access memory. Proc IEEE
104(10):1796–1830
Arrott AS (2005) Introduction to micromagnetics. In: Heinrich B, Bland JAC (eds) Ultrathin
magnetic structures IV. Springer, pp 101–148
Bhowmik D, Nowakowski ME, You L, Lee O, Keating D, Wong M, Bokor J, Salahuddin S (2015)
Deterministic domain wall motion orthogonal to current flow due to spin orbit torque. Sci Rep
5(1):1–10
Bhowmik D, Lee OJ, You L, Salahuddin S (2016a) Magnetization switching and domain wall
motion due to spin orbit torque. In: Atulasimha J, Bandyopadhyay S (eds) Nanomagnetic and
spintronic devices for energy-efficient memory and computing. Wiley
Bhowmik D, Lee OJ, You L, Salahuddin S (2016b) Magnetization switching and domain wall
motion due to spin orbit torque. In: Atulasimha J, Bandyopadhyay S (eds) Nanomagnetic and
spintronic devices for energy-efficient memory and computing. Wiley, pp 165–187
Bhowmik D, Saxena U, Dankar A, Verma A, Kaushik D, Chatterjee S, Singh U (2019) On-chip
learning for domain wall synapse based fully connected neural network. J Magn Magn Mater
489:165434
Bhowmik D, You L, Salahuddin S (2012) Possible route to low current, high speed, dynamic
switching in a perpendicular anisotropy CoFeB-MgO junction using spin Hall effect of Ta. Proc
Int Electron Dev Meet (IEDM) 29.7.1–29.7.4
Blundell S (2001) Magnetism in condensed matter. Oxford University Press
Coey JMD (2010) Magnetism and magnetic materials. Cambridge University Press
Emori S et al (2014) Spin Hall torque magnetometry of Dzyaloshinskii domain walls. Phys Rev B
90:184427
44 2 Introduction to Nanomagnetism and Spintronics

Emori S, Bauer U, Ahn SM, Martinez E, Beach GSD (2013) Current-driven dynamics of chiral
ferromagnetic domain walls. Nat Mater 12(7):611–616
Ikeda S et al (2010) A perpendicular-anisotropy CoFeB-MgO magnetic tunnel junction. Nat Mater
9:721–724
Kaushik D, Singh U, Sahu U, Sreedevi I, Bhowmik D (2020) Comparing domain wall synapse with
other non volatile memory devices for on-chip learning in analog hardware neural network. AIP
Adv 10(2):025111
Lee OJ, Liu LQ, Pai CF, Li Y, Tseng HW, Gowtham PG, Park JP, Ralph DC, Buhrman RA (2014)
Central role of domain wall depinning for perpendicular magnetization switching driven by spin
torque from the spin Hall effect. Phys Rev B 89:024418
Lee S-W, Lee K-J (2016) Emerging three-terminal magnetic memory devices. Proc IEEE
104(10):1831–1843
Liu L, Pai CF, Li Y, Tseng HW, Ralph DC, Buhrman RA (2012a) Spin-torque switching with the
giant spin Hall effect of tantalum. Science 336(6081):555–558
Liu L, Lee OJ, Gudmundsen TJ, Ralph DC, Buhrman RA (2012b) Current-induced switching of
perpendicularly magnetized magnetic layers using spin torque from the spin Hall effect. Phys
Rev Lett 109(9):096602
Maekawa S, Valenzuela SO, Saitoh E, Kimura T (2017) Spin current. Oxford University Press
Miron IM et al (2010) Current-driven spin torque induced by the Rashba effect in a ferromagnetic
metal layer. Nature Mater 9:230–234
Miron IM et al (2011) Perpendicular switching of a single ferromagnetic layer induced by in-plane
current injection. Nature 476:189–193
Parkin S et al (2008) Magnetic domain-wall racetrack memory. Science 320:5873
Ryu KS, Thomas L, Yang SH, Parkin S (2013) Chiral spin torque at magnetic domain walls. Nat
Nanotechnol 8(7):527–533
Vansteenkiste A, Leliaert J, Dvornik M, Helsen M, Garcia-Sanchez F, Waeyenberge BV (2014) The
design and verification of MuMax3. AIP Adv 4:107133
Chapter 3
Introduction to Artificial Neural
Networks (ANN) and Spiking Neural
Networks (SNN)

3.1 Introduction to Supervised Learning

Most of the learning/training of NNs we carry out in this chapter is in supervised

mode, i.e., there’s an instructor who guides the training by providing correct output
labels for different input samples. We define supervised learning a little more formally
next.
If a true world picture for the given data (on which we are trying to train the NN
or any other ML model for that matter) is given by . y→ = f (→ x ), then the job of the ML
classifier (NN in our case) is to come up with . →ŷ = f , (→
x , w) such that . →ŷ is the closest
representative of . y→.
Note that supervised learning is not merely curve-fitting, which some people often
mistake it to be. This is because our target is not only to obtain high classification
accuracy for the training data set but also for the test data set (data which is not known
during training). This means our job is not just to fit a curve on the training data but
to understand the inherent process that generates the training data and anticipate the
test data that will be generated by the same process. This is what we mean by world
picture above. For example, in the case of MNIST data set of handwritten digits
(Deng 2012), on which we will train our designed neural networks in this chapter,
we have to develop an understanding of the human hand writing process to train well
on the data set.
As a result, a probabilistic interpretation is also popular for supervised learning,
which is as follows. The true world picture has the underlying probability distribution:
→ = y→|→
. p(Y x ).
The classifier’s job is to come up with . p(Y→ = →ŷ|→x , w) such that it is a true repre-
sentative of . p(Y→ = y→|→
x ). In the context of trying to grasp the world picture and not
just fit the model on the training data, we need to avoid the following two phenomena
while training ML models:
– Under-fitting: This happens when the model performs poorly on training data
itself. While our target is not to simply do “curve-fitting” on the training data (as
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 45
D. Bhowmik, Spintronics-Based Neuromorphic Computing, Springer Tracts in Electrical
and Electronics Engineering, https://doi.org/10.1007/978-981-97-4445-9_3
46 3 Introduction to Artificial Neural Networks …

stated above), we still need to classify the training data with reasonable accuracy,
which doesn’t happen in the case of under-fitting. Possible reasons for under-fitting
are that the ML model (NN say) is not complex enough, the hyper-parameters (like
learning ratio, as we will see later) of the model aren’t set at correct values, or the
training hasn’t been carried out for sufficient number of iterations or epochs.
– Over-fitting: Over-fitting happens when the model yields very high classification
accuracy on training data but much lower accuracy on test data. This is basically
what has been referred to earlier here as fitting the model on training data without
grasping the inherent process/true world picture which generates the data. Possible
reason for over-fitting is making the model too complex (e.g., a very large NN with
too many synaptic weights, more details below).

For a more in-depth analysis on ML/NN, the readers are recommended to go

through machine learning textbooks by Bishop (2006), or Goodfellow et al. (2016),
or Trappenberg (2020). In this chapter, we will discuss specific NN algorithms with
respect to the spintronics-based neuromorphic implementations that we will discuss
later in the book.

3.2 ANN: Fully Connected Neural Network (FCNN)

Without a Hidden Layer/Single Layer Perceptron (SLP)

We discuss a single-layer perceptron here (as shown in Fig. 3.1), i.e., there is no
hidden layer present. It can also be called fully connected neural network (FCNN),
without a hidden layer. We discuss this network here in the context of classifying
hand-written digits from the MNIST database (Deng 2012), but this method can be
extended to classifying other kinds of data, like that in other simple ML data sets
like Fisher’s Iris (Fisher 1936), Wisconsin Breast Cancer (WBC) data set (Agarap
2017), etc.
For the case of MNIST, number of nodes in input layer .= number of pixels of
each input image of a digit .= 28 .× 28 .= 784. Input to the nodes of the input layer
.{x 1 , x 2 , x 3 . . . . . .x 784 } corresponds to the intensities of the pixels. Number of nodes
in output layer .= number of digits .= 10. The desired output when input image is
of digit 0 is given by .{Y1 , Y2 , Y3 , . . . Y10 } = {1, −1, −1, . . . − 1}, for digit 1 is given
by .{Y1 , Y2 , Y3 , . . . Y10 } = {−1, 1, −1, . . . − 1}, and so on. The target of training the
network is such that for a given input, the output .{y1 , y2 , y3 , . . . y10 } at the output
layer of the network matches the desired output.
Following the standard FCNN training algorithm, output at any node n is given by

2
y = f (z n ) =
. n − 1;
1 + e−λzn
z n = wn,1 x1 +wn,2 x2 + · · · · · · wn,784 x784 + wn,0
= (∑m=1
m=784
wn,m xm ) + wn,0 , (3.1)
3.2 ANN: Fully Connected Neural Network (FCNN) ... 47

Fig. 3.1 Schematic of a fully connected neural network (FCNN) without a hidden layer

where . f is the non-linear activation function (tanh function in this case), and
wn,1 , wn,2 . . . . . . wn,784 , wn,0 -s are the synaptic weights, .wn,0 being the bias weight.
.
Multiplication between the input vector and the synaptic-weight matrix is called
a vector matrix multiplication (VMM) operation. This is the forward computation
process for this network.
We next describe how the weights of the FCNN are updated to train the net-
work. We use the very popular Stochastic Gradient Descent (SGD) algorithm, which
attempts to reach the global minimum of the error landscape (the weight parameters
are the axes here in this hyper-dimensional space) by travelling along the gradient
of the landscape. Error at output node .n is calculated as follows:

1
∈ =
. n (Yn − yn )2 (3.2)
2
The weight of the synapse connecting input node .m with output node .n is updated
between iteration .i and .i + 1 as follows:

∂∈n
wn,m
.
i+1
= wn,m
i
− Δwn,m = wn,m
i
−η
∂wn,m
ηλ
= wn,m
i
− (Yn − yn )(1 − yn2 )xm , (3.3)
2
and weight of the bias synapse for output node .n is updated as follows:
48 3 Introduction to Artificial Neural Networks …

∂∈n
wn,0
.
i+1
= wn,0
i
− Δwn,0 = wn,0
i
−η
∂wn,0
ηλ
= wn,0
i
− (Yn − yn )(1 − yn2 ), (3.4)
2
where .η is the learning rate (hyper-parameter in the model, taken to be 0.1 here). The
training sample is changed at every iteration to exhaust all examples in the training
set. Then this process is repeated several times, each repetition being called an epoch.
Thus, total number of iterations .= number of epochs .× number of training samples.
In Fig. 3.3a, we show our results for training this SLP/FCNN without a hidden
layer on the MNIST data set of handwritten digits. 10,000 samples, almost uniformly
distributed between the ten output classes (ten digits), are used for training, while
1,000 samples, distinct from the training samples, are used for testing. Training
accuracy versus epoch and test accuracy versus epoch are plotted in Fig. 3.3. While
the final training accuracy is high (.>90%), test accuracy saturates at around 70.%.
Without adding any hidden layer to the neural network, it is not possible to increase
the test accuracy considerably above this value. We discuss addition of a hidden layer
to FCNN next.

3.3 ANN: Fully Connected Neural Network (FCNN) with

Hidden Layers/Multi-layer Perceptron (MLP)

In a FCNN, the input vector (.x1 ,.x2 ,….xm ….x M ) corresponds to the data sample which
needs to be classified, as mentioned before. During forward computation for an
FCNN with one hidden layer (also called Multi-Layer Perceptron, or MLP), it is first
multiplied with the synaptic-weight matrix (VMM operation), and then a non-linear
activation function (“tanh” in our case again) is applied on the outcome of the MVM
operation. This generates the input for the next/hidden layer: (.z 1 , .z 2 ,…….z p ….z P ).
This is shown in the equations below and in Fig. 3.2:

z¯ = w p,1 x1 + w p,2 x2 · · · · · · + w p,m xm · · · +w p,M x M + w p,0

. p

2
z p = f (z¯p ) = −1
1 + (e−λz¯p )
(3.5)

As shown in Fig. 3.2, the inputs to the hidden layer are then multiplied with the
next synaptic-weight matrix as follows, and then “tanh” function is applied to yield
the final output: . y1 , . y2 ,……. y p …. y N .

y¯ = vn,1 z 1 + vn,2 z 2 · · · · · · + vn, p z p · · · +vn,P z P + vn,0 ;

. n

2
yn = f ( y¯n ) = − 1 (3.6)
1 + (e−λ y¯n )
3.3 ANN: Fully Connected Neural Network (FCNN) with Hidden … 49

Fig. 3.2 Schematic of a fully connected neural network (FCNN) with one hidden layer. (Reprinted
with permission from: Sahu U, Sisodia N, Muduli PK, Bhowmik D (2022) Ferrimagnetic synapse
devices for fast and energy-efficient on-chip learning on crossbar-array-based neural networks (a
device-circuit-system co-study). IEEE Trans Electr Dev 69(4):0018–9383 .© 2022 Sahu et al. 2022)

During the learning process, the final output . y1 , . y2 ,……. y p …. y P is then compared
with the desired output.Y1 ,.Y2 ,…….Y p ….Y P to generate the error, as shown in Fig. 3.2a.
The synaptic weights in Eqs. 3.5 and 3.6 are updated based on the error.
Update of the synaptic-weight matrix in Eq. 3.6 is given by

ηλ
Δvn, p =
. (Yn − yn )(1 − yn2 ))z p = Δvn z p (3.7)
2
where .Δvn = the common part of weight update generated at each node .n of the
output layer: (. ηλ
2
(Yn − yn )(1 − yn2 )). Learning rate (.η) is again taken to be 0.1.
Update of the synaptic-weight matrix in Eq. 3.5 is given by

λ
Δw p,m = (∑n=1
.
n=N
Δvn vn, p ) (1 − z 2p ))xm (3.8)
2
As evident from Eq. 3.8, weight updates for the synaptic weights between the first
layer and the hidden layer depend not only the common part of weight update at that
stage . λ2 (1 − z 2p )) but also on the synaptic weights, at the next stage, which connect
the hidden layer with the output layer (.(∑n=1 n=N
Δvn vn, p )). This is what is known as
back-propagation in machine learning literature (Fig. 3.2).
In Fig. 3.3b, we show our results for training this MLP/FCNN with one hidden
layer on the same MNIST data set. 10,000 samples, almost uniformly distributed
between the ten output classes (ten digits), are used for training, while 1,000 samples,
50 3 Introduction to Artificial Neural Networks …

Fig. 3.3 a Training and test accuracy of an FCNN with no hidden layer versus epoch number during
training. b Training and test accuracy of an FCNN with one hidden layer (with 250 nodes) versus
epoch number during training. Learning rate (hyper-parameter) is taken to be 0.1 in both cases

distinct from the training samples, are used for testing. Number of nodes in the hidden
layer in our design is 250. Training accuracy versus epoch and test accuracy versus
epoch are plotted in Fig. 3.3. While the final training accuracy is high (.>90%), test
accuracy saturates at around 70.%.

3.4 ANN: Convolutional Neural Network (CNN)

The FCNNs discussed in the previous sections involve multiplying different synaptic
weights with different input features, which can be different pixels of an input image.
However, this is not found to be a very efficient way of image classification because
there are too many synaptic weights involved, and everything is left up to the updates
of these weights, which can go in any trajectory hampering the whole training. In
CNNs, which have been found to be much more efficient for image classification
(once the images become much more complex compared to MNIST), instead of sev-
eral layers of VMM operation like in the FCNNs before, several layers of convolution
operation are first carried out on each image, as shown in Fig. 3.4. For each CNN

Fig. 3.4 Schematic of a convolutional neural network (CNN) showing the convolution operations
between a particular segment of the input image and convolutional weight kernels. (Reprinted
with permission from: Desai VB, Kaushik D, Sharda J, Bhowmik D (2022) On-chip learning of a
domain-wall-synapse-crossbar-array-based convolutional neural network. Neuromorphic Comput
Eng 2(2):024006 Desai et al. 2022)
3.5 Implementation of ANNs Through Crossbar Arrays … 51

operation, weight kernels corresponding to different features are first convolved with
different segments of the image as described below through the example in Fig. 3.4.
64 convolutional weight kernels, corresponding to 64 features (which need
to be extracted), are used. Each kernel is of .(3 × 3 × 3) size. The RGB com-
ponents of the first .(3 × 3 × 3) segment of the input image are denoted by
.(x 1 , . . . , x 9 , y1 , . . . , y9 , z 1 , . . . , z 9 ) say. The synaptic weights of the convolutional
kernels are represented by .u ik , vki , wki , where .i indicates the kernel number (from 1
to 64) and .k indicates the position in the kernel matrix (from 1 to 9), and .u, .v, and .w
correspond to R, G, and B, respectively. The output .oi , where .i indicates the kernel
number (from 1 to 64) and .βi denotes the bias term (one for each kernel) is given by

∑
9 ∑
9 ∑
9
o =
. i xk u ik + yk vki + z k wki + βi (3.9)
k=1 k=1 k=1

This way, the convolutional weight kernel (while keeping the weights unchanged)
is moved over various segments whose pixel indices differ from the first segment by
.(X, Y ) say. In that case, the above operation turns out to be

∑
9 ∑
9 ∑
9
o X,Y =
. i xk+X u ik + yk+Y vki + z k wki + βi (3.10)
k=1 k=1 k=1

Thus, a set of .oi values is created which is transferred to the next convolutional
layer, where some other convolutional kernels are used. Between the two convolution
operations, operations like max pooling and drop-out are often performed. In max
pooling, the highest among four adjacent .oi values is often chosen. This helps in
reducing the size of the image in the subsequent convolution operations. The purpose
of drop-out operation is to randomly drop some.oi values. This helps in reducing over-
fitting (mentioned in Sect. 3.1). Finally, the matrix is flattened into a single column
vector, which is applied on the FCNN (like in previous sections) to generate a 10-
dimensional column vector for output like before. The weights of the FCNN layer
and the weights of the convolutional kernels are all updated through gradient descent
method (following back-propagation) as before. More details on CNN training can
be found in (Desai et al. 2022).

3.5 Implementation of ANNs Through Crossbar Arrays of

Non-volatile Memory (NVM) Synapses

We next briefly discuss how the ANNs shown in the previous sections can be imple-
mented through crossbar arrays of non-volatile memory (NVM) synapse devices
(Christensen et al. 2022; Tsai et al. 2018; Sebastian et al. 2018; Chakraborty et al.
2020). The spintronic domain-wall synapse mentioned in Chap. 1 and to be covered
in more details later in the book is one such NVM synapse. A crossbar array stores
52 3 Introduction to Artificial Neural Networks …

Fig. 3.5 Instantaneous vector matrix multiplication (VMM) (shown by blue arrows) makes forward
computation on a crossbar array very fast; hence it speeds up both on-chip inference (a) and on-
chip learning (b). Instantaneous outer product calculation (shown by red arrows) enables parallel
weight-update operation in crossbar arrays; hence it speeds up on-chip learning (b) (more details
in Sadashiva et al. 2024)

the weights of the ANN in its synaptic devices (Fig. 3.5) and executes Vector Matrix
Multiplication (VMM) between the input vector and the corresponding weights in
a parallel fashion using Ohm’s law. This makes forward computation, needed both
for inference with trained networks or training of networks, very fast and energy
efficient. To train the networks in the crossbar arrays, the equations mentioned above
for training of ANNs need to be implemented on the crossbar array itself. As can be
seen from Eq. (3.3), the weight update at any synapse connecting input node .m with
output node .n is a product of a common part of weight update which depends on the
output node .n (. ηλ
2
(Yn − yn )(1 − yn2 )) and the input at node .m (.xm ). Such multiplica-
tion can happen simultaneously on all the synapses of a crossbar array (outer product
calculation), making crossbar arrays suitable for on-chip learning (training in custom
hardware instead of conventional computer) as well. While implementing the back-
propagation algorithm, crossbar arrays can provide further acceleration by imple-
menting Eqs. 3.6 and 3.7 very fast, as discussed in detail in Chap. 5 where we look
into spintronic domain-wall synapse-based crossbar array implementations of ANN.
The ANN training algorithms discussed in this chapter may need to be
suitably modified to simplify the peripheral circuitry needed to apply weight-
update/programming pulses on the NVM synapses and train the crossbar array. One
such modification is applying thresholding functions on the weight-update equations
mentioned above (Eqs. 3.7 and 3.8). This limits the weight update of any synapse
device at any iteration to .+1 bit (long-term potentiation, or LTP), .−1 bit (long-term
depression, or LTD), or no update. We discuss the thresholding algorithm in detail
in Chap. 5.
3.6 Spiking Neural Network (SNN) 53

Fig. 3.6 a LTP and LTD characteristics of an ideal NVM synapse, showing complete linearity and
symmetry b LTD and LTD characteristics of a non-ideal NVM synapse for different non-linearity
factors of LTP (.α P ) and LTD (.α D ) (more details in Sadashiva et al. 2024)

Another important thing that needs to be taken into consideration is the non-
ideality of weight-update characteristics of the NVM synapse: long-term potentiation
or LTP (positive weight update) and long-term depression or LTD (negative weight
update). For an ideal NVM synapse, LTP and LTD are perfectly linear and symmetric
(as shown in Fig. 3.6a), which leads to the highest possible accuracy of the crossbar
array. But for practical NVM synapses, including the domain-wall synapse discussed
in the book, asymmetry and non-linearity of LTP and LTD are non-zero as shown
in Fig. 3.6b. This leads to significant drop in accuracy as discussed in Chap. 5. Ways
to improve the non-linearity and symmetry of the domain-wall synapse, through
incorporation of notches or defects at the edges of the device, are discussed in Chap. 4.

3.6 Spiking Neural Network (SNN)

Thus far, we have discussed non-spiking version of neural networks, popularly known
as artificial neural networks (ANNs) as mentioned above. ANNs are much more
commonly used in conventional machine learning (ML), compared to their spike-
based counterparts (SNNs). But the inspiration that ANN derives from the brain is
rather loose: in the brain, neurons are connected to each other through synapses, and
in ANNs also, one layer of non-linear activation functions (called neurons) feeds to
the next layer of non-linear activation functions modulated by weights which can be
thought of as synapses (Fig. 3.1). But apart from this, there’s not much resemblance
between ANN and the brain. For example, ANNs deal with floating point numbers
while the brain processes in the form of temporal spikes. Weights of synapses in
ANNs are updated by minimizing the global error through gradient descent (back-
propagation algorithm, as discussed above). There is very little evidence to suggest
that such a thing happens in the brain; instead, the synaptic strengths in the brain
get modified through local spike-based rules as discussed next. As a result, in order
54 3 Introduction to Artificial Neural Networks …

to design brain-inspired or neuromorphic hardware, it’s imperative that we develop

a mathematical model for a neural network which is much closer to the brain. We
do that here by discussing spiking neural networks (SNNs). Much like the brain, an
SNN deals with information in the form of temporal spikes as discussed next.

3.6.1 Leaky Integrate Fire (LIF) Neurons

Depending upon how closely we want the neurons in SNN to represent biological
neurons, different mathematical models like Hodgkin-Huxley (HH) model, Integrate-
and-Fire (IF) model, Leaky-Integrate-and-Fire (LIF) model, Izhikevich model, etc.
can be chosen (Dayan and Abbott 2005; Trappenberg 2010). Among these, the HH
model very closely models the actual behaviour of the neuron (including how the
neuron spikes, as a result of the opening of channels which selectively allow sodium
and potassium ions between inside and outside of the neuron) but is computationally
complex. For the purpose of modelling of SNNs on a conventional computer or
implementation of SNNs on neuromorphic hardware, a much simpler model like
LIF is a better fit. Unlike the HH model, the LIF model doesn’t include the spike-
generation mechanism and cannot predict the shape of the spike (over time). But the
LIF model can predict the precise time at which a spike will be generated, and in
most cases, that is all that is needed from a neuron model in SNNs.
Following the LIF model, the membrane potential .v(t) of each neuron is governed
by the following equation:

dv(t)
C
. = −G L (v(t) − E L ) + I (t) (3.11)
dt

where . I (t) is input current to the neuron, .G L is membrane conductance, . E L is

resting potential of neuron (Fig. 3.7). Once .v(t) reaches the threshold potential (.Vth ),
the neuron generates a spike and .v(t) drops to . E L .
Based on the above equation, upon the application of a fixed DC current . I0 , the
final steady-state potential reached is . GI0L + E L . Only when this potential is higher
than threshold voltage .Vth , does the neuron spike.

Fig. 3.7 Equivalent circuit

for the LIF model of a neuron
3.6 Spiking Neural Network (SNN) 55

I0
. + E L ≥ Vth ⇒ I0 ≥ G L (Vth − E L )
GL

Thus the neuron spikes only when the input current is higher than the threshold
current .G L (Vth − E L ). Under the application of a fixed input current . I (t) = I0 ,
when . I0 is higher than a threshold .G L (Vth − E L ), the neuron spikes at a frequency
proportional to the magnitude of the current (. I0 ).

3.6.2 Spike Time-Dependent Plasticity (STDP) of Synapses

As mentioned earlier, unlike in ANNs implemented on conventional computer,

synapses in the brain do not update their strengths/weights by back-propagation but
by local learning rules. In the brain, the strength of each synapse is adjusted based
on the difference between the spike times of the pre-neuron and the post-neuron that
are connected through the synapse (Dayan and Abbott 2005; Trappenberg 2010).
Strengthening of the synaptic weight can be short term or long term (Dayan and
Abbott 2005). In the SNNs we deal with here, we use long-term potentiation and
depression only (LTP, LTD).
Spike Time-Dependent Plasticity (STDP) is such a local-rule-based weight-update
mechanism found experimentally inside the rat’s hippocampus (Bi and Poo 1998).
Say pre-neuron .i is connected to the post-neuron . j, and say .w j,i is the weight stored
in the synapse. Then following the STDP rule, when pre-neuron .i spikes at .tspike,i
and post-neuron . j spikes at .tspike, j , the weight update .Δw j,i is as follows:
tspike, j −tspike,i
−( )
Δw j,i = Γ1 e τ1
i f tspike, j > tspike,i ;
. t −t (3.12)
−( spike,i τ spike, j )
Δw j,i = −Γ2 e 2 i f tspike, j < tspike,i

where .Γ1 , .Γ2 , .τ1 , and .τ2 are constants related to STDP (Diehl and Cook 2015; Dayan
and Abbott 2005; Trappenberg 2010).
The weight update following the above STDP equation is plotted in Fig. 3.8.
As per this weight-update rule, when the post-neuron spikes after the pre-neuron,

Fig. 3.8 STDP

characteristic of a synapse
connecting a post-neuron
and a pre-neuron where the
post-neuron spikes at time
.t post and the pre-neuron
spikes at time .t pr e
56 3 Introduction to Artificial Neural Networks …

synaptic weight goes up (LTP) and when the post-neuron spikes before the pre-
neuron, synaptic weight goes down (LTP). Shorter the duration between the pre-
neuron spike and the post-neuron spike, larger is the magnitude of LTP or LTD.

3.6.3 SNN Composed of LIF Neurons and STDP-Following

Synapses

While the models of neurons and synapses described above are based on concrete
biological evidence, the design of SNNs composed on such neurons and synapses
(such SNN is meant to be implemented on neuromorphic hardware) is less based
on biological evidence (perhaps because of lack of biological data due to difficulty
in precisely mapping the brain at the neuronal level) and more based on the final
application/learning task at hand. Much like ANN, here also there are multiple layers
of neurons (spiking neurons in this case) connected by synapses. But the synaptic
weights are updated following the local STDP rule as mentioned above.
Diehl and Cook’s seminal paper on STDP-enabled training of SNN for classifica-
tion of MNIST data set (Diehl and Cook 2015) has provided the standard template for
designing such STDP-enabled SNN for subsequent neuromorphic research. Various
kinds of neuromorphic hardware, using different kinds of emerging devices, have
been proposed which implement this SNN algorithm. We show this SNN structure
briefly here and discuss further details of the SNN-training algorithm in Chap. 6
while discussing a spintronics-based implementation of this algorithm.
A schematic of the SNN structure used by Diehl and Cook (2015) is shown in
Fig. 3.9. The SNN has two layers: the input layer and the processing layer. Each node
of the processing layer contains an excitatory neuron as well as an inhibitory neuron.
All the neurons are modelled by LIF. When the input is applied on the input layer,
the signal propagated to the excitatory neurons of the processing layer is in the form
of a spikes, the frequency of which is directly proportional to the intensity of input.
Each excitatory neuron is connected to an inhibitory neuron in a one-to-one fashion,
whereas the inhibitory neuron is connected to all the excitatory neurons except for
the one it receives the input signal from (as shown in Fig. 3.9). The synapses which
connect input-layer neurons to the excitatory neurons of the processing layer store
weights which are updated through the STDP rule. The SNN learns patterns by
forming impressions of these patterns on the synaptic weights, i.e., if the SNN learns
to identify the digit “0” say, the synaptic weights are adjusted such that the intensity
map of the synapse array actually shows a “0.”
While neuromorphic hardware using various kinds of emerging devices has been
proposed using this SNN algorithm (Sengupta et al. 2016; Wang et al. 2018; Sahu
et al. 2019), the algorithm is not known to scale very well. Once hidden layers are
added, STDP-based training doesn’t work very well, and so STDP-enabled SNN is
not known to train well on complex data sets like CIFAR-10, CIFAR-100, etc. Hence,
other methods which don’t use STDP have been used to achieve high classification
3.6 Spiking Neural Network (SNN) 57

Fig. 3.9 Structure of a

STDP-enabled spiking
neural network, as proposed
by Diehl and Cook (more
details in Diehl and Cook
2015)

accuracy for SNN on complex data sets. One such method is training an ANN first
and then converting the ANN with final weight values to SNN (Sengupta et al. 2019;
Roy et al. 2020). The other method is training the SNN directly by using a modified
version of gradient descent/back-propagation algorithm which handles spikes by
converting them to smooth functions (without that, spikes can’t be differentiated and
back-propagation algorithm can’t be used) (Eshraghian et al. 2023). This technique
is known as surrogate gradient descent. The interested reader can go through recent
reports by Sengupta et al. (2019), Roy et al. (2020), and Eshraghian et al. (2023) to
know more about these techniques. In Chap. 6 of this book, we nonetheless focus
on STDP-enabled training of SNN because STDP being a local rule, it’s easier to
implement STDP in spintronic neuromorphic hardware compared to these other
techniques.

Suggested Exercises

Design a Fully Connected Neural Network (FCNN)/Multi-Layer Perceptron

(MLP) with one hidden layer, meant for training on the MNIST data set. Use
at least 100 nodes for the hidden layer.
Q1. Using the equations provided in this text, code up the whole FCNN/MLP
in a programming language of your choice and train it on the MNIST data set of
handwritten digits. Use about 50,000 samples for training and 10,000 samples for
testing. Plot training and test accuracy as functions of epochs.
58 3 Introduction to Artificial Neural Networks …

Now, assume that you will implement the same FCNN/MLP using a crossbar array
of synaptic devices with the following synaptic property for long-term potentiation
(LTP) or positive weight update:

G max − G min − An
,
. GP = −n ,
(1 − e P ) + G min , (3.13)
− Amax
1−e P

and the following for long-term depression (LTD) or negative weight update:

G max − G min n , −n ,max

. GD = − −n ,
(1 − e AD
) + G max , (3.14)
− Amax
1−e D

where .G max = maximum normalized conductance (.=1) corresponding to highest

value of conductance of the synapse, .G min = minimum normalized conductance
(.=0) corresponding to lowest value of conductance of the synapse,. Pmax = maximum
number of pulses (normalized) to sweep from .G min to .G max , .G P is the normalized
conductance during conductance increase (LTP), .G D is the normalized conductance
during conductance decrease (LTD), .n , is the pulse number normalized by maximum
number of pulses to sweep from .G min to .G max , and . A P and . A D are curve-fitting
constants (more details on this in Chap. 4).
Q2. Assume a very large positive value for . A P and a very large negative value
for . A D . Show analytically that this corresponds to nearly linear and symmetric LTP
and LTD curves, like that shown in Fig. 3.6a.
Q2. Now, decrease the magnitudes of. A P and. A D without changing their polarities.
Study numerically the impact of this change in magnitudes on the linearity and
symmetry of LTP and LTD behaviour. Show relevant plots. You will obtain plots like
that shown in Fig. 3.6b. Note that in Fig. 3.6b, different non-linearity coefficients .α P
and .α D are used. Mapping between .α P and . A P as well as between .α D and . A D can
be found in the manual provided with the NeuroSim simulator mentioned in Chaps. 4
and 5 (Luo et al. 2019).
Q3. Train the FCNN/MLP you designed in Q1 for different magnitudes of . A P
and . A D without changing their polarities (assume all synapses take the same values
of AP and AD for the given network). Plot/tabulate train and test accuracy numbers
for these different values of . A P and . A D .
Q4. Now, assume linear and symmetric LTP and LTD for the synapses, but the
conductance being limited to. N -bit resolution. The conductance gets updated by 1 bit
only both for LTP and LTD. Plot/tabulate train and test accuracy numbers for different
values of N. Starting from what value of. N does the accuracy of the network resemble
that you have obtained before, when the conductance and conductance updates had
infinite bit resolution?
Q5. Now, study the impact of limited bit resolution and non-linearity and asymme-
try (different magnitudes of . A P and . A D without changing their polarities), all acting
simultaneously, on the train and test accuracy numbers of the designed FCNN/MLP.
Plot or tabulate your results.
References 59

You will find the analysis you have carried out here and the related results use-
ful while reading Chaps. 4 and 5 of the book, where we will study the impact of
non-idealities in the domain-wall device’s LTP and LTD properties (non-linearity,
asymmetry, limited bit resolution) on the accuracy of the implemented neural
networks.

References

Agarap AF (2017) On breast cancer detection: an application of machine learning algorithms on

the Wisconsin diagnostic dataset. arXiv:1711.07831
Bi G, Poo M (1998) Synaptic modifications in cultured hippocampal neurons: dependence on spike
timing, synaptic strength, and postsynaptic cell type. J Neurosci 18(24):10464–10472
Bishop CM (2006) Pattern recognition and machine learning. Springer
Chakraborty I et al (2020) Resistive crossbars as approximate hardware building blocks for machine
learning: opportunities and challenges. Proc IEEE 108(12):2276–2310
Christensen DV et al (2022) 2022 roadmap on neuromorphic computing and engineering.
Neuromorphic Comput Eng 2(2):022501
Dayan P, Abbott LF (2005) Theoretical neuroscience: computational and mathematical modeling
of neural systems (computational neuroscience). MIT Press
Deng L (2012) The MNIST database of handwritten digit images for machine learning research
[Best of the Web]. IEEE Signal Proc Mag 29(6):141–142
Desai VB, Kaushik D, Sharda J, Bhowmik D (2022) On-chip learning of a domain-wall-synapse-
crossbar-array-based convolutional neural network. Neuromorphic Comput Eng 2(2):024006
Diehl PU, Cook M (2015) Unsupervised learning of digit recognition using spike-timing-dependent
plasticity. Front Comput Neurosci 9:99
Eshraghian JK, Ward M, Neftci EO, Wang X, Lenz G, Dwivedi G, Bennamoun M, Jeong DS,
Liu WD (2023) Training Spiking neural networks using lessons from deep learning. Proc. IEEE
111(9):1016–1054
Fisher RA (1936) The use of multiple measurements in taxonomic problems. Ann Eugenics
7(2):179–188
Goodfellow I, Bengio Y, Courville A (2016) Deep learning. MIT Press
Luo Y, Peng X, Yu S (2019) MLP+NeuroSimV3.0. In: Proceedings of the international conference
on neuromorphic systems (ACM)
Roy K, Jaiswal A, Panda P (2020) Towards spike-based machine intelligence with neuromorphic
computing. Nature 575(7784):607–617
Sadashiva A, Holla A, Bhowmik D (2024) Impact of non-idealities of synapse devices on on-
chip inference and learning performance of crossbar arrays, as studied through “Crossbar Sim”
simulator. Under review
Sahu U, Pandey A, Goyal K, Bhowmik D (2019) Spike time dependent plasticity (STDP) enabled
learning in spiking neural networks using domain wall based synapses and neurons. AIP Adv
9(12)
Sahu U, Sisodia N, Muduli PK, Bhowmik D (2022) Ferrimagnetic synapse devices for fast and
energy-efficient on-chip learning on crossbar-array-based neural networks (a device-circuit-
system co-study). IEEE Trans Electron Dev 69(4)
Sebastian A, Le Gallo M, Burr GW, Kim S, BrightSky M, Eleftheriou E (2018) Tutorial: brain-
inspired computing using phase-change memory devices. J Appl Phys 124(11)
Sengupta A, Banerjee A, Roy K (2016) Hybrid spintronic-CMOS spiking neural network with
on-chip learning: devices, circuits, and systems. Phys Rev Appl 6:064003
Sengupta A, Ye Y, Wang R, Liu C, Roy K (2019) Going deeper in spiking neural networks: Vgg
and residual architectures. Front Neurosci 13
60 3 Introduction to Artificial Neural Networks …

Trappenberg T (2010) Fundamentals of computational neuroscience. Oxford University Press

Trappenberg T (2020) Fundamentals of machine learning. Oxford University Press
Tsai H, Ambrogio S, Narayanan P, Shelby RM, Burr GW (2018) Recent progress in analog memory-
based accelerators for deep learning. J Phys D: Appl Phys 51(28):283001
Wang Z, Crafton B, Gomez J, Xu R, Luo A, Krivokapic Z, Martin L, Datta S, Raychowdhury A,
Khan AI (2018) Experimental demonstration of ferroelectric spiking neurons for unsupervised
clustering. In: Proceedings of international electron devices meeting (IEDM), pp 13.3.1–13.3.4
 Part III
Neuromorphic Computing Using
Spintronic Devices
Chapter 4
The Ferromagnetic Domain-Wall
Synapse Device

4.1 Working Principle of a Domain-Wall Synapse Device

The ferromagnetic domain-wall synapse device, as briefly discussed in Chap. 1, is

essentially an improvisation on the spin-orbit torque magnetic random access mem-
ory (SOTMRAM) (Fullerton and Childress 2016; Apalkov et al. 2016; Lee and Lee
2016). Much like SOTMRAM, spin-orbit torque due to in-plane current flowing
through the heavy-metal layer underneath the ferromagnetic-metal layer generates
spin-orbit torque in the domain-wall device and manipulates the magnetization. But
unlike each SOTMRAM device, which is a binary device fundamentally and stores a
zero or a one (one bit), the domain-wall device is nearly analog and can store multiple
bits as explained in this chapter (Chakroborty et al. 2020; Sengupta et al. 2016).
But much like SOTMRAM and its predecessor, spin-transfer torque MRAM
(STTMRAM), the domain-wall device utilizes the magnetic tunnel junction (MTJ)
structure for read-out of information. An MTJ device is a nanoscale device con-
sisting of two layers of ferromagnetic metal (free layer and fixed layer), separated
by an ultra-thin layer of insulator. The insulating layer is so thin that electrons can
tunnel through the barrier when an appropriate bias voltage is applied between the
two metal electrodes. As a consequence of spin-dependent tunnelling, the resis-
tance/conductance of the device depends on the relative orientation between the
magnetizations of the two ferromagnetic-metal layers. This phenomenon is called
tunnelling magneto-resistance (TMR) (Apalkov et al. 2016; Lee and Lee 2016).
The non-volatility of the MTJ device (and of STTMRAM, SOTMRAM, domain-
wall device in extension) arises from the fact that once the magnetization of a layer
is manipulated by an external pulse (magnetic field or electrical current), the magne-
tization doesn’t change further in the absence of any further pulse. This non-volatile
nature makes it suitable to be used as a non-volatile memory (NVM) binary device
for conventional computing (as cache memory or a stand-alone memory chip) or as
an NVM synaptic device for neuromorphic computing applications (Apalkov et al.
2016; Lee and Lee 2016).

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 63
D. Bhowmik, Spintronics-Based Neuromorphic Computing, Springer Tracts in Electrical
and Electronics Engineering, https://doi.org/10.1007/978-981-97-4445-9_4
64 4 The Ferromagnetic Domain-Wall Synapse Device

A heavy-metal/ferromagnetic-metal (free layer)/oxide/ferromagnetic-metal (fixed

layer) hetero-structure-based MTJ device can be modified to act as a NVM synaptic
device, as shown in Fig. 4.1 (Chakroborty et al. 2020; Sengupta et al. 2016). The
basic physics behind the device is that of SOT-driven domain-wall (DW) motion.
When in-plane “write” current flows through the heavy-metal layer, the domain wall
inside the free ferromagnetic layer experiences spin-orbit torque (SOT) (Miron et al.
2010; Emori et al. 2013; Ryu et al. 2013; Bhowmik et al. 2015). If the ferromag-
netic layer exhibits perpendicular magnetic anisotropy (PMA), and Dzyaloshinskii-
Moriya interaction (DMI) is also present at the interface, then the domain wall can
be of Neel type. This has been experimentally shown in Emori et al. (2013, 2014),
Ryu et al. (2013) and Sampaio et al. (2013). Those experimental reports also show
that such a domain wall moves due to SOT from the in-plane current, even in the
absence of magnetic field (Emori et al. 2013, 2014; Ryu et al. 2013).
The induced SOT moves the domain wall, and the conductance of the “read” path
(MTJ structure in Fig. 4.1) of the device changes as explained next. Say the domain
wall is initially located on the leftmost side of the free ferromagnetic layer. This
means the average magnetization of the ferromagnetic free layer/track, in which the
domain wall moves, is vertically upward. Since the average moment of the magnetic
fixed layer is vertically downward, following the physics of tunnelling magneto-
resistance (TMR) of an MTJ device, the conductance of the “read” path is the lowest
in this case (due to anti-parallel orientation of the moments of the fixed and free
layers): .G min .
Now as the domain wall moves rightward due to SOT, the average magnetization of
the free layer switches from the vertically up direction to the vertically down direction.
As a result, the relative orientation of the free and fixed layer gradually shifts from
anti-parallel to parallel configuration, resulting in increase of the conductance of

Fig. 4.1 Schematic of a heavy-metal/ferromagnet hetero-structure-based domain-wall synapse

device is shown here (Singh et al. 2020). White arrows correspond to the magnetic moments of the
ferromagnetic layers here: the lower green layer in which the domain wall (DW) is present is the
“free” layer and the upper green layer is the “fixed” layer of the MTJ structure. Rightward domain-
wall motion due to positive “write” current pulses through the heavy-metal layer results in increase
of conductance of the “read” path (MTJ structure): LTP (as shown in Figs. 4.2 and 4.3. Leftward
domain-wall motion due to negative “write” current pulses through the heavy-metal layer results
in decrease of conductance of the “read” path (MTJ structure): LTD (as shown in Figs. 4.2 and
4.3). (Reprinted with permission from: Kaushik D, Singh U, Sahu U, Sreedevi I, Bhowmik D (2020)
Comparing domain-wall synapse with other non volatile memory devices for on-chip learning in
analog hardware neural network. AIP Adv 10(2):025111 Singh et al. 2020)
4.2 Micromagnetic Simulation of the Domain-Wall Synapse Device 65

the read path (Bhowmik et al. 2019; Singh et al. 2020). When the domain wall is
at the rightmost end, conductance is maximum: .G max . This is known as long-term
potentiation (LTP), or positive weight update, of the synapse device.
However, upon application of in-plane “write” current of the opposite polarity, the
domain wall moves towards the left side of the free layer. This causes conductance
to decrease (i.e., the average magnetization of the free layer goes upward, aligning
antiparallel with that of the fixed layer). This is known as long-term depression
(LTD), or negative weight increase, of the synapse device.
While performing on-chip learning in neuromorphic crossbar arrays, the network
is trained by updating the weights of synapses after every iteration (LTP and LTD),
following the standard gradient descent algorithm, as explained in Chap. 3. Hence,
SOT-based domain-wall synaptic device is used for on-chip learning, with the weight
of the synapse being proportional to the conductance of the “read” path of the domain-
wall device, and the weight update carried out by applying current pulses through
the “write” path (Bhowmik et al. 2019).
While performing on-chip learning, the network is trained by updating the weights
of synapses after every iteration following the standard gradient descent algorithm.
SOT-based domain-wall synaptic device is used for on-chip learning. Weights are
updated by applying an appropriate vertical “write” current pulse to the heavy-metal
layer so that the domain wall will move accordingly. Hence, the required change in
the conductance of the synapse can be achieved (Bhowmik et al. 2019). In the next
section, we demonstrate the working principle of the domain-wall synapse device
(as described above) through micromagnetic simulation.

4.2 Micromagnetic Simulation of the Domain-Wall

Synapse Device

The physics behind the micromagnetic framework and the mathematical formal-
ism associated with it have been discussed in Chap. 2. Using that micromagnetic
framework, Divya Kaushik et al. (Singh et al. 2020) have carried out and reported
micromagnetic simulation of the domain-wall device on the GPU-accelerated micro-
magnetic package “mumax3,” which we discuss here. The device of lateral dimen-
sions (1000 nm .× 50 nm) is used for performing micromagnetic simulations here
(Emori et al. 2014). Platinum(Pt) is used as the heavy-metal layer because of its
advantage of low resistivity over other similar materials (reduces the Joule heating,
therefore energy needed per current pulse the domain-wall movement is reduced)
and high spin Hall angle (Liu et al. 2011, 2012a). The thickness of the free ferro-
magnetic layer above the Pt layer is taken to be 1 nm. The spin hall angle of Pt is
0.07. Since thickness of the heavy-metal (Pt) layer (10 nm) is higher than the spin
diffusion length (2–3 nm) (Berger et al. 2018; Qu et al. 2014), the spin current den-
sity injected into the free ferromagnetic layer by the Pt layer (. Js ) .= in-plane charge
current density (. Jc ) .× 0.07 (spin Hall angle) (Liu et al. 2011, 2012a, b).
66 4 The Ferromagnetic Domain-Wall Synapse Device

For the ferromagnetic free layer, the following micromagnetic simulation param-
eters are used: saturation magnetization (. Ms ) .= .7 × 105 A/m, anisotropy constant
(K) .= .8 × 105 J/m.3 , exchange-correlation constant (A) .= .1 × 10−11 J/m, damping
factor .= 0.3 (Emori et al. 2014), and Dzyaloshinskii-Moriya Interaction (DMI) .=
−3
.1.2 × 10 J/.m2 (Singh et al. 2020). The easy axis of the magnetization is in the
out-of-plane direction, i.e., the free layer exhibits perpendicular magnetic anisotropy
(PMA).
In these device simulations (Singh et al. 2020), the domain wall is allowed to move
by applying a current pulse of fixed duration (3 ns) and fixed magnitude (25 .µA),
corresponding to a current density of.≈5 × 106 A/cm.2 , which is applied to the heavy-
metal layer. The movement of the domain wall obtained by our micro-magnetic
simulations is shown in Fig. 4.2. Initially, the domain wall is present at the leftmost
edge of the free ferromagnet layer. A current pulse of fixed magnitude corresponding
to the threshold current density (.≈5 × 106 A/cm.2 ) is allowed to pass through the
heavy-metal layer. This, in turn, changes the conductance of the synaptic device either
by increasing or decreasing it by a fixed step value depending upon the polarity of the
in-plane “write” current applied. The current pulse of fixed magnitude (25 .µA) with
positive polarity and duration 3 ns applied to the heavy metal moves the domain wall
towards the right side of the free layer by fixed step, increasing the conductance by a
fixed step value, whereas the current pulse of fixed magnitude (25 .µA) with negative
polarity and duration 3 ns applied to the heavy metal moves the domain wall towards
the left side of the free layer by the fixed step and decreases the conductance by a
fixed step value. In Fig. 4.2, the blue colour represents the magnetic moments pointing
into the plane, or .m z = −1 (vertically downwards), and the red colour represents the
magnetic moments pointing out of the plane, or .m z = 1 (vertically upwards). The
region where the colour changes from blue to red corresponds to the domain wall.
The corresponding conductance versus pulse plot (LTP, LTD plot), as a result
of these LTP-causing and LTD-causing current pulses, is shown in Fig. 4.3. Based
on the .G max and .G min values calculated for the given MTJ device, it is found that
a positive polarity current pulse of fixed magnitude, i.e., 25 .µA, will increase the
conductance by .ΔG, i.e., .≈0.071 .× .10−3 mho (since the domain wall move towards
the right edge of the free layer by .≈20 nm due to each current pulse). Similarly, a
negative polarity current pulse of fixed magnitude, i.e., .−25 µA, will decrease the
conductance by the same amount, i.e., .≈−0.071 × 10−3 mho and hence the domain
wall will move towards the left edge of the free layer by .≈20 nm. This means that the
conductance value can be (.≈ − 0.071 × 10−3 mho, 0, .≈0.071 .× .10−3 mho), thereby
affecting the domain-wall movement by moving it by fixed step of approximately
20 nm either towards the left edge of the free layer or unchanged or towards the right
edge of the free layer, respectively.
4.2 Micromagnetic Simulation of the Domain-Wall Synapse Device 67

Fig. 4.2 a “Write” current pulses applied on the heavy-metal layer of constant magnitude and
opposite polarities, as a function of time. b Domain-wall motion in the ferromagnetic layer for
different time instants corresponding to different current pulses of fixed magnitude as shown in (a).
Domain wall moves from its initial position when in-plane current flows through heavy metal due to
spin-orbit torque on the magnetization of the ferromagnetic layer above it. The cross product of.Mavg
and.σ decides the direction in which the domain wall will move. Blue colour corresponds to moments
being vertically down and red colour represents moments vertically up (with respect to Fig. 4.1.
Rightward domain-wall motion corresponds to LTP (Fig. 4.3); leftward motion corresponds to LTD
(Fig. 4.3)). (Adapted with permission from: Kaushik D, Singh U, Sahu U, Sreedevi I, Bhowmik
D (2020) Comparing domain-wall synapse with other non volatile memory devices for on-chip
learning in analog hardware neural network. AIP Adv 10(2):025111 Singh et al. 2020)
68 4 The Ferromagnetic Domain-Wall Synapse Device

Fig. 4.3 Conductance response of “read” path versus in-plane “write” current pulse of the domain-
wall device in Fig. 4.1, which is its LTP/LTD characteristic. Conductance increases or decreases
in fixed steps (.ΔG) due to the application of fixed magnitude current pulses of opposite polarities
(positive polarity: LTP due to rightward domain-wall motion in Fig. 4.2, negative polarity: LTD due
to leftward domain-wall motion in Fig. 4.2). (Reprinted with permission from: Kaushik D, Singh
U, Sahu U, Sreedevi I, Bhowmik D (2020) Comparing domain-wall synapse with other non volatile
memory devices for on-chip learning in analog hardware neural network. AIP Adv 10(2):025111
Singh et al. 2020)

4.3 Experimental Demonstrations of the Domain-Wall

Synapse Device

The main application of LTP and LTD of non-volatile memory (NVM) devices, like
the domain-wall device, is as a synaptic element in a neuromorphic crossbar array.
As explained in Chap. 3 (and again in more detail in Chap. 5), neural networks are
trained on the crossbar array (on-chip learning) by updating the synaptic weights
through a thresholded version of the back-propagation algorithm (Kaushik et al.
2020). Thereby, positive weight update by 1 bit is deemed LTP and negative weight
update by 1 bit LTD. Non-linearity and asymmetry in LTP and LTD lead to significant
reduction in classification accuracy for the crossbar array because the actual weight
update on the device deviates from that calculated through the back-propagation
algorithm (relevant equations without thresholding provided in Chap. 3 and with
thresholding provided in Chap. 5) (Sun and Yu 2019).
In RRAM, LTP and LTD happen through the creation and annihilation of oxygen
filaments in the oxide layer of the device due to electrical pulses (Chakroborty et al.
2020). In PCM, LTP and LTD happen through amorphization and crystallization of
the phase change material in the device due to electrical processes (Tsai et al. 2018;
Sebastian et al. 2018). These processes are not particularly symmetric, i.e., it is not
like reversing the polarity of the pulse will straight up lead to crystallization or vice
versa. As a result, asymmetry is a major issue in these devices, particularly for PCM.
Non-linearity is also a major issue both for RRAM and PCM, with not all synaptic
states having equal gaps in conductance on the LTP-LTD curves. The conductance
tends to saturate rather quickly with application of pulses due to the inherent physics
of the device.
4.3 Experimental Demonstrations of the Domain-Wall Synapse Device 69

Micromagnetic simulations in the previous section predict a very linear and synap-
tic LTP and LTD characteristic for the domain-wall synapse device. If such synaptic
characteristic is indeed obtained experimentally, then this can be a major advantage
for the domain-wall synapse compared to other NVM synapse devices (resistive
random access memory (RRAM), phase change memory (PCM), etc.) in terms of
obtaining high classification accuracy for on-chip learning in crossbar arrays (Singh
et al. 2020). Also, the LTP-causing pulses are identical to each other and LTD-causing
pulses are identical to each other for the domain-wall synapse, making the associated
peripheral circuit design for on-chip learning in the crossbar array all the more simple
for the domain-wall synapse device (as we will see in the next chapter) (Kaushik
et al. 2020).
In this section, we briefly review various experimental demonstrations of the
domain-wall synapse device and compare the experimentally reported LTP and
LTD with that predicted from simulation in the previous section. While spin-
orbit-torque(SOT)-driven domain-wall motion in heavy-metal-ferromagnetic-metal
hetero-structures has been demonstrated through several experiments over the last
decade (Emori et al. 2013, 2014; Ryu et al. 2013; Bhowmik et al. 2015; Conte
et al. 2015), experimental demonstration of LTP and LTD in such devices through
domain-wall motion in steps is more recent.
In one of the early experimental demonstrations of the domain-wall synapse
(Zhang et al. 2019), Shuai Zhang et al. have deposited the following thin film
stack: Si substrate/Ta (10 nm)/CoFeB (1.2 nm)/MgO (1.6 nm)/Ta (20 nm), using
room-temperature sputtering without a post-annealing process. CoFeB acts as the
ferromagnetic layer in which the domain wall moves, Ta layer below it acts as the
heavy-metal current flowing through which generates SOT, and MgO layer promotes
PMA in CoFeB. 50.× 400.µm.2 Hall bar devices are fabricated from the stack through
photolithography and argon-ion milling. By passing small magnitude “read” current
through one of the bars and measuring anomalous Hall voltage across the orthogo-
nal bar, anomalous Hall resistance, proportional to out-of-plane magnetization in the
CoFeB layer, is measured. LTP- and LTD-causing “write” pulses are applied through
the same Hall bar through which “read” current pulses are applied, just that “write”
current pulses are much larger in magnitude than “read” current pulses and hence the
former can generate SOT to move the domain wall in the CoFeB layer and change the
anomalous Hall resistance of the device but the latter can’t. It is to be noted that for
this work and the other two works we discuss subsequently in this sub-section, the
final synaptic device structure with the “read” path (MTJ) is not fabricated. Hence,
anomalous Hall resistance, which is also proportional to the average out-of-plane
magnetization of the ferromagnetic layer in which the domain wall moves, is used
as a substitute for conductance to obtain the synaptic characteristic (LTP, LTD) of
the device.
Zhang et al. (2019) have experimentally measured anomalous Hall resistance of
the fabricated Hall bar which is plotted as a function of the “write” current pulse
number (positive pulses: LTP, negative pulses: LTD). They observe that while for the
initial pulses LTP and LTD look quite linear, as the pulse number goes up, anomalous
Hall resistance saturates (LTP and LTD become non-linear). Shuai Zhang et al. have
70 4 The Ferromagnetic Domain-Wall Synapse Device

also imaged the domain-wall motion in the ferromagnetic layer, corresponding to

the LTP and LTD steps, through magneto-optic Kerr effect (MOKE) imaging. The
MOKE images show experimentally that indeed LTP and LTD happen due to domain-
wall motion in these devices.
In another experimental demonstration of such a synapse device, Qiqi Zhang et al.
have first deposited a high-quality epitaxial Pt (heavy metal) layer through molecular
beam epitaxy (MBE) and then deposited Ti (2 nm)/CoFeB (1.2 nm)/MgO (2 nm)/W
(1 nm) stack on the Pt layer through magnetron sputtering (Zhang et al. 2022). From
the thin film stacks, they have fabricated Hall bar devices and characterized LTP
and LTD in them through anomalous Hall resistance measurement, much like Shuai
Zhang et al. LTP and LTD plots obtained by Qiqi Zhang et al. are much more linear
and symmetric than that obtained by Zhang et al. (2019). But unlike Shuai Zhang
et al., Qiqi Zhang et al. have used LTP-causing pulses of increasing magnitude and
LTD-causing pulses of increasing magnitude (Zhang et al. 2022). As mentioned
earlier, using non-identical pulses demands a much more complicated peripheral
circuit design because based on the current weight in a particular synapse device, an
appropriate pulse needs to be applied to carry out weight update.
Ram Singh Yadav et al. have characterized the non-linearity and asymmetry of
LTP (under identical pulses) and LTD (under identical pulses) of similar Pt (heavy
metal)/Co (ferromagnetic-metal)/SiO.2 -based domain-wall synapse devices (Yadav
et al. 2023a) through similar anomalous Hall resistance measurement (Fig. 4.4a),
followed by fitting the experimentally obtained LTP and LTD data to parameterized
equations provided in the literature of the NeuroSim simulator (Luo et al. 2019). This
same fitting method has been used to characterize various kinds of NVM synapse
devices (RRAM, PCM, ferroelectric devices) and thereby compare the linearity and
symmetry of their LTP and LTD behaviours.
The parameterized equation for LTP is
G max − G min − An
,
. GP = −n ,
(1 − e P ) + G min , (4.1)
− Amax
1−e P

and that for LTD is

G max − G min n , −n ,max
. GD = − −n ,
(1 − e AD
) + G max , (4.2)
− Amax
1−e D

where .G max = maximum normalized conductance (.=1) corresponding to highest

value of anomalous Hall resistance, .G min = minimum normalized conductance (.=0)
corresponding to lowest value of anomalous Hall resistance,. Pmax = maximum num-
ber of pulses (normalized) to sweep from .G min to .G max , .G P is the normalized con-
ductance during LTP, .G D is the normalized conductance during LTD, .n , is the pulse
number normalized by maximum number of pulses to sweep from .G min to .G max
(60), and . A P and . A D are curve-fitting constants.
Based on the experimentally obtained LTP and LTD on their device (Fig. 4.4a)
and the above equations, Ram Singh Yadav et al. obtain . A P = 0.2809 and . A D =
4.3 Experimental Demonstrations of the Domain-Wall Synapse Device 71

Fig. 4.4 a, b Experimentally obtained change in anomalous Hall resistance (. Rx y ) of the domain-
wall device as a function of “write” current pulse number under identical positive current pulses (in
red) for LTP (in red), and identical negative current pulses (in black) for LTD (in black), as reported
by Yadav et al. (2023a). c Cycle-to-cycle variation in LTP and LTD for the same device. (Reprinted
with permission from: Yadav RS, Gupta P, Holla A, Ali Khan KI, Muduli PK, Bhowmik D (2023)
Demonstration of synaptic behaviour in a heavy-metal-ferromagnetic-metal-oxide-heterostructure-
based spintronic device for on-chip learning in crossbar-array-based neural networks. ACS Appl
Electr Mat 5(1):484–497. American Chemical Society (ACS), Copyright 2023 American Chemical
Society Yadav et al. 2023a)

−0.2843 (Yadav et al. 2023a). Subsequently, using the table that maps between
this fitting constant and the non-linearity coefficient provided in the manual for
NeuroSim simulator (Luo et al. 2019), they obtain the non-linearity factor for LTP
(.α P ) to be 3.875 and that for LTD (.α D ) to be .−3.84. The corresponding asymmetry
.= |α P − α D | = 7.715.

When compared to other NVM synapse devices (as listed in Table 4.1), the non-
linearity and asymmetry factors obtained for the domain-wall device are not much
higher but not very low at the same time. For example, in electrochemical RAM
(ECRAM) device, non-linearity factors are as low as 0.01 (.α P ) and .−0.34 (.α D ).
In the following sub-sections, we first explain the cause of significant non-linearity
and asymmetry in LTP and LTD of domain-wall devices, which was not predicted
72 4 The Ferromagnetic Domain-Wall Synapse Device

Table 4.1 Summary of LTP and LTD characteristics for various NVM synaptic technologies, based
on recent experimental reports (Yadav et al. 2023a; Sahu et al. 2023; Lee et al. 2022; Goh et al.
2021)
NVM synapse Synaptic bit resolution Non-linearity of LTP Asymmetry
technology (.α P ) (.|α P − α D |)
RRAM (HfO.2 based) .≈5 4.88 8.58
Electrochemical RAM .≈7 0.01 0.35
Ferroelectric junction .≈5 6.76 14.14
(HZO based)
Domain-wall device .≈5 3.875 7.715

by the earlier micromagnetic simulation we covered, based on the report by Kaushik

et al. (2020). In a recent report Yadav et al. (2023b), carry out some modifications
in the micromagnetic simulation to model this non-linearity and asymmetry of LTP
and LTD. Then, in their report, Ram Singh Yadav et al. suggest modifications to the
domain-wall device structure so that the linearity and symmetry of LTP and LTD can
be improved (Yadav et al. 2023b). We discuss their analysis here briefly and then
summarize some recent experimental efforts in that direction.

4.4 Micromagnetic Simulation to Explain Experimentally

Obtained Non-linearity of LTP and LTD

In the micromagnetic simulation of Sect. 4.2, 3 ns long pulses are consecutively

applied on the domain-wall device without any delay between the pulses. As a result,
the entire domain-wall motion that is observed in the simulation (Fig. 4.2) is due to
SOT during each current pulse. Since the current magnitude is fixed for all the pulses,
the domain-wall velocity is fixed. And since the duration of each pulse is fixed, the
distance moved by the domain wall for every LTP pulse (and every LTD pulse) is
the same. So the LTP and LTD characteristics are linear and symmetric.
But in the experiments that have been reported above (Zhang et al. 2019, 2022;
Yadav et al. 2023a), between any two consecutive LTP pulses, and between any two
consecutive LTD pulses, there is a delay which is typically longer than the duration
of each pulse. In this section, as per the recent report by Yadav et al. (2023b),
micromagnetic simulation is shown taking into account a delay of 100 ns between
any two consecutive LTP pulses (and consecutive LTD pulses), each pulse being of
1 ns duration. The corresponding domain-wall motion is shown in Fig. 4.5 (taken
from the report by Yadav et al. 2023b).
Here, in Fig. 4.5, cyan colour represents moments vertically down and red colour
represents moments vertically up. A domain wall is initialized at the right end of
the ferromagnetic track. The domain wall moving right to left leads to LTP here
(magnetic moments gradually switch from vertically up to vertically down, leading
4.4 Micromagnetic Simulation to Explain Experimentally … 73

Fig. 4.5 a Magnetization configuration images “just after” programming pulse.n is applied for LTP
and “before” the next pulse (.n + 1) pulse is applied for LTP (the time gap between the two is the
delay time between pulses: 100 ns). Here, .n = 2, 3, 10, 11. The cyan colour represents vertical up
magnetic moments, i.e.,.m z = 1, and red represents vertically down state.m z = −1. b Magnetization
configuration images “just after” programming pulse .n is applied for LTD and “before” the next
pulse (.n + 1) pulse is applied for LTD (the time gap between the two is the delay time: 100 ns).
Here,.n = 22, 23, 30, 31. (Reprinted with permission from: Yadav RS, Sadashiva A, Holla A, Muduli
PK, Bhowmik D (2023) Impact of edge defects on the synaptic characteristic of a ferromagnetic
domain-wall device and on on-chip learning. Neur Comput Eng 3(3) Yadav et al. 2023b)

to increase in conductance of the “read” path, as per the configuration of the MTJ
shown in Fig. 4.1). The domain wall moving left to right (due to “write” current pulses
of opposite polarity) leads to LTD here (magnetic moments gradually switch from
vertically down to vertically up, leading to decrease in conductance of the “read”
path, as per the configuration of the MTJ shown in Fig. 4.1).
74 4 The Ferromagnetic Domain-Wall Synapse Device

As observed from the magnetization configuration images of Fig. 4.5, the domain
wall moves here at the velocity of 25 m/s during each pulse. Since each pulse is of 1
ns duration, this means the domain wall moves by 25 nm during a pulse, left or right
depending upon the current polarity (between “before” .nth pulse and “just after”
.n + 1th pulse, .n = 2, 3, 10, 11 here). But Fig. 4.5 also shows that during the delay of
100 ns between one pulse and the next, the domain wall may also move. During LTP,
between “just after 2nd pulse” and “before 3rd pulse” (the delay between 2nd and
3rd pulse basically), domain wall moves by 37 nm. Similarly in the delay between
3rd and 4th pulse, domain wall moves by 22 nm; between 10th and 11th pulse, it
moves by 2 nm; and between 11th and 12th pulse, it moves by 3 nm.
Thus, as per the report by Yadav et al. (2023b), it is observed that for LTP, the
domain wall moves much more during the delay times between the initial pulses
compared to that between the later pulses. The physical phenomenon responsible for
this is described next (Yadav et al. 2023b). During the initial pulses, the domain wall
is located towards the right end of the device; during the delay time between pulses,
when SOT is non-existent, the domain wall still wants to move right towards the centre
of the device to minimize the energy due to dipole-dipole interaction (magnetostatic
energy). Dipole-dipole interaction energy is minimum when the domain wall is at
the middle, with half of the magnetic layer polarized up and half polarized down, as

Fig. 4.6 LTP and LTD of the domain-wall device corresponding to the micromagnetic-simulation-
based snapshots of domain-wall motion in Fig. 4.5. 100 ns delay is applied between two consecutive
LTP pulses or consecutive LTD pulses. Average out-of-plane (OOP) magnetization (.m z ) at the end
of the delay is plotted as a function of pulse number. Thus, .m z here includes domain-wall motion
both during the pulse and during the delay. Hence, the LTP and LTD curves are non-linear and
asymmetric as explained in text. b For non-linearity and asymmetry quantification, .m z in (a) is first
converted to conductance (through TMR effect) and then normalized (minimum conductance .=
0, maximum conductance .= 1). Then, by fitting curves based on the equations provided with the
NeuroSim simulator (Luo et al. 2019), the following numbers are obtained: non-linearity factor of
LTP .α p .= 2.47, non-linearity factor of LTD .αd = −2.45, and asymmetry .= .|α P − α D | = 4.92.
(Reprinted with permission from: Yadav RS, Sadashiva A, Holla A, Muduli PK, Bhowmik D (2023)
Impact of edge defects on the synaptic characteristic of a ferromagnetic domain-wall device and
on on-chip learning. Neur Comput Eng 3(3) Yadav et al. 2023b)
4.5 Improving Linearity and Symmetry of LTP and LTD 75

shown both through magneto-optic Kerr effect (MOKE) imaging experiments and
micromagnetic simulations in the report by Bhowmik et al. (2015). During the later
pulses, the domain wall is already near the centre of the device; so its tendency
to move during the delay time, when charge current and spin current are zero, is
much less. This leads to non-linearity of the LTP curve as observed in Fig. 4.6a,
again taken from the simulation-based report by Yadav et al. (red plot: LTP, black
plot: LTD) (Yadav et al. 2023b). Using the method associated with the NeuroSim
simulator (Luo et al. 2019), we obtain the following: non-linearity coefficient for
LTP .α P = 2.47, non-linearity coefficient for LTD .α D = −2.45, and thus asymmetry
coefficient (.|α P − α D |).= 4.92 (Fig. 4.6b: red plot for LTP, black plot for LTD). These
numbers are comparable to that reported experimentally by Yadav et al. (in a previous
report by the same group, already discussed above) (Yadav et al. 2023a). This means
the modified micromagnetic modelling here can indeed explain the non-linearity and
asymmetry reported experimentally in similar devices (unlike the micromagnetic
modelling in the previous sub-section).

4.5 Improving Linearity and Symmetry of LTP and LTD

The micromagnetic simulation in the previous sub-section shows that motion of

the domain wall between LTP/LTD pulses, due to dipole fields, leads to the non-
linearity of LTP and LTD characteristics of the domain-wall synapse device. Such
stray domain-wall motion can be prevented by adding defects to the edges of the
device. These defects can be of two types: regions where the magnetic material is
removed (like notches or grooves) or regions where PMA is modified. Either way,
these defects are known to pin the domain wall and prevent its motion when current
pulse is not applied. This improves the linearity and symmetry of the LTP and LTD
of the device, as predicted from micromagnetic simulations including these defects.
For the case of material defects (defects with modified PMA), lack of domain-wall
motion between pulses is shown in Fig. 4.8 (taken from the same simulation-based
report by Yadav et al. 2023b). The corresponding LTP and LTD plots show much
more linearity and symmetry, as expected (Fig. 4.7, taken from Yadav et al. 2023b).
Similar micromagnetics-based analysis has been carried out for grooves or notches
pinning the domain wall by Liu et al. (2021) and Dhull et al. (2023).
We next briefly review some recent experimental demonstrations of the domain-
wall synapse device, in which notches or grooves have been engineered to enable
domain-wall pinning unlike in the experimental demonstrations discussed earlier in
this chapter. Durgesh Kumar et al. have deposited W (6 nm)/ CoFeB (1 nm)/MgO
(1 nm)/Ru (2 nm) thin film stacks (which exhibit PMA) through sputtering (Kumar
et al. 2023). Here, W acts as the heavy-metal layer. In order to improve the SOT
efficiency and thereby increase the domain-wall velocity (and reduce the current and
energy for LTP and LTD), they have used HP-W (3 nm)/LP-W (3 nm) for the W
layer, where HP-W refers to tungsten (W) deposited under high pressure and LP-W
means W deposited under low pressure. Subsequently, they have fabricated meander
76 4 The Ferromagnetic Domain-Wall Synapse Device

Fig. 4.7 LTP and LTD of the domain-wall device, including material defects, corresponding to the
micromagnetic-simulation-based snapshots of domain-wall motion in Fig. 4.5. 100 ns delay is still
applied between two consecutive LTP pulses or consecutive LTD pulses. (Reprinted with permission
from: Yadav RS, Sadashiva A, Holla A, Muduli PK, Bhowmik D (2023) Impact of edge defects on
the synaptic characteristic of a ferromagnetic domain-wall device and on on-chip learning. Neur
Comput Eng 3(3) Yadav et al. 2023b)

microwires of dimensions 10 .µm by 50 .µm, where neighbouring segments are offset

by a certain distance (meandering structure). These meanders act as domain-wall
pinning centres. Consequently, Durgesh Kumar et al. have demonstrated domain-
wall motion for LTP and LTD by steps of about 1 micron under very low current
density (only 10.3 A/cm.2 ) and have reported an energy budget as low as 27 aJ/bit for
the device. The domain-wall motion has been imaged here through magneto-optic
Kerr effect microscopy (Kumar et al. 2023).
Thomas Leonard et al. have fabricated an MTJ stack (with Ta as the heavy-metal
layer, CoFeB as the ferromagnetic-metal layer) exhibiting PMA and then fabricated
devices from it (Leonard et al. 2022). Geometric notches are fabricated on the two
edges of the device using a combination of electron beam lithography and ion milling.
Taking advantage of these notches which pin the domain wall, 3–5 extremely stable
conductance/synaptic states have been experimentally demonstrated in this work
(Leonard et al. 2022).

4.6 Alternative Spintronic Synapse Devices

Apart from the domain-wall synapse device, other spintronic devices have also been
proposed as NVM synapses in neuromorphic crossbar arrays. Innovations in this
regard mainly try to address two drawbacks of the domain-wall synapse device. One
drawback is that the domain wall typically has a width of 40–50 nm; so in order to
have say a 5-bit device, the domain wall needs to move in 32 LTP/LTD steps. This
means the length of the domain-wall device needs to be more than a micron. This
makes this technology difficult to scale. Another drawback of the domain-wall device
4.6 Alternative Spintronic Synapse Devices 77

Fig. 4.8 a For the device with material defects, magnetization configuration images are taken “just
after” programming pulse .n is applied for LTP and “before” the next pulse (.n + 1) pulse is applied
for LTP, with 100 ns delay existing between two consecutive LTP pulses. Here, .n = 2, 3, 10, 11.
The cyan colour represents vertical up magnetic moments, i.e.,.m z = 1, and red represents vertically
down state, i.e., .m z = −1. b Magnetization configuration images “just after” programming pulse
.n is applied for LTD and “before” the next pulse (.n + 1) pulse is applied for LTD, with 100 ns
delay existing between consecutive pulses. Here, .n = 22, 23, 30, 31. (Reprinted with permission
from: Yadav RS, Sadashiva A, Holla A, Muduli PK, Bhowmik D (2023) Impact of edge defects on
the synaptic characteristic of a ferromagnetic domain-wall device and on on-chip learning. Neur
Comput Eng 3(3) Yadav et al. 2023b)

is that, with the exception of the report by Kumar et al. (2023) mentioned above, most
ferromagnetic domain-wall devices still need significant amount of current to move
the domain wall. Typically current density of 10.6 A/cm.2 or 10.7 A/cm.2 is required
to move the domain wall, which means the energy consumed and Joule heating in
78 4 The Ferromagnetic Domain-Wall Synapse Device

the device per LTP/LTD pulse is not that low when compared to alternative synapse
devices like ferroelectric devices say, where an electric field is applied across an
insulator device (so very low current flow) to cause LTP/LTD.
The ferromagnetic/spintronic skyrmionic device, which is an alternative to the
domain-wall device, promises to overcome both these drawbacks. The skyrmion is
another topologically stable structure like the domain wall which occurs in ultra-thin
ferromagnetic layers but is much smaller in size compared to the domain wall. Hence
skyrmionic synapse devices, in which conductance changes due to the movement of
skyrmions as opposed to the domain wall, are supposed to be much smaller than the
domain-wall device. It has also been shown through both micromagnetic simulations
and experiments that skyrmions need much less current density to move compared to
the domain wall (Sampaio et al. 2013; Bhattacharya et al. 2019; Saxena et al. 2018).
Here, we briefly review a recent experimental demonstration of the skyrmionic
device. Kyung Mee Song et al. have fabricated a skyrmionic device based on a
Pt (heavy metal)/GdFeCo (ferrimagnet)/MgO multi-layer stack (Song et al. 2020).
Creation and motion of skyrmions in the device are captured through Hall resistiv-
ity measurement and scanning transmission X-ray microscopy (STXM). Upon the
application of current pulses of increasing magnitude in the order of 10.6 A/cm.2 , it
is shown through STXM imaging that skyrmions are created successively and they
move due to SOT from the current. As a result, the Hall resistivity changes as a func-
tion of pulse number for the device: LTP and LTD (Song et al. 2020). The LTP and
LTD here look quite linear and symmetric for this device, compared to experimental
reports on various domain-wall devices discussed earlier. But as just mentioned, the
LTP pulses are not identical to each other (same for LTD pulses). This will make
the design of the peripheral circuit for on-chip learning in crossbar arrays using
these synapse devices difficult. So this issue with the skyrmionic device needs to be
addressed.
In addition to the skyrmionic device, some recent reports promote the use of
the ferrimagnetic domain-wall device as opposed to the ferromagnetic domain-wall
device as an NVM synapse for neuromorphic computing (Sahu et al. 2022). Both
through micromagnetic simulations and experimental demonstrations (Bläsing et al.
2018), it has been shown that the SOT-driven velocity of the domain wall in the fer-
rimagnetic device is much higher than that in the ferromagnetic device for the same
value of current density. As a result, it has been shown that the use of ferrimagnetic
domain-wall synapse device in crossbar arrays can lead to lower energy consump-
tion and higher speed for on-chip learning, when compared to its ferromagnetic
counterpart (Sahu et al. 2022).
References 79

References

Apalkov D, Dieny B, Slaughter JM (2016) Magnetoresistive random access memory. Proc IEEE
104(10):1796–1830
Berger AJ, Edwards ERJ, Nembach HT, Karis O, Weiler M, Silva TJ (2018) Determination of
the spin Hall effect and the spin diffusion length of Pt from self-consistent fitting of damping
enhancement and inverse spin-orbit torque measurements. Phys Rev B 98(2)
Bhattacharya T, Li S, Huang Y, Kang W, Zhao W, Suri M (2019) Low-power (1t1n) skyrmionic
synapses for spiking neuromorphic systems. IEEE Access 7:5034–5044
Bhowmik D, Nowakowski ME, You L, Lee O, Keating D, Wong M, Jeffrey B, Salahuddin S (2015)
Deterministic domain wall motion orthogonal to current flow due to spin orbit torque. Sci Rep
5(1):1–10
Bhowmik D, Saxena U, Dankar A, Verma A, Kaushik D, Chatterjee S, Singh U (2019) On-chip
learning for domain wall synapse based fully connected neural network. J Magn Magn Mat
489:165434
Bläsing R, Ma T, Yang SH, Garg C, Dejene FK, N’Diaye AT, Chen G, Liu K, Parkin SSP (2018)
Exchange coupling torque in ferrimagnetic Co/Gd bilayer maximized near angular momentum
compensation temperature. Nat Commun 9(1):1–8
Chakroborty I et al (2020) Resistive crossbars as approximate hardware building blocks for machine
learning: opportunities and challenges. Proc IEEE 1–35
Conte L et al (2015) Role of B diffusion in the interfacial Dzyaloshinskii-Moriya interaction in
Ta/Co20Fe60B20/MgO nanowires. Phys Rev B 91:014433
Dhull S, Misba W, Nisar A, Atulasimha J, Kaushik BK (2023) Quantized magnetic domain wall
synapse for efficient deep neural networks. TechRxiv Preprint. techrxiv.21982208.v1
Emori S, Bauer U, Ahn SM, Martinez E, Beach GSD (2013) Current-driven dynamics of chiral
ferromagnetic domain walls. Nat Mat 12(7):611–616
Emori S, Martinez E, Lee K-J, Lee H-W, Bauer U, Ahn S-M, Agrawal P, Bono DC, Beach GSD
(2014) Spin Hall torque magnetometry of Dzyaloshinskii domain walls. Phys Rev B 90(18)
Fullerton ER, Childress JR (2016) Spintronics, magnetoresistive heads, and the emergence of the
digital world. Proc IEEE 104(10):1787–1795
Goh Y et al (2021) High performance and self-rectifying Hafnia-based ferroelectric tunnel junc-
tion for neuromorphic computing and TCAM applications. In: 2021 IEEE international electron
devices meeting (IEDM), San Francisco, CA, USA, pp 17.2.1–17.2.4
Kaushik D, Sharda J, Bhowmik D (2020) Synapse cell optimization and back-propagation algorithm
implementation in a domain wall synapse based crossbar neural network for scalable on-chip
learning. Nanotechnology 31(36)
Ryu KS, Thomas L, Yang SH, Parkin S (2013) Chiral spin torque at magnetic domain walls. Nat
Nanotechnol 8(7):527-533
Kumar D et al (2023) Ultralow energy domain wall device for spin-based neuromorphic computing.
ACS Nano 17(7):6261–6274
Lee S-W, Lee K-J (2016) Emerging three-terminal magnetic memory devices. Proc IEEE
104(10):1831–1843
Lee J, Nikam RD, Kim D, Hwang H (2022) Highly scalable (30 nm) and ultra-low-energy (5fJ/pulse)
Vertical sensing ECRAM with ideal synaptic characteristics using ion-permeable graphene elec-
trodes. In: 2022 international electron devices meeting (IEDM), San Francisco, CA, USA
Leonard T et al (2022) Shape-dependent multi-weight magnetic artificial synapses for neuromorphic
computing. Adv Electron Mater 2200563
Liu L, Pai C-F, Li Y, Tseng HW, Ralph DC, Buhrman RA (2012a) Spin-torque switching with the
giant spin Hall effect of tantalum. Science 336(6081):555–558
Liu S, Xiao TP, Cui C, Incorvia JAC, Bennett CH, Marinella MJ (2021) A domain wall-magnetic
tunnel junction artificial synapse with notched geometry for accurate and efficient training of
deep neural networks. Appl Phys Lett 118(20):202405
80 4 The Ferromagnetic Domain-Wall Synapse Device

Liu L, Lee OJ, Gudmundsen TJ, Ralph DC, Buhrman RA (2012b) Current-induced switching of
perpendicularly magnetized magnetic layers using spin torque from the spin Hall effect. Phys
Rev Lett 109(9)
Liu L, Moriyama T, Ralph DC, Buhrman RA (2011) Spin-torque ferromagnetic resonance induced
by the spin Hall effect. Phys Rev Lett 106(3)
Luo Y, Peng X, Yu S (2019) MLP+NeuroSimV3.0. In: Proceedings of the international conference
on neuromorphic systems (ACM)
Miron IM et al (2010) Current-driven spin torque induced by the Rashba effect in a ferromagnetic
metal layer. Nat Mat 9:230–234
Qu D, Huang SY, Miao BF, Huang SX, Chien CL (2014) Self-consistent determination of spin Hall
angles in selected 5 d metals by thermal spin injection. Phys Rev B 89(14)
Sahu DP, Park K, Chung PH, Han J, Yoon T-S (2023) Linear and symmetric synaptic weight update
characteristics by controlling filament geometry in oxide/suboxide HfOx bilayer memristive
device for neuromorphic computing. Sci Rep 13(1)
Sahu U, Sisodia N, Sharda J, Muduli PK, Bhowmik D (2022) Ferrimagnetic synapse devices for
fast and energy-efficient on-chip learning on an analog-hardware neural network. IEEE Trans
Electron Dev 69(4):1713–1720
Sampaio J, Cros V, Rohart S, Thiaville A, Fert A (2013) Nucleation, stability and current-induced
motion of isolated magnetic skyrmions in nanostructures. Nat Nanotechnol 8(11):839–844
Saxena U, Kaushik D, Bansal M, Sahu U, Bhowmik D (2018) Low energy implementation of
feed-forward neural network with back-propagation algorithm using a spin orbit torque driven
skyrmionic device. IEEE Trans Magn 54 (11)
Sebastian A et al (2018) Tutorial: brain-inspired computing using phase-change memory devices.
J Appl Phys 124:111101
Sengupta A, Shim Y, Roy K (2016) Proposal for an all-spin artificial neural network: emulating
neural and synaptic functionalities through domain wall motion in ferromagnets. IEEE Trans
Biomed Circuits Syst 10(6)
Singh U, Sahu U, Sreedevi I, Bhowmik D (2020) Comparing domain wall synapse with other
non volatile memory devices for on-chip learning in analog hardware neural network. AIP Adv
10(2):025111
Song KM et al (2020) Skyrmion-based artificial synapses for neuromorphic computing. Nature
Electron 3:148–155
Sun X, Yu S (2019) Impact of non-ideal characteristics of resistive synaptic devices on implementing
convolutional neural networks. IEEE J Emerg Sel Top Circuits Syst 9(3):570–579
Tsai H et al (2018) Recent progress in analog memory-based accelerators for deep learning. J Phys
D: Appl Phys 51:283001
Yadav RS, Gupta P, Holla A, Ali Khan KI, Muduli PK, Bhowmik D (2023a) Demonstration of synap-
tic behavior in a heavy-metal-ferromagnetic-metal-oxide-heterostructure-based spintronic device
for on-chip learning in crossbar-array-based neural networks. ACS Appl Electr Mat 5(1):484–497.
American Chemical Society (ACS) )
Yadav RS, Sadashiva A, Holla A, Muduli PK, Bhowmik D (2023b) Impact of edge defects on
the synaptic characteristic of a ferromagnetic domain-wall device and on on-chip learning. Neur
Comput Eng 3(3)
Zhang Q et al (2022) Perpendicular magnetization switching driven by spin-orbit torque for artificial
synapses in epitaxial Pt-based multilayers. Adv Electron Mater 2200845
Zhang S et al (2019) A Spin-orbit-torque memristive device. Adv Electron Mater 1800782
Chapter 5
Design of Artificial Neural Networks
(ANN) with Domain-Wall Synapse
Devices

5.1 Thresholding-Based Modification to the Neural

Network Training Algorithm

The equations related to inference (forward pass) and training of a fully connected
neural network (FCNN) with a hidden layer have been provided in Chap. 3. When
the FCNN is implemented on a conventional computer like a CPU or a GPU, which
is the common practice in machine-learning-and-AI-related fields. But when such
an algorithm is implemented on a neuromorphic crossbar array, constraints imposed
by the NVM synapse devices in the array need to be kept in mind. The design of
the peripheral circuit to bring about the weight updates in the synapse devices also
needs to be kept simple. As a result, modifications need to be made to the FCNN
algorithm before implementing it on the crossbar array. One such modification, as
reported by Kaushik et al. (2020), is adding thresholding functions to the equations
to limit to the synaptic-weight update to .+W , .−W or 0 per iteration, where .W is a
fixed positive number. We discuss this modification here.
With respect to the FCNN in Fig. 3.2 of Chap. 3 (also redrawn here: Fig. 5.1), the
following are the equations for weight update of the synapses at any iteration when
no thresholding is applied:

ηλ
Δvn, p =
. (Yn − yn )(1 − yn2 ))z p = Δvn z p , (5.1)
2
where .Δvn = is the common part of weight update generated at each node .n of the
output layer: (. ηλ
2
(Yn − yn )(1 − yn2 ))

λ
Δw p,m = (Σn=1
.
n=N
Δvn vn, p ) (1 − z 2p ))xm (5.2)
2

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 81
D. Bhowmik, Spintronics-Based Neuromorphic Computing, Springer Tracts in Electrical
and Electronics Engineering, https://doi.org/10.1007/978-981-97-4445-9_5
82 5 Design of Artificial Neural Networks (ANN) …

Fig. 5.1 Schematic of a fully connected neural network (FCNN) with one hidden layer. (Reprinted
with permission from: Sahu U, Sisodia N, Sharda J, Muduli PK, Bhowmik D (2022) Ferrimag-
netic synapse devices for fast and energy-efficient on-chip learning on an analog-hardware neural
network. IEEE Trans Electron Dev 69(4):1713–1720, 0018-9383 .© 2022 IEEE (2022) Sahu et al.
2022)

In order to limit these weight updates to.+W ,.−W , or 0, the following thresholding
functions are used:

. Q 1 (X ) = θ1 ∀ X > q1 ;
Q 1 (X ) = 0 ∀ − q1 ≤ X ≤ q1 ;
Q 1 (X ) = −θ1 ∀ X < −q1
(5.3)

and

. Q 2 (Y ) = θ2 ∀ Y > q2 ;
Q 2 (Y ) = 0 ∀ − q2 ≤ Y ≤ q2 ;
Q 2 (Y ) = −θ2 ∀ Y < −q2 ,
(5.4)

where.q1 ,.q2 ,.θ1 , and.θ2 : hyper-parameters. Making use of the thresholding functions,
the equations for synaptic-weight update are now given as follows:
( )
ηλ
Δvn, p = Q 2
. (Yn − yn )(1 − yn ) Q 1 (z p ) = Q 2 (Δvn )Q 1 (z p ),
2
(5.5)
2
5.1 Thresholding-Based Modification to the Neural Network Training Algorithm 83

and
λ
Δw p,m = Q 2 ((Σn=1
.
n=N
Δvn vn, p ) (1 − z 2p ))Q 1 (xm ), (5.6)
2
where . Q 1 and . Q 2 : thresholding functions mentioned above.
Based on the equations above, at any iteration during the training process, update
of each element of the synaptic-weight matrix can take one of only three possible
values:.θ1 θ2 , or.−θ1 θ2 , or 0. We can call it:.+W ,.−W or 0. This makes implementation
on analog hardware much more easy since the multiplier associated with each synapse
device can be much simpler that way. We see this in the next section.
Since the weight update at each synapse is now limited to .+W , .−W , or 0 per iter-
ation, the weight update curves (synaptic characteristics) of the spintronic domain-
wall device reported in Chap. 4 become particularly relevant. Thereby, assuming the
synaptic characteristic of the device is linear and symmetric, identical programming
pulses for long-term potentiation (LTP) leading to domain-wall motion in one direc-
tion and fixed increase in weight (.W ) per pulse, and identical programming pulses
for long-term depression (LTD) lead to a fixed decrease in weight (.W ) per pulse.
However, as explained in Chap. 4, the LTP/LTD characteristic of a domain-wall
device (or for any NVM synapse device for that) is not perfectly linear and symmetric.
Non-linearity and asymmetry of LTP and LTD are known to reduce the classification
accuracy of the neural network, when it is implemented on the crossbar array endowed
with the on-chip learning facility that makes use of these non-ideal LTP and LTD
characteristics of the synapse device at every iteration of the training.
In Chap. 4, we have discussed a report by Yadav et al. (2023) where experi-
mentally obtained LTP and LTD of a domain-wall device have been reported, and
using curve-fitting following the method provided with the NeuroSim simulator (Luo
et al. 2019), non-linearity coefficients .α P and .α D are obtained: Fig. 4.7 of Chap. 4.
Figure 5.2 (taken from the same report by Yadav et al. 2023) shows a comparison
of classification accuracy numbers for training of FCNN with a hidden layer for the
following four cases:
Case 1: Each synaptic weight is of 32-bit precision; weight update is linear and
symmetric. This is the case when the neural network is trained on a conventional
computer (green plot in Fig. 5.2).
Case 2: Each synaptic weight is of 5-bit precision (keeping in mind that the domain-
wall device has about 30 stable and distinguishable conductance/Hall resistance
levels, as reported by Yadav et al. 2023), but the weight update (LTP/LTD) is linear
and symmetric (black plot in Fig. 5.2).
Case 3: Each synaptic weight is of 5-bit precision, and weight update is also
as per the reported LTP and LTD by Yadav et al. (2023), i.e., .α P = 3.875 and
.α D = −3.84 (red plot in Fig. 5.2). No cycle-to-cycle variation is included.
Case 4: Experimentally obtained LTP and LTD from Yadav et al. (2023), while
including cycle-to-cycle variations (blue plot in Fig. 5.2). The same LTP and LTD
experiments are repeated on the same device for 20 cycles, to obtain cycle-to-cycle
variations.
84 5 Design of Artificial Neural Networks (ANN) …

Fig. 5.2 Train (a) and test (b) accuracy numbers on the MNIST data set when an ideal 64-bit
device is used per synapse cell (light green plots), two 5-bit ideal devices (with completely linear
and symmetric LTP and LTD characteristics) are used per synapse cell in the VMM operation of
the designed FCNN/MLP of Fig. 5.1 (black plots), and when two 5-bit spintronic devices with
experimentally obtained LTP and LTD characteristics (reported by Yadav et al. 2023) are used per
synapse cell (red plots: without cycle-to-cycle variation, blue plots: after including cycle-to-cycle
variation of Fig. 5.1c). Other than the ideal-64-bit synapse case, for all the other three cases, the
thresholding-based algorithm discussed in the text is used. (Reprinted with permission from: Yadav
RS, Gupta P, Holla A, Ali Khan KI, Muduli PK, Bhowmik D (2023) Demonstration of synaptic
behavior in a heavy-metal-ferromagnetic-metal-oxide-heterostructure-based spintronic device for
on-chip learning in crossbar-array-based neural networks. ACS Appl Electr Mater 5(1):484–497,
Copyright 2023 American Chemical Society (ACS) Yadav et al. 2023)

MNIST data set of handwritten digits has been used for the purpose. We observe
that compared to Case 1 (ideal case), the drop in classification accuracy is very low for
Case 2, which means limiting the bit resolution of the synaptic weight doesn’t affect
the accuracy much if linearity and asymmetry are not affected. However, when the
linearity and asymmetry are affected, accuracy does drop significantly irrespective
of cycle-to-cycle variations, as evident from the plots for Case 3 and Case 4. This
makes the incorporation of grooves or notches or material defects in the domain-wall
device to pin the domain wall and improve the device’s non-linearity and asymmetry
of LTP and LTD (as discussed in Chap. 3) all the more important.

5.2 Design of the Synapse Cell

In this section, we discuss how the multiplication between thresholded version of

common part of weight update and thresholded version of input can be carried out
for every synapse device to implement Eq. 5.5. For this purpose, two transistors need
to be associated with the domain-wall device at each intersection of the horizontal
and vertical bars. The two transistors together act as a multiplier between two bits
(if the thresholding functions are not used in the network training algorithm, then
an analog multiplier needed to be designed here as opposed to a bit multiplier, and
this would have demanded much more than two transistors). The two transistors then
5.2 Design of the Synapse Cell 85

feed the necessary programming current pulse to the domain-wall device to cause
weight update by a bit (.+W , .−W , 0). The two transistors and the domain-wall device
together form a synapse cell, as shown in Fig. 5.4. At each intersection point of the
horizontal and vertical bars in the crossbar array, there needs to be a synapse cell. More
details on the design and working principle of the synapse cell are presented below.
The multiplier unit in this designed synapse cell, as reported by Kaushik et al.
(2020), contains two transistors: one NMOS (T1) and the other PMOS (T2). When
. Q 1 (z p ) > 0 is applied to the circuit, T1 turns on, and T2 turns off. As a result, the volt-
age that injects the programming/“write” current into the domain-wall synapse device
is given by: .V1 = Q 2 (Δvn ). Since this voltage is positive and quantized/thresholded,
it means quantized/thresholded “write” current of positive polarity will flow through
the heavy metal layer of the synapse device (current into the device). This results in
a positive weight update (also called long-term potentiation or LTP) in the synapse.
When . Q 1 (z p ) < 0 instead, T1 turns off, and T2 turns on. Hence, the voltage that
drives the “write” current into the synapse device is: .V1 = −Q 2 (Δvn ). So quan-
tized/thresholded “write” current of negative polarity flows through the heavy metal
layer of the domain-wall device (current flows out of the device), leading to negative
weight update (also called long-term depression or LTD) in the synapse device. And
when . Q 1 (z p ) = 0, both the transistors (T1 and T2) are turned off, and .V1 is 0, which
means no “write” current will flow; thus, there is no weight update.
The above description can be summarized as follows:
⎧
⎪
⎨ Q 1 (z p ) > 0: T1 turns on, T2 off, V1 = Q 2 (Δvn ), LTP.
. Q 1 (z p ) = 0: T1 and T2 off, V1 = 0; no weight update.
⎪
⎩
Q 1 (z p ) < 0: T1 turns off, T2 on, V1 = −Q 2 (Δvn ), LTD.
LTP/LTDat each synapse of the second layer of the FCNN is given by Eq. 5.5.
Based on the above three scenarios (when. Q 1 (z p ) can be positive or negative or zero),
there are nine possible cases for weight update (LTP/LTD), three corresponding to
each scenario (where . Q 2 (.Δvn ) can be positive or negative or zero).
⎧ Hence, .Δvn, p now only takes the value .±θ1 θ2 (.= ±ΔW ). .Δ .wn,m =
⎪
⎪ Case 1 : ΔW, Q 1 (z p ) > 0 & Q 2 (Δvn ) > 0
⎪
⎪
⎪
⎪ Case 2 : 0, Q 1 (z p ) = 0 & Q 2 (Δvn ) > 0
⎪
⎪
⎪
⎪ Case 3 : − ΔW, Q 1 (z p ) < 0 & Q 2 (Δvn ) > 0
⎪
⎪
⎨ Case 4 : 0, Q 1 (z p ) > 0 & Q 2 (Δvn ) = 0
. Case 5 : 0, Q 1 (z p ) = 0 & Q 2 (Δvn ) = 0
⎪
⎪
⎪
⎪ Case 6 : 0, Q 1 (z p ) < 0 & Q 2 (Δvn ) = 0
⎪
⎪
⎪
⎪ Case 7 : − ΔW, Q 1 (z p ) > 0 & Q 2 (Δvn ) < 0
⎪
⎪
⎪
⎪ Case 8 : 0, Q 1 (z p ) = 0 & Q 2 (Δvn ) < 0
⎩
Case 9 : ΔW, Q 1 (z p ) < 0 & Q 2 (Δvn ) < 0
SPICE simulation of the synapse cell including a domain-wall synapse device and
two transistors at 65 nm technology node (Fig. 5.4) has been reported by Kaushik et al.
(2020). The linear and symmetric conductance response of the domain-wall synapse
device, obtained through micromagnetic simulation without allowing for delay
between pulses (no stray domain-wall motion hence) and shown in Fig. 5.3 (very
similar to Fig. 4.3 of Chap. 4), has been used for the purpose for the sake of simplicity.
86 5 Design of Artificial Neural Networks (ANN) …

Fig. 5.3 a Schematic of the domain-wall synapse device b Conductance response of “read” path
versus in-plane “write” current pulse of the domain-wall synapse device, which is its LTP/LTD
characteristic. Conductance increases or decreases in fixed steps (.ΔG) due to the application of
fixed-magnitude current pulses of opposite polarities. Delay between pulses is ignored here; so
there’s no stray domain-wall motion between pulses and the LTP and LTD characteristics are linear
and symmetric. (Reprinted with permission from: Kaushik D, Sharda J, Bhowmik D (2020) Synapse
cell optimization and back-propagation algorithm implementation in a domain-wall-synapse-based
crossbar neural network for scalable on-chip learning. Nanotechnology 31(36) Kaushik et al. 2020)

A Verilog A module with three terminals and the conductance between two terminals
(“read” path) being controlled by “write” current through the third terminal, based on
the exact same synaptic characteristic as shown in Fig. 5.3, is inserted in the SPICE
simulation of the synapse cell.
For Cases 1 and 9 above, .V1 is positive based on the operation of T1 and T2 as
discussed above. Positive .V1 drives “write” current of magnitude 100 .µA and of
positive polarity (from left to right) in the domain-wall synapse (Fig. 5.4). Hence,
conductance increases by .ΔG (Fig. 5.5) and weight increases by .ΔW , as observed
from SPICE simulations of the synapse cell. Case 1 has been identified in the figure.
Positive values of . Q 1 (z p ) and . Q 2 (Δvn ) lead to positive “write” current and positive
conductance update (LTP), as expected.
Similarly, for Cases 3 and 7, .V1 is negative based on the operation of T1 and T2
as discussed above. Negative .V1 drives “write” current of magnitude 100 .µA and
of negative polarity in the domain-wall synapse (Fig. 5.4). As evident from SPICE
5.2 Design of the Synapse Cell 87

Fig. 5.4 Schematic of optimized synapse cell including a two-transistor-based multiplier which
multiplies two bits as sown and a domain-wall-based synaptic device into which the transistors feed
“write” current for conductance/weight update. (Reprinted with permission from: Kaushik D,
Sharda J, Bhowmik D (2020) Synapse cell optimization and back-propagation algorithm imple-
mentation in a domain-wall-synapse-based crossbar neural network for scalable on-chip learning.
Nanotechnology 31(36) Kaushik et al. 2020)

Fig. 5.5 SPICE simulation of the synapse cell in Fig. 5.4 is carried out here. Inputs to the cell
. Q 1 (z p )
and . Q 2 (Δvn ) are plotted as a function of time. “Write” current flowing into the domain-
wall synapse device in the cell and the corresponding change in conductance in its “read” path are
also plotted as functions of time. (Reprinted with permission from: Kaushik D, Sharda J, Bhowmik
D (2020) Synapse cell optimization and back-propagation algorithm implementation in a domain-
wall-synapse-based crossbar neural network for scalable on-chip learning. Nanotechnology 31(36)
Kaushik et al. 2020)
88 5 Design of Artificial Neural Networks (ANN) …

simulations in Fig. 5.5 (Case 7 has been identified), this leads to conductance decrease
by .ΔG and hence weight decrease by .ΔW (LTD), as expected.
For Cases 2, 4, 5, 6, and 8, .V1 = 0. Hence, no current flows through the domain-
wall synapse and its conductance/weight is not updated, as also seen from SPICE
simulations in Fig. 5.5 (Cases 4 and 5 have been identified).

5.3 Design of Crossbar Arrays to Implement the

Back-Propagation Algorithm

Having shown the design of each single synapse cell in the crossbar array, we now
focus our attention on the system-level design of crossbar arrays such that equations
for the thresholded version of the back-propagation algorithm (Eqs. 5.5 and 5.6) can
be implemented to carry out appropriate weight updates for all the domain-wall
synapse devices in the crossbar array. While carrying the design, which has been
reported by Kaushik et al. (2020), one needs to keep in mind that we would like to
update the weights of all the synapses in a crossbar array simultaneously/at once to
get the maximum benefit from the neuromorphic system.
For the FCNN with a single hidden layer of Fig. 5.1, the corresponding imple-
mentation through crossbar arrays (including implementation of weight update/back-
propagation for on-chip learning) involves three crossbar arrays and is shown in
Fig. 5.6. The equations for forward computation are provided in Chap. 3 and are
repeated here, to link them with their crossbar-array implementation.
The equations for forward computation at the first stage of the FCNN of Fig. 5.1 are

z¯ = w p,1 x1 + w p,2 x2 · · · · · · + w p,m xm · · · +w p,M x M + w p,0

. p

2
z p = f (z¯p ) = −1
1 + (e−λz¯p )
(5.7)

The vector-matrix multiplication (VMM) operation above is implemented through

crossbar array I in Fig. 5.6, using a combination of Ohm’s Law and Kirchoff’s Current
Law as explained in previous chapters. Additionally, operational-amplifier-based
circuits can be designed for the non-linear activation functions used above.
The equations for forward computation at the second stage of the FCNN of
Fig. 5.1 are

. ny¯ = vn,1 z 1 + vn,2 z 2 · · · · · · + vn, p z p · · · + vn,P z P + vn,0 ;

2
yn = f ( y¯n ) = −1
1 + (e−λ y¯n )
(5.8)
5.3 Design of Crossbar Arrays to Implement the Back-Propagation Algorithm 89

Fig. 5.6 Schematic of the domain-wall-synapse-device-based crossbar array system to implement

both forward computation and thresholded weight update (back-propagation) equations with respect
to the FCNN with one hidden layer shown in Fig. 5.1. (Reprinted with permission from: Kaushik
D, Sharda J, Bhowmik D (2020) Synapse cell optimization and back-propagation algorithm imple-
mentation in a domain-wall-synapse-based crossbar neural network for scalable on-chip learning.
Nanotechnology 31(36) Kaushik et al. 2020)

This VMM operation is carried out in crossbar II, with outputs of crossbar I (.z 1 ,
z …….z P ) being feds as inputs to crossbar II for the second VMM operation. LTP and
. 2
LTD of the synapses in crossbar II, after thresholding, are given by Eq. 5.5 mentioned
above. To implement this, the common part of weight update is calculated and thresh-
olded at the output nodes of crossbar II: . Q 2 ( ηλ
2
(Yn − yn )(1 − yn2 )), or . Q 2 (Δvn ).
For thresholding, a dedicated amplifier-based circuit (more details in Kaushik
et al. 2020) can be implemented at each output node, and also at each input node
to threshold the input (Kaushik et al. 2020). The advantage of this scheme is that
such thresholding circuit needs to be present only at each input node and each output
node as per this design, and not at every synapse cell. This makes the total number
of thresholding circuits required equal to the sum of the number of input nodes and
that of output nodes, as opposed to the product of the number of input nodes and that
of output nodes. Thus, this design doesn’t require the implementation of too many
thresholding circuits.
To carry out the multiplication in Eq. 5.5, . Q 2 ( ηλ
2
(Yn − yn )(1 − yn2 )), or . Q 2 (Δvn ),
is applied at the other set of vertical bars going into the synapse cells (not the ones
through which currents flow corresponding to forward computation. VMM), as
shown in Fig. 5.6. As also shown in Fig. 5.6, . Q 1 (z p ) is applied at each horizontal
90 5 Design of Artificial Neural Networks (ANN) …

bar labelled . p. The multiplication between . Q 1 (z p ) at the horizontal bar labelled . p

and . Q 2 (Δvn ) at the alternate vertical bar labelled .n is carried out in the synapse
cell connecting input node . p with output node .n using two transistors, as described
in the previous section. Then the domain-wall synapse device in the synapse cell is
updated as per Eq. 5.5.
Implementing parallel weight updates on the synapses in crossbar I is much more
involved because as per Eq. 5.6, . Q 2 (Σn=1 n=N
Δvn vn, p ) needs to be calculated which
depends upon the synaptic weights of crossbar II at that iteration. Following the
design reported by Kaushik et al. (2020), the common part of weight update at each
output node .n (.Δvn ), before thresholding, is applied as an input to another crossbar
(crossbar III in Fig. 5.6) with . N input nodes and . P output nodes, in order to calculate
.Σn=1 Δvn vn, p . Crossbar III’s synapses have same weight values as crossbar II’s
n=N

synapses, and both weight values are updated simultaneously after every iteration.
At a synapse in crossbar III, connecting output node . p with input node .n, current
through the “read” path of the domain-wall synapse is equal to the product of input
.Δvn and weight .vn, p (Ohm’s Law). Currents for all the synapses connecting node . p

with all the input nodes (from .n = 1 to . N ) add up following Kirchoff’s Current Law
(KCL) (shown by.Σ in Fig. 5.6), similar to the forward VMM computation in crossbar
I and II. Thus, at each node . p, sum of “read” currents is equal to .Σn=1 n=N
Δvn vn, p .
That current is then sent back to crossbar I, as shown in Fig. 5.6. Then multiplication
between the thresholded version of.Σn=1 n=N
Δvn vn, p and thresholded input to crossbar I
. x m takes place in each synapse cell of crossbar I, just like in the case of multiplications
in synapse cells of crossbar II as described earlier. Thus, Eq. 5.6 can be implemented
in hardware, and LTP/LTD of the synaptic weights of crossbar array I can be carried
out in parallel. SPICE simulations of this entire three-crossbar system (Fig. 5.6),
implementing on-chip learning of the FCNN with one hidden layer of Fig. 5.1 and
resulting in high classification accuracy for the Fisher’s Iris data set of flowers, have
been reported by Kaushik et al. (2020).
To conclude, it is to be noted that all the design choices discussed in this chapter
(thresholding algorithm, synapse cell design, and crossbar array design) should be
applicable not only to the domain-wall non-volatile memory (NVM) synapse device
but also to any three-terminal synapse device, where the “read” conductance between
two terminals is modulated by “write” current flowing from the third terminal. To
keep the circuit design simple, we have used identical pulses for positive weight
update (LTP) and identical pulses for negative weight update (LTD). So, as long
as this can be achieved in any three-terminal NVM device with decently linear and
symmetric LTP and LTD characteristics, the design and analysis shown here will be
applicable to that device.
References 91

References

Kaushik D, Sharda J, Bhowmik D (2020) Synapse cell optimization and back-propagation algorithm
implementation in a domain wall synapse based crossbar Neural Network for scalable on-chip
learning. Nanotechnology 31(36)
Luo Y, Peng X, Yu S (2019) MLP+NeuroSimV3.0. In: Proceedings of the international conference
on neuromorphic systems (ACM)
Sahu U, Sisodia N, Sharda J, Muduli PK, Bhowmik D (2022) Ferrimagnetic synapse devices for
fast and energy-efficient on-chip learning on an analog-hardware neural network. IEEE Trans
Electron Dev 69(4):1713–1720
Yadav RS, Gupta P, Holla A, Ali Khan KI, Muduli PK, Bhowmik D (2023) Demonstration of
synaptic behavior in a heavy-metal-ferromagnetic-metal-oxide-heterostructure-based spintronic
device for on-chip learning in crossbar-array-based neural networks. ACS Appl Electr Mater
5(1):484–497. American Chemical Society (ACS)
Chapter 6
Design of Spiking Neural Networks
(SNN) with Domain-Wall Devices

6.1 Design and Working of the STDP-Enabling Circuit

In Chap. 3, spike-time-dependent plasticity (STDP), important for training of SNNs,

has been introduced and necessary equations have been provided (Bi and Poo 1998).
Here, we first repeat those basic equations/rules for STDP, along with some basic
rules/equations for the domain-wall device discussed in the previous chapters, and
then go on to discuss how an analog circuit, in combination with the domain-wall
device, can be designed to implement these equations.
Not only for domain-wall devices but also for any non-volatile memory (NVM)
synapse device, change in conductance (.ΔG) of the synaptic device is related to the
change in weight that the synapse device stores (.Δw) as follows:
G max − G min
ΔG =
. Δw (6.1)
wmax − wmin

For a wide range of current density flowing through the domain-wall device,
both micromagnetic simulations and actual experiments show that the domain-wall
velocity is proportional to the current density (Sahu et al. 2019; Martinez et al.
2014; Emori et al. 2014). Hence magnitude of the “write”/programming current
pulse (. Iwrite ) needed to bring about a certain change in conductance (.ΔG) is given
as follows:
∂ Iwrite
. Iwrite = ΔG, (6.2)
∂G

where. ∂ I∂G
write
is the slope of a straight line that fits the conductance versus write current
characteristic of the domain-wall device, which can be either obtained experimentally
or through micromagnetic simulations.
As discussed in Chap. 3 and shown in Fig. 6.1 taken from the same chapter, in an
SNN, a neuron of the input layer (pre-neuron) is connected to a neuron of the output

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 93
D. Bhowmik, Spintronics-Based Neuromorphic Computing, Springer Tracts in Electrical
and Electronics Engineering, https://doi.org/10.1007/978-981-97-4445-9_6
94 6 Design of Spiking Neural Networks (SNN) with Domain-Wall Devices

Fig. 6.1 STDP

characteristic of a synapse
connecting a post-neuron
and a pre-neuron where the
post-neuron spikes at time
.t post and the pre-neuron
spikes at time .t pr e

layer(post-neuron) through a non-volatile synapse and the weight of the synapse is

updated by the STDP rule as follows (Bi and Poo 1998):
t post −t pr e
Δw = Γ e−( τ )
i f t post > t pr e ;
. t −t , (6.3)
−( pr e τ post )
Δw = −Γ e i f t post < t pr e

where.Δw is the change in weight of the synapse,.t pr e is the time when the pre-neuron
spikes, .t post is the time when the post-neuron spikes, .Δt is the difference between
.t post and .t pr e , .Γ is a constant of proportionality, and .τ is the STDP time constant for

the synapse.
Thus, as per this STDP rule, when the pre-neuron spikes before the post-neuron
does, the weight of the synapse goes up (synapse gets strengthened, or potentiation).
When the pre-neuron spikes after the post-neuron, the synaptic weight goes down
(depression). The magnitude of the weight change is inversely proportional to the
timing difference between the two spikes. Hence, this is essentially a positive feed-
back mechanism because the pre-neuron spiking promotes the post-neuron to spike
as long as the weight is positive, and if the post-neuron spikes after the pre-neuron,
the weight goes up further so that, next time, the post-neuron spikes immediately
after the pre-neuron.
In order to enable the non-volatile domain-wall synapse device with STDP prop-
erty, as per Eqs. 6.1, 6.2, and 6.3, for a given spike at the pre-neuron and another
spike at the post-neuron, “write”/programming current needs to be applied on the
device as follows in order to update the weight and train the SNN:
t post −t pr e
Iwrite = I0 e−( τ )
i f t post > t pr e
. t −t , (6.4)
−( pr e τ post )
Iwrite = −I0 e i f t post < t pr e

where . I0 is proportional to .Γ in Eq. 6.3. In order to achieve this, the domain-wall

synapse device needs to be integrated with a transistor-based circuit operating in
a sub-threshold regime, which injects appropriate “write” current pulse into the
domain-wall device, as shown in Figs. 6.2, 6.3, and 6.4 (Sahu et al. 2019; Sengupta
et al. 2016).
6.1 Design and Working of the STDP-Enabling Circuit 95

Fig. 6.2 Domain-wall (DW) device along with transistor-based circuit which together emulates
the STDP property of the synapse. (Reprinted with permission from: Sahu U, Pandey A, Goyal
K, Bhowmik D (2019) Spike-time-dependent plasticity (STDP)-enabled learning in spiking neural
networks using domain-wall-based synapses and neurons. AIP Adv 9(12) Sahu et al. 2019)

. Iwrite in Eq. 6.4 can be considered a sum of two components: . Iwrite = Iwrite,1 +
Iwrite,2 , where:
t post −t pr e
−( )
I f t post > t pr e , Iwrite,1 = I0 e τ1
; Iwrite,2 = 0
. t pr e −t post (6.5)
−( )
I f t post < t pr e , Iwrite,1 = 0; Iwrite,2 = −I0 e τ2

In the circuit of Fig. 6.2, drain current through transistor T4 corresponds to. Iwrite,1 ,
and drain current through T8 corresponds to . Iwrite,2 . Based on SPICE simulations
of the circuit in Fig. 6.2 as reported by Sahu et al. (2019) (similar circuit design
has also been reported by Sengupta et al. (2016, 2017), . Iwrite,1 , and . Iwrite,2 are
plotted as functions of time for spiking pattern 1 (Fig. 6.3) and spiking pattern 2
(Fig. 6.4).
For spiking pattern 1, pre-neuron spikes once (.t pr e ) followed by several post-
neuron spikes (.t post ) (Fig. 6.3a, b). Hence, .t post > t pr e here. The dominant current
that flows through the domain-wall device is . Iwrite,1 in this case, which is generated
by transistor T3 of the circuit in Fig. 6.2. T3 operates in the sub-threshold regime,
and so the current flowing through it is an exponential function of gate voltage. Since
the capacitor C1 starts charging as soon as the pre-neuron spikes at .t pr e (Fig. 6.3c),
96 6 Design of Spiking Neural Networks (SNN) with Domain-Wall Devices

Fig. 6.3 Voltage and current of different components of the circuit in Fig. 6.2 for spiking pattern
1: a Gate voltage of T2 versus time showing spiking pattern of pre-neuron b Gate voltage of T4
showing spiking pattern of post-neuron c Gate voltage of T3 and T7 d Drain current through T4
(. Iwrite,1 ) and T8 (. Iwrite,2 ). (Reprinted with permission from: Sahu U, Pandey A, Goyal K, Bhowmik
D (2019) Spike-time-dependent plasticity (STDP)-enabled learning in spiking neural networks using
domain-wall-based synapses and neurons. AIP Adv 9(12) Sahu et al. 2019)

voltage across the capacitor C1 and hence the gate voltage of transistor T3 exponen-
tially decays with time starting from .t pr e (Fig. 6.3d). This leads to implementation of
the component of the STDP rule in Eqs. 6.3 and 6.4 corresponding to the post-neuron
spiking after the pre-neuron (positive weight update or potentiation).
Similarly, for spiking pattern 2, post-neuron spikes once (Fig. 6.4b) and pre-neuron
spikes several times after that (Fig. 6.4a). Hence,.t post < t pr e here. In this case, capac-
itor C2 discharges with time starting from .t post , gate voltage of T7 hence changes
linearly with time, and since T7 operates in the sub-threshold regime, magnitude
of the current flowing through T7 which dominates the “write” current flowing into
the domain-wall device (. Iwrite,2 ) drops exponentially with time starting from .t post .
This leads to implementation of the component of the STDP rule in Eqs. 6.3 and
6.4 corresponding to the pre-neuron spiking after the post-neuron (negative weight
update or depression).
6.2 Design and Working of the Domain-Wall-Device-Based LIF Neuron 97

Fig. 6.4 Voltage and current of different components of the circuit in Fig. 6.2 for spiking pattern
2: a Gate voltage of T8 versus time showing spiking pattern of pre-neuron b Gate voltage of T6
showing spiking pattern of post-neuron c Gate voltage of T3 and T7 d Drain current through T4
(. Iwrite,1 ) and T8 (. Iwrite,2 ). (Reprinted with permission from: Sahu U, Pandey A, Goyal K, Bhowmik
D (2019) Spike-time-dependent plasticity (STDP)-enabled learning in spiking neural networks using
domain-wall-based synapses and neurons. AIP Adv 9(12) Sahu et al. 2019)

6.2 Design and Working of the Domain-Wall-Device-Based

LIF Neuron

The Leaky Integrate Fire (LIF) property of a neuron in an SNN has been introduced
briefly in Chap. 3 (Dayan and Abbott 2005). We provide the basic equation again
here, and then discuss how the domain-wall device discussed in earlier chapters can
have that functionality. In the LIF model of neuron (Dayan and Abbott 2005; Diehl
and Cook 2015), as discussed in Chap. 3, the neuron’s membrane potential .v(t) is
governed by the following equation:

∂v(t)
.C = −G L (v(t) − E L ) + I (t), (6.6)
∂t

where . I (t): input current to the neuron, .G L : membrane conductance, and . E L : resting
potential of neuron. Once .v(t) reaches the threshold potential (.Vth ), the neuron
generates a spike and .v(t) drops to . E L .
98 6 Design of Spiking Neural Networks (SNN) with Domain-Wall Devices

Fig. 6.5 The design of the domain-wall (DW) device, in association with a transistor-based circuit,
which together act as a LIF neuron. (Reprinted with permission from: Sahu U, Pandey A, Goyal
K, Bhowmik D (2019) Spike-time-dependent plasticity (STDP)-enabled learning in spiking neural
networks using domain-wall-based synapses and neurons. AIP Adv 9(12) Sahu et al. 2019)

Fig. 6.6 Time gap between consecutive spikes as a function of the input current into the neuron,
as obtained by solving the equation for the LIF model of the neuron as well by numerically solving
micromagnetics-based domain-wall (DW) dynamics and SPICE-based circuit dynamics of the neu-
ron. (Reprinted with permission from: Sahu U, Pandey A, Goyal K, Bhowmik D (2019) Spike-time-
dependent plasticity (STDP)-enabled learning in spiking neural networks using domain-wall-based
synapses and neurons. AIP Adv 9(12) Sahu et al. 2019)

In Fig. 6.6, the time gap between two consecutive spikes generated by the neuron
is plotted as a function of input dc current to the neuron after solving Eq. 6.6 for a
set of LIF parameter values as reported by Sahu et al. (2019).
The DW neuron, integrated with a transistor-based circuit that we design in
Fig. 6.5, satisfies the desired LIF characteristic described in Fig. 6.6. The working
principle of the neuron unit is as follows: “write” current moves the domain wall from
one end of the device to another, much like in the domain-wall synapse device. How-
ever, the magnetic tunnel junction (MTJ) is located only at the other end of the device
here, unlike over the entire device as in the case with the domain-wall synapse device.
In this case, when the domain wall reaches the other end of the device, tunnelling
6.3 Performance of SNN with Domain-Wall-Device-Based Neurons and Synapses 99

magneto-resistance (TMR) of the MTJ changes abruptly and a spike is generated

(Sengupta and Roy 2017). Thus, the time period between two consecutive spikes is
equal to the time taken by the domain wall to cover the length of the device (Fig. 6.5),
which is in turn integral of domain-wall velocity over time. And since in this device,
for a wide range of current density, the domain-wall velocity is linearly proportional
to the current density, effectively the time gap between two spikes is equal to the
integral of current over time, with some constant of proportionality, which is same
as the LIF model of the neuron (Eq. 6.6).
The precise spike-generation mechanism of the circuit is the following. Once the
domain wall reaches the other end, TMR of the MTJ increases due to switching of
moment in the free layer. As a result, gate voltage of the transistor T1 in Fig. 6.5
increases, turning it on. This gate voltage is the output of the neuron (.Vout ) if it
is a post-neuron since the spike required for a post-neuron is positive (Figs. 6.3b
and 6.4b). For a pre-neuron, another two-transistor-based standard inverter circuit is
connected to the gate voltage of this transistor T1 (Fig. 6.5). The pMOS transistor T2
and nMOS transistor T3 together form the inverter circuit. The output of the inverter
circuit shows a negative spike when the domain wall in the neuron device reaches the
MTJ, as required for a pre-neuron (Figs. 6.3a and 6.4a). The ON current of transistor
T1 flows in the reverse direction of input current in the DW neuron and moves DW to
its initial position. This is equivalent to .v(t) in the LIF model of the neuron dropping
to . E L after a spike: the reset operation of the neuron.

6.3 Performance of SNN with Domain-Wall-Device-Based

Neurons and Synapses

Using the above micromagnetics-SPICE-combined models of domain-wall-device-

based neurons and synapses, Upasana Sahu et al. have generated simulation results
related to the performance of these devices (Sahu et al. 2019, 2021). Then they incor-
porated these obtained synapse and neuron characteristics in codes written on high-
level language Python and designed SNNs in the process. We discuss the reported
SNN designs and performance results briefly here.
Learning happens in the designed SNNs in two possible ways: partially supervised
(Biswas et al. 2016, 2017) and completely unsupervised modes (Diehl and Cook
2015). The STDP property of the synapse, which is very local in nature and doesn’t
work based on optimization of global error unlike the back-propagation algorithm
in ANNs discussed before, always leads to some degree of lack of supervision in
these SNN designs. In the completely unsupervised mode, absolutely no information
about the labels/categories of the inputs is used during the training process. The LIF
neurons at the output layer have a homoeostasis property instead (Zhang and Linden
2003):
t− tspike
−( )
. Vth (t) = Vth (tspike ) + ΔVth e τhomeo
, (6.7)
100 6 Design of Spiking Neural Networks (SNN) with Domain-Wall Devices

where .Vth is the threshold voltage needed for spiking, .tspike is the time when the
neuron spikes, and .τhomeo is the homoeostasis constant. Because of this homoeostasis
property, when each neuron spikes, its threshold voltage for spiking goes up by a
certain magnitude and then gradually decays. This ensures that if a neuron spikes,
its probability of spiking again is low, and thus the various LIF neurons at the output
layer share the different input images of various categories among them (Sahu et al.
2019; Diehl and Cook 2015).
On the other hand, in the partially supervised mode, particular neurons are made
to fire and the remaining neurons are made not to fire (by applying inhibitory currents
on them) at the time of training based on the labels of the training inputs. Thus, some
amount of supervision is carried out during training (while STDP still contributes
to some amount of lack of supervision), leading to the name: partially supervised
learning.
For the Fisher’s Iris data set of flowers (Sahu et al. 2019), classification accuracy
numbers on the training set of 150 samples (train accuracy) and test set of 50 sam-
ples, as reported by Sahu et al. (2019), are plotted as a function of training epoch
number in Fig. 6.7. Both completely unsupervised and partially supervised modes
are considered.
However, the accuracy numbers are much lower for an SNN of the same design if
more involved data sets like MNIST data set of handwritten digits are considered. For
example, since the Fisher’s Iris data set contains data for three categories of flowers,
only three output neurons are considered in the SNN corresponding to the results in
Fig. 6.7. MNIST data set has images of ten digits. However, if only ten output neurons
are used for the SNN corresponding to MNIST, the obtained accuracy will be as low
as 10.% (same as random guess). This is because this kind of STDP-enabled SNN
trains through the process of synaptic weights learning impressions of the digits. For
MNIST, since there are many variations in the images corresponding to the same
digit (since different people have different handwriting), these different variations
get superposed on the weights connecting the different input neurons to the single

Fig. 6.7 Training and test accuracy for the designed SNN, consisting of domain-wall-device-based
neurons and synapses, for classification using the Fisher’s Iris data set of flowers. (Reprinted with
permission from: Sahu U, Pandey A, Goyal K, Bhowmik D (2019) Spike-time-dependent plastic-
ity (STDP)-enabled learning in spiking neural networks using domain-wall-based synapses and
neurons. AIP Adv 9(12) Sahu et al. 2019)
6.3 Performance of SNN with Domain-Wall-Device-Based Neurons and Synapses 101

Fig. 6.8 SNN with 10 output neurons, 1 neuron per digit: a After the 1st epoch of partially super-
vised learning on the MNIST data set, weights of all the synapses connecting the 784 pre-neurons
to a post-neuron are plotted as a 28.×28 pixel greyscale image. Intensity of each pixel is propor-
tional to the value of each weight. b Same plot of synaptic weights after the 25th epoch of partially
supervised learning

output neuron for the digit, resulting in the formation of blobs, as shown in Fig. 6.8.
This blob formation automatically makes classification by the SNN inaccurate.
To prevent this from happening, in a follow-up report (Sahu et al. 2021), Upasana
Sahu et al. have considered 400 neurons in the output layer of the SNN that is trained
102 6 Design of Spiking Neural Networks (SNN) with Domain-Wall Devices

Fig. 6.9 SNN with 400 output neurons, 40 neurons per digit: a After the 1st epoch of completely
unsupervised learning on the MNIST data set, weights of all the synapses connecting the 784 pre-
neurons to a post-neuron are plotted as a 28.×28 pixel greyscale image. Intensity of each pixel is
proportional to the value of each weight. Images corresponding to all such 400 post-neurons are then
plotted together. b Same plot of synaptic weights after the 5th epoch of completely unsupervised
learning

on the MNIST data set. Forty neurons are assigned per digit, as a result. Both in
the partially supervised and completely unsupervised modes, each output neuron
is endowed with the homoeostasis property (as explained before) so that different
neurons can take care of different variations of the same digit. The synaptic weights
are shown as 28 .× 28-pixel images in Fig. 6.9 for completely unsupervised mode
and Fig. 6.10 for partially supervised mode. It can be observed that the kind of blob
formation that happens in the case of one output neuron per digit doesn’t happen
anymore now; weights corresponding to different neurons for the same digit capture
6.3 Performance of SNN with Domain-Wall-Device-Based Neurons and Synapses 103

Fig. 6.10 SNN with 400 output neurons, 40 neurons per digit: a After the 1st epoch of partially
supervised learning on the MNIST data set, weights of all the synapses in the SNN are plotted as
a greyscale image b Plot of synaptic weights after the 2nd epoch of partially supervised learning.
(Reprinted with permission from: Sahu U, Goyal K, Bhowmik D (2021) Training of a Spiking Neural
Network on spintronics-based analog hardware for handwritten digit recognition. In: Proceedings
of international conference on emerging electronics (ICEE). 978-1-7281-8660-3/20/$31.00 .©2020
IEEE Sahu et al. 2021)

different handwriting variations of the same digit. As a result, the classification

accuracy also improves for both modes, as shown in Figs. 6.11and 6.12.
However, we can also observe that the classification accuracy of this STDP-
enabled SNN is lower than that obtained through non-spiking neural networks
(ANNs) designed in the previous chapters. While SNN is deemed to be much more
energy-efficient than ANN because of the event-driven nature of the former, this
typical low accuracy associated with SNNs creates a major impediment for practical
implementation of SNNs. One reason for this low accuracy is the local nature of the
104 6 Design of Spiking Neural Networks (SNN) with Domain-Wall Devices

Fig. 6.11 SNN with 400 output neurons, 40 neurons per digit: Classification accuracy as a function
of epoch number for 1000 train images and 100 test images from the MNIST data set, corresponding
to the completely unsupervised learning mode

Fig. 6.12 SNN with 400 output neurons, 40 neurons per digit: Classification accuracy as a function
of epoch number for 1000 train images and 100 test images from the MNIST data set, corresponding
to the partially supervised learning mode. (Reprinted with permission from: Sahu U, Goyal K,
Bhowmik D (2021) Training of a Spiking Neural Network on spintronics-based analog hardware
for handwritten digit recognition. In: Proceedings of the International Conference on Emerging
Electronics (ICEE), 978-1-7281-8660-3/20/ $31.00 .©2020 IEEE Sahu et al. 2021)

training through STDP (no global error is being minimized here). Related to that,
another factor that creates a major issue for STDP-enabled SNNs is that it’s diffi-
cult to implement hidden layers in the design, again because there’s no concept of
minimization of global error here. On the contrary, ANNs largely leverage from the
hidden layers present in them to achieve high classification accuracy even for data
sets much more complex than MNIST (CIFAR-10, CIFAR-100, etc.).
Multiple approaches have been pursued to overcome this difficulty with SNNs,
with one aspect in common: moving away from STDP. Though STDP is biologically
motivated (experimental evidence of STDP has been found in rat’s hippocampus
Bi and Poo 1998), achieving high classification accuracy is of higher priority than
biological feasibility when it comes to implementations of SNN on neuromorphic
References 105

hardware. Such non-STDP approaches are briefly discussed at the end of Chap. 3;
the interested reader is recommended to go through the references mentioned there
to know more about these non-STDP approaches.

References

Bi G, Poo M (1998) Synaptic modifications in cultured hippocampal neurons: dependence on spike

timing, synaptic strength, and postsynaptic cell type. J Neurosci 18(24):10464–10472
Biswas A, Prasad S, Lashkare S, Ganguly U (2016) A simple and efficient SNN and its performance
and robustness evaluation method to enable hardware implementation. arXiv:1612.02233
Biswas A, Shukla A, Prasad S, Lashkare S, Ganguly U (2017) A simple Spiking Neural Network for
supervised learning for energy efficient hardware implementation. In: Proceedings in international
conference on artificial neural networks
Dayan P, Abbott LF (2005) Theoretical neuroscience: computational and mathematical modeling
of neural systems (computational neuroscience). MIT Press
Diehl PU, Cook M (2015) Unsupervised learning of digit recognition using spike-timing-dependent
plasticity. Front Comput Neurosci 9(99)
Emori S, Martinez E, Lee K-J, Lee H-W, Bauer U, Ahn S-M, Agrawal P, Bono DC, Beach GSD
(2014) Spin Hall torque magnetometry of Dzyaloshinskii domain walls. Phys Rev B 90(18)
Martinez E, Emori S, Perez N, Torres L, Beach GSD (2014) Current-driven dynamics of Dzyaloshin-
skii domain walls in the presence of in-plane fields: full micromagnetic and one-dimensional
analysis. J Appl Phys 115(21)
Roy K, Jaiswal A, Panda P (2019) Towards spike-based machine intelligence with neuromorphic
computing. Nature 575(7784):607–617
Sahu U, Goyal K, Bhowmik D (2021) Training of a Spiking Neural Network on spintronics based
analog hardware for handwritten digit recognition. In: Proceedings of international conference
on emerging electronics (ICEE)
Sahu U, Pandey A, Goyal K, Bhowmik D (2019) Spike time dependent plasticity (STDP) enabled
learning in spiking neural networks using domain wall based synapses and neurons. AIP Adv
9(12)
Sengupta A, Banerjee A, Roy K (2016) Hybrid spintronic-CMOS spiking neural network with
on-chip learning: devices, circuits, and systems. Phys Rev Appl 6(6):064003
Sengupta A, Roy K (2017) Encoding neural and synaptic functionalities in electron spin: a pathway
to efficient neuromorphic computing. Appl Phys Rev 4(4):041105
Zhang W, Linden DJ (2003) The other side of the engram: experience-driven changes in neuronal
intrinsic excitability. Nat Rev Neurosci 4(11):885–900
Chapter 7
Spintronic Oscillators, Their
Synchronization Properties, and
Applications in Oscillatory Neural
Networks (ONNs)

7.1 Auto-oscillations in SHNOs of Nano-pillar Geometry

In Chap. 2 of this book, we showed through numerical simulation of Landau–

Lifschitz–Gilbert–Slonczweski (LLGS) equations that spin transfer torque and spin-
orbit torque that are generated in ferromagnetic metal/oxide/ferromagnetic metal
hetero-structures and heavy metal/ferromagnetic metal/oxide hetero-structures,
respectively, are capable of switching the magnetic moments in the ferromagnetic
metal layer from vertically up to vertically down and the reverse (for perpendicular
magnetic anisotropy or PMA systems). The domain-wall motion phenomenon dis-
cussed in the subsequent chapters is essentially an extension of that phenomenon
because motion of a domain wall means magnetic moments in selected regions of
the device are switching from up to down (or down to up) as a result of spin-orbit
torque from the current flowing through the device.
Here, we consider a nanoscale device fabricated from the same heavy metal/
ferromagnetic metal/oxide hetero-structure, exhibiting PMA and show through
numerical simulation of the same LLGS equations that spin-orbit torque can cause
sustained magnetic oscillations in the ferromagnetic layer of this device. In the nano-
pillar geometry (Chen et al. 2016), as shown in Fig. 7.1 (also shown in Chap. 1),
since we consider uniform-mode oscillations only, solving LLGS equations within
the macro-spin model framework is sufficient and we do not need to solve LLGS
equations in the micromagnetic framework, unlike in the case of the domain-wall
device discussed before which deals with non-uniform magnetization configura-
tions (some moments vertically up and some moments down). Here we focus on
the heavy metal/ferromagnetic metal/oxide hetero-structure-based spin Hall nano-
oscillator (SHNO) device (uniform-mode spin Hall nano-oscillator, or SHNO), where
spin-orbit torque at the heavy metal/ferromagnetic metal interface causes magnetic
oscillations, the same LLGS-based analysis we carry out here is applicable for
uniform-mode spin torque nano-oscillator (STNO), corresponding to the ferromag-
netic metal/oxide/ferromagnetic metal hetero-structure, with suitable modification

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 107
D. Bhowmik, Spintronics-Based Neuromorphic Computing, Springer Tracts in Electrical
and Electronics Engineering, https://doi.org/10.1007/978-981-97-4445-9_7
108 7 Spintronic Oscillators, Their Synchronization Properties …

Fig. 7.1 Schematic of spin Hall nano-oscillator (SHNO) in the nano-pillar geometry: In a heavy-
metal/ferromagnet-based hetero-structure, in-plane current flowing through the heavy-metal layer
can induce auto-oscillation in the ferromagnetic layer. Magnetic moments of the ferromagnetic layer
precess around the vertical axis, along which the external DC magnetic field (. Happ ) is applied.
(Reprinted with permission from: Singh U, Garg N, Kumar S, Muduli PK, Bhowmik D (2021)
Learning of classification tasks with an array of uniform-mode spin Hall nano-oscillators. AIP Adv
11(4) Singh et al. 2021)

to the torque term in the LLGS equation. However, while dealing with SHNOs in
the nano-constriction geometry, LLGS equations in the micromagnetic framework
need to be used (discussed later here).
Using the macro-spin/single-domain model of the SHNO in the nano-pillar geom-
etry corresponding to Fig. 7.1, all the magnetic moments of the ferromagnetic layer in
the SHNO can be modelled by one giant macro-spin vector .m → (Arrott 2005). In-plane
current, flowing through the heavy-metal layer underneath the ferromagnetic layer,
leads to accumulation of electrons with spin polarization (.m→p ) directed in-plane,
orthogonal to current direction, at the heavy metal-ferromagnet interface due to spin
Hall effect (as discussed in detail in Chap. 2). The dynamics of the macro-spin vector
.m→ over time .t, under the effect of .m→p , is given by the Landau–Lifschitz–Gilbert–
Slonczweski (LLGS) equation is given as follows (discussed in Chap. 2):

→
dm γ α
. =− → × H→e f f ) − γ τ→SL − γ
(m (m → × H→e f f ),
→ ×m (7.1)
dt 1 + α2 1 + α2

where .γ: gyromagnetic ratio, .α: Gilbert Damping parameter, .τ→SL : Slonczweski spin
torque term, . H→e f f = − ∂∂ mE→ , and E is the energy density given by
Msat (Hk − μ0 M) 2
. E = −Msat Happl cos θ − cos θ (7.2)
2
Here, . Msat : saturation magnetization, . Happl : applied field which along the z-axis
(out-of-plane direction: Fig. 7.1). The ferromagnetic layer exhibits PMA (hence,
7.1 Auto-oscillations in SHNOs of Nano-pillar Geometry 109

z-axis is the easy axis) (Taniguchi et al. 2013).. Hk is proportional to the PMA strength.
θ is the angle .m
. → makes with the z-axis.
The strength of the spin-orbit-torque term (.τ→SL ) in Eq. 7.1 above depends on the
charge current density (. J ), flowing through heavy metal layer, as follows:
∈ α∈
τ→
. SL =β (m
→ × (m
→ p × m))
→ −β (m
→ ×m
→ p) (7.3)
1+α 2 1 + α2

J θS H E h
β=
. (7.4)
Msat ed

Λ2
.∈= (7.5)
(Λ2 + 1) + (Λ2 − 1)(m
→ · m→p )

Here, . J : in-plane current density through the heavy metal, .θ S H E : spin Hall angle
of the heavy metal, .d: thickness of the ferromagnetic layer, .α: damping factor, and
.Λ: Slonczweski parameter (Taniguchi et al. 2013; Garg et al. 2021).

Equation 7.1 can be converted into spherical coordinates (.r, θ, φ) such that . ddtm→
takes the following form:

→
dm d e→r dθ dφ
. = = e→θ + sin θ e→φ (7.6)
dt dt dt dt

Here, .e→r , .e→θ and .e→φ are the basis vectors in the spherical-coordinate system. .e→r is
taken along the direction of .m → (Roma et al. 2014). Since the norm of the macro-spin
→ is conserved and is always equal to 1, .r = 1 and . dr
vector .m dt
= 0. Using the above
expression for . ddtm→ , Eq. 7.1 translates into the following differential equations for the
direction of .m, → in terms of .θ and .φ, as a function of time (.t):

dθ
(1 + α2 )
. = −γαHe f f sin(θ) − γβ∈cos(θ)cos(φ) − γαβ∈sin(φ) (7.7)
dt

dφ
sin(θ)(1 + α2 )
. = γ He f f sin(θ) + γβ∈sin(φ) − γαβ∈cos(θ)cos(φ) (7.8)
dt
As reported by Garg et al. (2021), numerical simulation of Eqs. 7.7 and 7.8 yields
steady-state oscillations when current density . J in the equations is above a certain
threshold value. For these current values, based on the numerically obtained values
of .θ(t) and .φ(t), the macro-spin vector is obtained again in the Cartesian coordinates
as a function of time: .m x (t), .m y (t), .m z (t). .m x (t), and .m y (t), plotted against time .t,
are essentially sinusoidal curve, showing that the macro-spin vector (representative
of all magnetic moments locked to each other) is precessing around the vertical (z)
axis.
110 7 Spintronic Oscillators, Their Synchronization Properties …

Fig. 7.2 Single-

domain/macro-spin
simulation of the SHNO in
nano-pillar geometry: a
Natural frequency versus
in-plane charge current
density, characteristic for a
single SHNO b Precession
angle .θ (angle between the
macro-spin vector and the
vertical axis) versus in-plane
charge current density.
(Reprinted with permission
from: Garg N, Bhotla SVH,
Muduli PK, Bhowmik D
(2021)
Kuramoto-model-based data
classification using the
synchronization dynamics of
uniform-mode spin Hall
nano-oscillators.
Neuromorphic Comput Eng
1(2) Garg et al. 2021)

Taking Fast Fourier Transform (FFT) of .m x (t), the natural frequency of auto-
oscillation can be obtained. Natural frequency of oscillation versus current density,
as reported by Garg et al. (2021), is shown in Fig. 7.2a, and angle of precession
versus current density is plotted in Fig. 7.2b. It is observed that as the current density
increases, the macro-spin vector makes a large angle of precession with the vertical
axis. Since now it needs to cover a larger periphery for one complete rotation, the time
period of oscillation increases, or the natural frequency of the oscillator decreases.
All relevant parameter values in Eqs. 7.7 and 7.8 can be found in the report by Garg
et al. (2021) and also that by Taniguchi et al. (2013).
7.3 Synchronization of Dipole-Coupled Nano-pillar SHNOs 111

7.2 Auto-oscillations in SHNOs of Nano-constriction

Geometry

The macro-spin model discussed above cannot be used to model SHNOs in the nano-
constriction geometry though, like that shown in Fig. 7.3a (Divinskiy et al. 2017).
This is because in the nano-constriction geometry, the in-plane current density is
much higher at the nano-constriction compared to the rest of the magnetic region,
and hence auto-oscillation is triggered only at the nano-constriction region and not
outside it. Given the non-uniform nature of the current distribution and magnetization
configuration in this geometry, LLGS equations in the micromagnetic framework are
needed to model SHNOs in this geometry.
In the report by Divinskiy et al. (2017), micromagnetic simulations have been
carried out of the SHNO in nano-constriction geometry, as shown in Fig. 7.3a. Rel-
evant simulation parameters can be found in the report. Figure 7.3b shows that as
the applied DC in-plane magnetic field . H// increases, the angle of precession of the
magnetization increases. (It is to be noted that .θ in the report and figure by Divinsky
et al. (Fig. 7.3a) is the angle the magnetic moment vector makes with the in-plane
(.x-axis), and hence the angle of precession (.θ as per our earlier convention) is equal
to 90.◦ – .θ as per Divinsky et al.’s convention.)
Figure 7.3b also shows that as the angle of precession goes up, the frequency of
oscillation goes down, which agrees with the prediction from macro-spin model for
the nano-pillar-geometry-based SHNO as well. The same report by
Divinsky et al. also contains experimental observation of auto-oscillations in the
nano-constriction-geometry SHNO, based on micro-focus Brillouin light scattering
(BLS). Figure 7.4c, as taken from the report, shows that the natural frequency of
auto-oscillation decreases with increase in current flowing through the device, much
like in the case of SHNO in nano-pillar geometry reported earlier. The interested
reader is recommended to go through the report by Divinsky et al. for more details
on the experimental measurement.

7.3 Synchronization of Dipole-Coupled Nano-pillar SHNOs

Having discussed how auto-oscillations can be triggered in individual spintronic

oscillators through the application of current, we now turn our focus on how these
oscillators can be synchronized. If multiple oscillators have slightly different natural
frequencies of auto-oscillation, then synchronization means making them oscillate at
a common frequency as opposed to different frequencies. When such synchronization
happens, a proper phase relationship can be defined between any two oscillators,
otherwise not.
Here, we first discuss the modelling of dipole-coupled uniform-mode SHNOs
(in the nano-pillar geometry) through coupled LLGS equations (macro-spin model)
and show their synchronization behaviour through the numerical solution of these
112 7 Spintronic Oscillators, Their Synchronization Properties …

Fig. 7.3 a Schematic of the

SHNO in nano-constriction
geometry b Based on
calculations, out-of-plane
angle of the equilibrium
magnetization versus static
field is plotted (circles), and
uniform ferromagnetic
resonance (FMR) frequency
versus static field is also
plotted (point-down
triangles). Point-up triangles
represent the FMR
frequencies determined from
Brillouin light scattering
(BLS) measurements of
thermally excited magnons.
c Calculated spectrum of
spin waves in the
ferromagnetic multilayer of
the nano-constriction SHNO
in (a), magnetized by an
in-plane field of 2000 Oe.
More details can be found in
Divinskiy et al. (2017).
(Figure reprinted with
permission from: Divinskiy
B, Demidov VE, Kozhanov A,
Rinkevich AB, Demokritov
SO, Urazhdin S (2017)
Nano-constriction spin Hall
oscillator with perpendicular
magnetic anisotropy. Appl
Phys Lett 111: 032405
Divinskiy et al. 2017)
7.3 Synchronization of Dipole-Coupled Nano-pillar SHNOs 113

Fig. 7.4 a Corresponding to

the SHNO in
nano-constriction geometry
of Fig. 7.3a, spectra of
magnetization
auto-oscillations are plotted
for different driving currents.
b Peak intensity (point-down
triangles) and intensity
integrated over the BLS
spectrum (point-up triangles)
versus DC current. c Current
dependence of the central
frequency of the detected
spectral peak. Arrow marks
the threshold current above
which auto-oscillations are
triggered. More details can
be found in Divinskiy et al.
(2017). (Figure reprinted
with permission from:
Divinskiy B, Demidov VE,
Kozhanov A, Rinkevich AB,
Demokritov SO, Urazhdin S
(2017) Nano-constriction
spin Hall oscillator with
perpendicular magnetic
anisotropy. Appl Phys Lett
111:032405 Divinskiy et al.
2017)

coupled LLGS equations. This analysis has been carried out in detail in the report by
Garg et al. (2021) and that by Sri Vasudha Hemadri Bhotla et al.; we cover it briefly
here.
As per the report by Amin et al. (2009), if the wavelength associated with the ema-
nating RF magnetic field due to precession of the magnetic moment of the uniform-
mode spin oscillator (STNO or SHNO) is given by .λ and the distance between the
spin oscillator and the point at which the field is observed is given by . R, there are
114 7 Spintronic Oscillators, Their Synchronization Properties …

Fig. 7.5 Schematic of a pair of dipole-field-coupled nano-pillar SHNOs separated by distance . R,

with the direction of the macro-spin vector for SHNO1 given by (.θ1 , .φ1 ) and that of the macro-spin
vector for SHNO2 given by (.θ1 , .φ1 ) (more details in Garg et al. 2021)

three terms in the expression of the RF field: . R13 , . λ(R1 2 ) , and . (λ21)R terms. In the far-
field regime (. R >> λ), the . (λ21)R term dominates, whereas in the near-field regime
(. R << λ), the . R13 term dominates (Amin et al. 2009).
Based on the macro-spin-model-based simulation of the nano-pillar SHNO pre-
sented earlier (Fig. 7.2), the natural frequency of the SHNO is between 3 and 4.6
GHz (Fig. 7.2). The corresponding .λ is between 65 and 100 mm. In the system of
dipole-coupled SHNOs considered by Neha Garg et al. to demonstrate synchroniza-
tion (Fig. 7.5), the centre-to-centre distance between two adjacent SHNOs (. R) is in
the range of 250–350 nm. Thus, . R << λ here, and the near-field regime’s RF field
formula (. R13 ) is the only term to be considered.
In their report, Neha Garg et al. have considered a two-SHNO system or an SHNO
pair: SHNO1 is at the position (0, 0, 0), and SHNO2 is at position (. R, 0, 0), in the
Cartesian coordinate system (as shown in Fig. 7.5). In-plane current density into
SHNO1 is given by . J1 and that into SHNO2 is given by . J2 . Using the expression for
the . R13 term provided in the report by Amin et al. (2009), when the magnetic moment
1 dφ1
of SHNO1 precesses around the z-axis with the instantaneous frequency (. 2π dt
) and
2,1
an angle of precession .θ1 , RF magnetic field . H R F experienced by SHNO2 due to
SHNO1 is given by

μ0 Msat V sin(θ1 )
. H R2,1F = (2cos(φ1 )x̂ − sin(φ1 ) ŷ), (7.9)
4π(R 3 )

where .V : volume of the SHNO, and . Msat : saturation magnetization. Similarly, the
RF magnetic field experienced by SHNO1 due to the precession of the macro-spin
vector in SHNO2 can also be expressed (. H R1,2F ). In this case, SHNO2’s instantaneous
1 dφ2
frequency = . 2π dt
, and its angle of precession = .θ2 .
7.3 Synchronization of Dipole-Coupled Nano-pillar SHNOs 115

After including . H R1,2F in . H→e f f in the LLGS equation for SHNO1 (Eq. 7.1) and
converting it to spherical coordinates, Nega Garg et al. obtained the following two
equations, where .θ1 and .φ1 correspond to coordinates of the macro-spin vector for
SHNO1, and .θ2 and .φ2 correspond to the coordinates of the macro-spin vector for
SHNO2:
dθ1
. (1 + α2 ) = − γαHe f f 1 sin(θ1 )
dt
− γβ1 ∈1 (cos(θ1 )cos(φ1 ) + αsin(φ1 ))
+ 2K sin(θ2 )cos(φ2 )(αcos(θ1 )cos(φ1 ) − sin(φ1 ))
− K sin(θ2 )sin(φ2 )(αcos(θ1 )sin(φ1 ) + cos(φ1 )) (7.10)

dφ1
. sin(θ1 )(1 + α2 ) = γ He f f 1 sin(θ1 )
dt
+ γβ1 ∈1 (sin(φ1 ) − αcos(θ1 )cos(φ1 ))
−2K sin(θ2 )cos(φ2 )(cos(θ1 )cos(φ1 ) + αsin(φ1 ))
− K sin(θ2 )sin(φ2 )(αcos(φ1 ) − cos(θ1 )sin(φ1 )),
(7.11)

μ0 Msat V J1 θ S H E h
where . K = 4π(R 3 )
, . He f f 1 = Happl + (Hk − μ0 Msat )cos(θ1 ), .β1 = Msat ed
, and .∈1
Λ2
= (Λ2 +1)+(Λ2 −1)sin(θ1 )cos(φ1 )
.

Similarly, after including . H R2,1F in . H→e f f in the LLGS equation for SHNO2 and
converting it to spherical coordinates, Neha Garg et al. obtained the following two
equations, where .θ1 and .φ1 correspond to the macro-spin vector for SHNO1, and .θ2
and .φ2 correspond to the macro-spin vector for SHNO2:

dθ2
(1 + α2 )
. = − γαHe f f 2 sin(θ2 )
dt
− γβ2 ∈2 (cos(θ2 )cos(φ2 ) + αsin(φ2 ))
+ 2K sin(θ1 )cos(φ1 )(αcos(θ2 )cos(φ2 ) − sin(φ2 ))
− K sin(θ1 )sin(φ1 )(αcos(θ2 )sin(φ2 ) + cos(φ2 )) (7.12)

dφ2
.sin(θ2 )(1 + α2 ) = γ He f f 2 sin(θ2 )
dt
+ γβ2 ∈2 (sin(φ2 ) − αcos(θ2 )cos(φ2 ))
−2K sin(θ1 )cos(φ1 )(cos(θ2 )cos(φ2 ) + αsin(φ2 ))
− K sin(θ1 )sin(φ1 )(αcos(φ2 ) − cos(θ2 )sin(φ2 )),
(7.13)
116 7 Spintronic Oscillators, Their Synchronization Properties …

Fig. 7.6 Macro-spin

modelling of two
dipole-coupled SHNOs in
the nano-pillar geometry a
Variation of the
instantaneous frequencies of
the two SHNOs, SHNO1 and
SHNO2, with the natural
frequency of SHNO2 (. F2 ),
while the natural frequency
of SHNO1 (. F1 ) is kept
constant at 3.94 GHz and the
separation between the two
SHNOs is 100 nm. b
Variation of the
synchronization range for
different values of distance
between the two SHNOs.
(Reprinted with permission
from: Garg N, Bhotla SVH,
Muduli PK, Bhowmik D
(2021)
Kuramoto-model-based data
classification using the
synchronization dynamics of
uniform-mode spin Hall
nano-oscillators.
Neuromorphic Comput Eng
1:2 Garg et al. 2021)

μ0 Msat V J2 θ S H E h
where . K = 4π(R 3 )
, and . He f f 2 = Happl + (Hk − μ0 Msat )cos(θ2 ), .β2 = Msat ed
and
Λ2
∈ =
. 2
(Λ2 +1)+(Λ2 −1)sin(θ2 )cos(φ2 )
.
Keeping the natural frequency of SHNO1 constant at 3.94 GHz by maintaining
a constant value of . J1 , Neha Garg et al. vary . J2 such that the natural frequency of
SHNO2 varies as shown in Fig. 7.6a. For each pair of current-density values (. J1 , . J2 ),
the four coupled Eqs. 7.10, 7.11, 7.12, and 7.13 are solved numerically (parameter
values used can be found in the report by Garg et al. 2021). Thus, they obtain how
.θ1 , .φ1 , .θ2 , and .φ2 vary with time, and from that through the FFT-based method

discussed earlier, they report the instantaneous frequencies (. f 1 and . f 2 ) of the two
SHNOs, which can be different from the natural frequencies (. F1 and . F2 ) of the two
SHNOs due to the coupling between the SHNOs.
Following this method, Neha Garg et al. report the variation of the instantaneous
frequencies of SHNO1 and SHNO2 (. f 1 and . f 2 ) as functions of the natural frequency
7.4 Synchronization of Spin-Wave-Coupled Nano-constriction SHNOs 117

of SHNO2 (. F2 ) (Fig. 7.6a) (natural frequency of SHNO1 is fixed at 3.94 GHz) (Garg
et al. 2021). The distance between the two SHNOs (. R) is 100 nm. Figure 7.6a shows
that within a certain range of natural frequency of SHNO2 (labelled as “synchroniza-
tion range”), instantaneous frequency of SHNO1 (. f 1 ) .= instantaneous frequency of
SHNO2 (. f 2 ) .= natural frequency of SHNO2, i.e., the two SHNOs are synchronized
in this range and have a well-defined phase relationship. Outside this range, the
instantaneous frequency of SHNO1 is the same as the natural frequency of SHNO1
(3.94 GHz), and instantaneous frequency of SHNO2 follows the natural frequency
of SHNO2, which clearly means that they are not synchronized anymore.
Neha Garg et al. repeat this method for different values of separation between
the two SHNOs (. R) and report the synchronization range for each such distance
of separation in Fig. 7.6b (Garg et al. 2021). The synchronization range is found to
decrease with increase in the distance between SHNOs, which makes intuitive sense
because the dipole coupling gets weaker (. R13 relationship) as the distance . R between
SHNOs increases.

7.4 Synchronization of Spin-Wave-Coupled

Nano-constriction SHNOs

We next discuss micromagnetics-based modelling of spin-wave-coupled SHNOs in

the nano-constriction geometry (as reported by Kendziorczyk et al. 2016), which
also exhibit similar synchronization behaviour as that of dipole-coupled nano-pillar
SHNOs discussed earlier (Fig. 7.6). Figure 7.7a from the report by Kendziorczyk
et al. (2016) shows the schematic of the two nano-constriction-based SHNOs, where
the separation between the two nano-constrictions can be changed by changing .Δx
and .Δy. Unlike in the case of nano-pillar SHNOs of Fig. 7.6, these nano-constriction
SHNOs share the same ferromagnetic layer and hence couple with each other through
spin waves, which are essentially propagation oscillations of the magnetic moments
of the ferromagnetic layer. A more detailed discussion of spin waves can be found
in different magnetism textbooks (Coey 2010; Blundell 2001).
In the report by Kendziorczyk et al. (2016), there are two ways in which the
strength of coupling between the two nano-constriction SHNOs is modulated. In the
first method, the separation between the nano-constrictions is changed by changing
.Δx and .Δy (Fig. 7.7a). For the unoptimized geometry (.Δx = 0), as .Δy is varied,
we observe for almost values of .Δy, two distinct frequencies in the power spectral
density (PSD) of the nano-constrictions (two cyan lines amidst blue background in
Fig. 7.8e), which means that the two nano-constriction SHNOs are not synchronized.
On the other hand, for the optimized geometry (.Δx = 202 nm), for a wide range
of .Δy, only one frequency is observed corresponding to PSD of the two nano-
constrictions (only one cyan line amidst blue background in Fig. 7.8e), signifying
complete synchronization between the two SHNOs.
118 7 Spintronic Oscillators, Their Synchronization Properties …

Fig. 7.7 a Schematic of the spin-wave-coupled nano-constriction SHNO array simulated in the
report by Kendziorczyk and Kuhn (2016) b Current density and c internal field in the ferromagnetic
layer as a result of the current flow. More details are in Kendziorczyk and Kuhn (2016). (Reprinted
figure with permission from: Kendziorczyk T, Kuhn T (2016) Mutual synchronization of nano-
constriction-based spin Hall nano-oscillators through evanescent and propagating spin waves.
Phys Rev B 93:134413, Copyright 2016 by American Physical Society, https://doi.org/10.1103/
PhysRevB.93.134413 Kendziorczyk and Kuhn 2016)

In the second method, the separation between the nano-constrictions is kept fixed,
and the magnitude of the current (. Iwir e ) through the red wire in the schematic of
Fig. 7.7a is varied in the simulations. Since one of the two nano-constrictions is
physically much closer to the wire than the other, its natural frequency is modulated
by . Iwir e , while the other one’s not. This enables variation of the natural frequency
of one nano-constriction SHNO while keeping the other SHNO’s natural frequency
fixed, just like Neha Garg et al. varied the natural frequency of one nano-pillar
SHNO while keeping the other one’s frequency fixed (Fig. 7.6). Thus, in the case of
nano-constriction SHNOs, the allowed difference in natural frequencies so that the
SHNOs stay synchronized can be calculated (much like in the case of nano-pillar
SHNOs) and expressed in terms of .ΔIwrite . . Iwrite is shown as a function of the angle
of separation between the nano-constrictions (.φ) in Fig. 7.9c and as a function of the
distance between the nano-constrictions in Fig. 7.9d.
7.5 Neuromorphic Computing Using Synchronized Spintronic … 119

Fig. 7.8 a–d Spatial amplitude profile of the out-of-plane magnetic moment for two synchronized
nano-constriction SHNOs (shown in Fig. 7.7a) (more details in Kendziorczyk and Kuhn 2016)
(e) and (f) Sum of the calculated power spectral densities (PSD) from signals of the two nano-
constriction SHNOs for (e) unoptimized geometry: .Δx = 0 in Fig. 7.7a, and f optimized geometry:
.Δx = 202 nm in Fig. 7.7a. More details are in Kendziorczyk and Kuhn (2016). (Reprinted figure
with permission from: Kendziorczyk T, Kuhn T (2016) Mutual synchronization of nano-constriction-
based spin Hall nano-oscillators through evanescent and propagating spin waves. Phys Rev B
93:134413, Copyright 2016 by American Physical Society, https://doi.org/10.1103/PhysRevB.93.
134413 Kendziorczyk and Kuhn 2016)

On the experimental front, spin-wave-based synchronization between SHNOs in

the nano-constriction geometry has been reported not just for a pair of SHNOs but
for much larger SHNO arrays. As per the report by Zahedinejad et al. (2020), 64
SHNOs have been experimentally demonstrated to be synchronized.

7.5 Neuromorphic Computing Using Synchronized

Spintronic Oscillators: Oscillatory Neural Networks
(ONNs)

When spintronic oscillators are considered for applications such as RF and microwave
power generators, then the target is to synchronize as many SHNOs as possible to
create constructive interference of their oscillations (the oscillation amplitudes add
up) and maximize their power output. But for neuromorphic computing applications,
the natural frequency values of the SHNOs need to be tuned such that for some val-
ues, some SHNOs are in sync and others are not, while for other values, some other
SHNOs are in sync and others are not. These different combinations of SHNOs in
sync correspond to different classes into which the input data need to be classified
(e.g., the ten digits for the MNIST data set discussed before).
120 7 Spintronic Oscillators, Their Synchronization Properties …

Fig. 7.9 a and b Sum of the PSDs calculated from the signals of the two nano-constriction SHNOs
in Fig. 7.7a, under the influence of a nanowire with current flowing through it: . Iwir e , for different
geometries c Variation of the synchronization region, determined in terms of .ΔIwir e , with different
angles of separation between the SHNOs (.φ), for different values of currents exciting the SHNOs.
d Variation of the synchronization region, determined in terms of .ΔIwir e , with different distances
between the SHNOs for the optimized geometry (angle of separation.= 120.◦ ). Current through each
SHNO .= 2 mA. The insets show the spatial amplitude profile for the second harmonic signal of the
out-of-plane magnetization. More details are in Kendziorczyk and Kuhn (2016). (Reprinted figure
with permission from: Kendziorczyk T, Kuhn T (2016) Mutual synchronization of nano-constriction-
based spin Hall nano-oscillators through evanescent and propagating spin waves. Phys Rev B
93:134413, Copyright 2016 by American Physical Society, https://doi.org/10.1103/PhysRevB.93.
134413 Kendziorczyk and Kuhn 2016)
7.5 Neuromorphic Computing Using Synchronized Spintronic … 121

Fig. 7.10 Schematic for offline learning (to be implemented on a conventional computer) in a
two-input-two-output oscillatory neural network (ONN) for binary classification, more details in
Vodenicarevic et al. (2018), Vodenicarevic (2017), and Hemadri Bhotla et al. (2023). (Reprinted
with permission from: Hemadri Bhotla SV, Garg N, Aggarwal T, Muduli PK, Bhowmik D (2023)
An oscillator-synchronization-based offline learning algorithm, with on-chip inference on an array
of spin Hall nano-oscillators. IEEE Trans Nanotechnol 22:136–148; 1536-125X O2023 c IEEE
Hemadri Bhotla et al. 2023)

The algorithm that guides data classification using synchronized oscillators (also
known as oscillatory neural networks (ONNs)) is different from the algorithms for
ANN and SNN we have discussed before in this book. Like in ANNs and SNNs
though, there are two aspects of the neural networks: training and testing. But unlike
in those cases, researchers don’t usually consider training ONNs on the actual hard-
ware array of spintronic oscillators. Instead, the ONN is trained on a conventional
computer, and then the final trained network is implemented on the array of spintronic
oscillators.
Details of the training algorithm for ONNs can be found in reports by
Vodenicarevic et al. (2018), Vodenicarevic (2017) and Hemadri Bhotla et al. (2023).
We cover it briefly here. As shown in Fig. 7.10, the original input data in higher
dimensions (27 for Wisconsin Breast Cancer (WBC) data set: Fig. 7.11a, 4 for
Fisher’s Iris data set: Fig. 7.11b, 784 for MNIST data set: Fig. 7.11c) is first reduced
to two-dimensional data using a dimension-reduction algorithm such that in the new
two-dimensional space, each cluster corresponds to each output category/label of
122 7 Spintronic Oscillators, Their Synchronization Properties …

Fig. 7.11 Each input data

item in the Wisconsin Breast
Cancer (WBC) data set (a),
the Fisher’s Iris data set (b),
and the MNIST data set (c)
is transformed into two
dimensions: Feature 1,
Feature 2. Then Feature 2 is
plotted against Feature 1 for
each of the data sets to show
clustering of the data, based
on their labels/categories, in
reduced dimensions.
(Reprinted with permission
from: Singh U, Garg N,
Kumar S, Muduli PK,
Bhowmik D (2021) Learning
of classification tasks with an
array of uniform-mode spin
Hall nano-oscillators. AIP
Adv 11:4 Singh et al. 2021)

the data (Fig. 7.11). Popular dimension-reduction algorithms for this purpose are
principle component analysis (PCA), neighbourhood component analysis (NCA),
t-distributed stochastic neighbour embedding (tSNE), etc. (Singh et al. 2021).
Binary classification is considered in the report by Hemadri Bhotla et al. (2023),
and hence only four oscillators are considered: two for input and two for output
(Fig. 7.10). The dynamics of four coupled oscillators is modelled in the algorithm
through the Kuramoto model, which is essentially a mathematical model applica-
ble not only to spintronic oscillators but also to a wide range of oscillators, like
chemical oscillators, mechanical oscillators, electronic-circuit-based oscillators, etc.
(Vodenicarevic 2017). Out of the four oscillators, two are considered input oscilla-
tors and two are output oscillators. The input data, after transformation, is in two
7.5 Neuromorphic Computing Using Synchronized Spintronic … 123

Fig. 7.12 a Schematic for on-chip inference on a two-input-two-output nano-pillar SHNO array
for binary classification b Schematic of an individual nano-pillar SHNO in the array, more details
in Vodenicarevic et al. (2018), Vodenicarevic (2017) and Hemadri Bhotla et al. (2023). (Reprinted
with permission from: Hemadri Bhotla SV, Garg N, Aggarwal T, Muduli PK, Bhowmik D (2023)
An oscillator-synchronization-based offline learning algorithm, with on-chip inference on an array
of spin Hall nano-oscillators. IEEE Trans Nanotechnol 22:136–148; 1536-125X O2023 c IEEE
Hemadri Bhotla et al. 2023)

dimensions and is converted into the natural frequencies of the two input oscillators
(. F1in , . F2in ) (Fig. 7.10). Supervised learning scheme is followed here, with the final
target being the following: for one category/label of the input data (label A say), the
two output oscillators are synchronized; and for the other category/label (label B),
the two output oscillators are not synchronized. Accordingly, as every input sample
is fed to the four-oscillator model during training, an error is calculated based on
this desired level of synchronization and what is actually obtained, and then the nat-
ural frequencies of the output oscillators (. F3out , . F4out ) are updated following gradient
descent, like in the case of ANNs (Fig. 7.10). The exact equations corresponding to
this algorithm are provided in the reports by Vodenicarevic et al. (2018; 2017) and
Hemadri Bhotla et al. (2023), and interested readers are recommended to go through
them.
124 7 Spintronic Oscillators, Their Synchronization Properties …

Fig. 7.13 On-chip inference using an array of nano-pillar SHNOs, as shown in Fig. 7.12a and
modelled by coupled LLGS equations (macro-spin model): Current densities of appropriate magni-
tudes are applied on output SHNOs such that their natural frequencies are the same as that obtained
for the output oscillators after training through our proposed offline learning algorithm. Thus, the
synchronization maps obtained through training can be implemented on the array of SHNOs for
on-chip inference. The following binary classification cases are considered. a . J3out = 0.86 × 1011
A/m.2 ,. J4out = 0.8 × 1011 A/m.2 , resulting in. F3out = 4.04 GHz,. Fout
4 = 4.22 GHz, corresponding to

Iris: Setosa versus Versicolour b. J3 = 0.858 × 10 A/m. ,. J4 = 0.797 × 1011 A/m.2 , resulting
out 11 2 out
3 = 4.05 GHz, . F 4 = 4.23 GHz, corresponding to digit “1” versus digit “0” case within
in . Fout out
the MNIST data set, more details in Hemadri Bhotla et al. (2023), Vodenicarevic et al. (2018) and
Vodenicarevic (2017). (Reprinted with permission from: Hemadri Bhotla SV, Garg N, Aggarwal T,
Muduli PK, Bhowmik D (2023) An oscillator-synchronization-based offline learning algorithm, with
on-chip inference on an array of spin Hall nano-oscillators. IEEE Trans Nanotechnol 22:136–148;
1536-125X O2023c IEEE Hemadri Bhotla et al. 2023)
References 125

As a result of this training, once the final natural frequencies of the output oscilla-
tors (. F3out , . F4out ) get fixed, currents (. J3out , . J4out ) are applied on the output SHNOs in
the physical hardware such that their natural frequencies are. F3out and. F4out (Fig. 7.12).
Different input samples are fed into the oscillator hardware as natural frequencies
of input oscillators . F1in , . F2in (corresponding currents: . J1in , . J2in ). If the two output
oscillators synchronize, then the input sample is identified with label A, otherwise
label B. This is the inference scheme for this ONN.
Sri Vasudha Hemadri Bhotla et al. report simulation of this inference scheme on an
array of four dipole-coupled nano-pillar SHNOs, as shown in Fig. 7.12. The SHNOs
are modelled through coupled LLGS equations in the macro-spin model as discussed
before. Based on this simulation, final synchronization maps for binary classification
using the Fisher’s Iris data set (Setosa flower type vs. Versicolour flower type) are
shown in Fig. 7.13a. Output oscillators synchronizing (red region) correspond to
flowers of Setosa type, while output oscillators not synchronizing (yellow region)
correspond to flowers of Versicolour type. Final synchronization maps for binary
classification using MNIST data set (digit “1” vs. digit “0”) are shown in Fig. 7.13b.
Output oscillators synchronizing (red region) correspond to images of digit “1,” while
output oscillators not synchronizing (yellow region) correspond to images of digit
“0.”
This kind of inference scheme has been recently experimentally demonstrated not
on an array of dipole-coupled uniform-mode nano-pillar SHNOs but on an array of
coupled much-larger-size vortex-mode STNOs in the report by Romera et al. (2018).
Different vowel sounds have been distinguished through the fabricated ONN in their
report (Romera et al. 2018). In the aforementioned experimental report on spin-wave
coupled nano-constriction SHNOs by Zahedinejad et al., a similar inference scheme
has also been considered for their 4 .× 4 SHNO array.

References

Amin N, Xi H, Tang MX (2009) Analysis of electromagnetic fields generated by a spin-torque

oscillator. IEEE Trans Magn 45(10)
Arrott AS (2005) Introduction to micromagnetics. In: Heinrich B, Bland JAC (eds) Ultrathin mag-
netic structures IV: applications of nanomagnetism. Springer
Blundell S (2001) Magnetism in condensed matter. Oxford University Press
Chen T, Dumas RK, Eklund A, Muduli PK, Houshang A, Awad AA, Dürrenfeld P, Malm BG, Rusu
A, Åkerman J (2016) Spin-torque and spin-Hall nano-oscillators. Proc IEEE 104(10):1919–1945
Coey JMD (2010) Magnetism and magnetic materials. Cambridge University Press
Divinskiy B, Demidov VE, Kozhanov A, Rinkevich AB, Demokritov SO, Urazhdin S (2017)
Nanoconstriction spin-Hall oscillator with perpendicular magnetic anisotropy. Appl Phys Lett
111:032405
Garg N, Bhotla SVH, Muduli PK, Bhowmik D (2021) Kuramoto-model-based data classification
using the synchronization dynamics of uniform-mode spin Hall nano-oscillators. Neuromorphic
Comput Eng 1(2)
126 7 Spintronic Oscillators, Their Synchronization Properties …

Hemadri Bhotla SV, Garg N, Aggarwal T, Muduli PK, Bhowmik D (2023) An oscillator-
synchronization-based off-line learning algorithm, with on-chip inference on an array of spin
Hall nano-oscillators. IEEE Trans Nanotechnol 22:136–148
Kendziorczyk T, Kuhn T (2016) Mutual synchronization of nanoconstriction-based spin Hall nano-
oscillators through evanescent and propagating spin waves. Phys Rev B 93:134413
Roma F, Cugliandolo LF, Lozano GS (2014) Numerical integration of the stochastic Landau-
Lifshitz-Gilbert equation in generic time-discretization schemes. Phys Rev E 90:023203
Romera M, Talatchian P, Tsunegi S, Araujo FA, Cros V, Bortolotti P, Trastoy J, Yakushiji K,
Fukushima A, Kubota H, Yuasa S (2018) Vowel recognition with four coupled spin-torque nano-
oscillators. Nature 563:230–34
Singh U, Garg N, Kumar S, Muduli PK, Bhowmik D (2021) Learning of classification tasks with
an array of uniform-mode spin Hall nano-oscillators. AIP Adv 11:4
Taniguchi T, Arai H, Kubota H, Imamura H (2013) Theoretical study of spin-torque oscillator with
perpendicularly magnetized free layer. IEEE Trans Magn 50:1–4
Vodenicarevic D (2017) Rhythms and oscillations: a vision for nanoelectronics. Doctoral disserta-
tion, Université Paris Saclay (COmUE)
Vodenicarevic D, Locatelli N, Grollier J, Querlioz D (2018) Nano-oscillator-based classification
with a machine learning-compatible architecture. J Appl Phys 124:152117
Zahedinejad M, Awad AA, Muralidhar S, Khymyn R, Fulara H, Mazraati H, Dvornik M, Åkerman
J (2020) Two-dimensional mutually synchronized spin Hall nano-oscillator arrays for neuromor-
phic computing. Nat Nanotechnol 15:47–52