1

Artificial Intelligence for Quantum Error

Correction: A Comprehensive Review
Zihao Wang Hao Tang∗

Abstract—Quantum Error Correction (QEC) is the process of detecting and correcting errors in quantum systems, which are prone
to decoherence and quantum noise. QEC is crucial for developing stable and highly accurate quantum computing systems, therefore,
several research efforts have been made to develop the best QEC strategy. Recently, Google’s breakthrough shows great potential to
improve the accuracy of the existing error correction methods [1]. This survey provides a comprehensive review of advancements in
the use of artificial intelligence (AI) tools to enhance QEC schemes for existing Noisy Intermediate Scale Quantum (NISQ) systems.
arXiv:2412.20380v1 [quant-ph] 29 Dec 2024

Specifically, we focus on machine learning (ML) strategies and span from unsupervised, supervised, semi-supervised, to reinforcement
learning methods. It is clear from the evidence, that these methods have recently shown superior efficiency and accuracy in the QEC
pipeline compared to conventional approaches. Our review covers more than 150 relevant studies, offering a comprehensive overview of
progress and perspective in this field. We organized the reviewed literature on the basis of the AI strategies employed and improvements
in error correction performance. We also discuss challenges ahead such as data sparsity caused by limited quantum error datasets and
scalability issues as the number of quantum bits (qubits) in quantum systems kept increasing very fast. We conclude the paper with
summary of existing works and future research directions aimed at deeper integration of AI techniques into QEC strategies.

Index Terms—Artificial Intelligence (AI), Machine Learning (ML), Quantum Error Correction (QEC), Quantum Computing, Survey

1 I NTRODUCTION qubits to correct arbitrary single-qubit errors. Surface codes,

a type of topological code, arrange qubits on a lattice and
Q UANTUM computing has potential to revolutionize
computational capabilities, by harnessing the princi-
ples of quantum mechanics to address problems conven-
achieve high error thresholds. However, these approaches
face significant challenges, particularly regarding scalability
and resource requirements. Implementing surface codes,
tionally intractable for classical computers [2], [3]. Its appli-
for instance, may require thousands of physical qubits to
cations span various domains, including cryptography [4],
encode a single logical qubit, making large-scale quantum
[5], optimization [6], and simulation of physical quantum
computing infeasible [15]. Additionally, the error correction
systems [7], [8]. However, practical implementation of quan-
process involves complex syndrome measurements and de-
tum computing faces significant challenges, primarily due
coding algorithms, which become computationally inten-
to the vulnerability of quantum systems to errors caused by
sive as system size increases [14]. Furthermore, conventional
decoherence and quantum noise [9], [10].
methods often struggle to adapt to the dynamic nature
Decoherence refers to the loss of quantum coherence as
of quantum errors in large-scale, real-time quantum sys-
qubits interact with their environment, causing the deteri-
tems. Significant efforts have been made to overcome these
oration of superposition and entanglement—key resources
challenge, recently, Google Quantum AI team announced
for quantum computation [11]. Quantum noise, on the other
that their Willow processor is the first quantum processor
hand, includes various unwanted disturbances, such as bit-
ever, where error-corrected qubits improve exponentially as
flip and phase-flip errors, that can alter qubit states during
system size increases [1], this is a milestone in the road of de-
computation [12]. Without effective error correction, these
veloping large-scale realiable quantum computing system.
errors accumulate, rendering quantum computations unre-
Around same time, China launches new quantum proces-
liable. Quantum error correction (QEC) has been developed
sor called Zuchongzhi 3.0 [16], with 105 superconducting
as a critical framework to safeguard quantum information
qubits, and demonstrates its ability to achieve quantum
against such errors without directly measuring or disturbing
computational advantage by performing random circuit
the delicate states of quantum bits (qubits).
sampling significantly faster than the world’s most powerful
Conventional QEC methods, such as the Shor code [13]
classical supercomputers. These advancements clearly show
and surface codes [14], encode a logical qubit into multiple
potential of developing realiable and large-scale quantum
physical qubits to detect and correct errors. For example,
computing systems.
the Shor code encodes one logical qubit into nine physical
Machine learning (ML) techniques have emerged as
promising tools. Renowned for their capabilities in pattern
• Zihao Wang is with the School of Engineering and Applied Science, recognition [17], predictive modeling [18], and adaptive
University of Pennsylvania, Philadelphia 19104, PA, US. E-mail: zi-
[email protected] learning [19], ML offers significant potential to transform
• Hao Tang is with the School of Computer Science, Peking University, QEC. These techniques can automate the detection and cor-
Beijing 100871, China. E-mail: [email protected] rection of quantum errors, often with lower computational
∗ Corresponding author: Hao Tang. overhead compared to conventional methods. For instance,
 2

Fig. 1: Layout of this survey.

neural network decoders have demonstrated superior ac- The paper is organized as follows (see Fig.1): Section 2
curacy and speed over classical decoding algorithms in provides background information on quantum computing,
surface codes applications [20]. Reinforcement learning has and common types of errors, including bit-flip, phase-flip,
been used to optimize error correction protocols by training and depolarizing errors. It also reviews the principles and
agents to make optimal decisions in dynamic quantum limitations of conventional QEC methods and offers a brief
environments [21]. overview of common ML techniques. Section 3 delves into
However, integrating ML into QEC introduces its own the detailed workings of conventional QEC methods. Sec-
challenges. Key issues include scarcity of quantum error tion 4 presents analysis of the limitations of the conventional
data for training, scalability of ML models to handle large methods discussed in Section 3. Section 5 explores the
qubit systems, and ensuring stability of learning algorithms motivations for applying ML to QEC and reviews recent
in fluctuating quantum environments. Addressing these advancements. Section 6 describes evaluation metrics of ML
challenges is crucial for realizing full potential of ML in and conventional methods for QEC tasks, with a focus on
QEC. Therefore, there is a pressing need for a comprehen- performance metrics and scalability considerations. Section
sive survey that synthesizes recent developments, identifies 7 concludes the paper by summarizing existing AI-enabled
key trends, and outlines future research directions. methods, and analyzing future directions in the field.
In this paper, we aim to address this need by providing This paper seeks to bridge the gap between QEC and
an in-depth survey of the integration of ML techniques ML, offering new perspectives on enhancing quantum error
into QEC over recent years. We analyze how these modern correction techniques. Through a comprehensive review,
computational tools can enhance the reliability and per- we synthesize current research, identify key trends, and
formance of quantum computers, highlighting both their address emerging challenges, providing valuable insights
potential and limitations. Our survey stands out by not only for both researchers and practitioners. Our ultimate goal is
reviewing existing literature but also categorizing the stud- to contribute to the development of robust, fault-tolerant
ies based on the ML strategies employed and the resulting quantum computing. This paper is designed for readers
improvements in error correction performance. who may not have an extensive background in quantum
We begin by examining the fundamentals of both quan- mechanics or QEC. However, a basic understanding of
tum computing, QEC, and ML. This includes a summary of related mathematical equations and concepts is expected. By
quantum mechanisms, various types of quantum errors, and presenting the content, we aim to engage a wider audience
ML paradigms. Then we explore conventional approaches and promote interdisciplinary collaboration.
of QEC, such as stabilizer codes and topological codes,
offering concise explanations for readers who may not be 2 P RELIMINARIES
familiar with these concepts. Next, we focus on recent In this section, we present a brief introduction to the key
advancements in ML that have introduced innovative QEC topics and concepts, establishing a foundation for the sub-
strategies. sequent discussion on QEC.
 3

z z
z |ψ⟩

|ψ⟩ Z |ψ⟩
|ψ⟩
θ

ϕ y y y
x x x
X |ψ⟩

(1): Bloch sphere. (2): Bit-flip error. (3): Phase-flip error.

Fig. 2: Bloch sphere representations for a single qubit, highlighting quantum states and common quantum errors: (1) The
Bloch sphere representation of a qubit state |ψ⟩ defined by the polar angle θ and azimuthal angle ϕ in spherical coordinates,
visualizing the superposition of quantum states. (2) A bit-flip error caused by the Pauli-X operation, which flips the state
vector |ψ⟩ through the xy -plane, inverting the z -component to produce X |ψ⟩. (3) A phase-flip error caused by the Pauli-Z
operation, which flips the sign of the x- and y -components of the state vector |ψ⟩, altering the relative phase to produce
Z |ψ⟩. These visualizations provide insight into the geometric interpretation of quantum states and common errors in
quantum computation.

2.1 Quantum Computing When multiple qubits are involved, their combined state
Quantum computing is a rapidly emerging field that uti- is described by the tensor product of the individual qubit
lizes the principles of quantum mechanics to perform com- states. This results in an exponential increase in the number
putations that are beyond the reach of classical comput- of possible states a quantum computer can represent as the
ers. By harnessing phenomena such as superposition and number of qubits grows. This exponential scaling enables
entanglement, quantum computers have the potential to quantum computers to efficiently represent and manipulate
solve complex problems exponentially faster than their highly complex states, allowing them to perform computa-
classical counterparts [22]. However, the inherent fragility tions that are intractable for classical computers.
of quantum states, which are highly susceptible to errors Another uniquely quantum phenomenon is entangle-
caused by environmental noise and operational imperfec- ment, where qubits become correlated such that the state of
tions, presents a significant barrier to practical quantum one qubit cannot be described independently of the others,
computing. To overcome these challenges, Quantum Error even when they are physically separated. Entanglement
Correction (QEC) techniques have been developed to detect is essential for quantum algorithms and QEC because it
and correct errors, ensuring reliable and robust quantum creates correlations that are impossible to achieve with clas-
computations. sical systems. By distributing information across multiple
The fundamental unit of quantum information is quan- entangled qubits, QEC can detect and correct errors with-
tum bit, or qubit. Unlike a classical bit, which exists defini- out directly measuring the quantum state—a fundamental
tively as either 0 or 1, a qubit can exist in a superposition principle of its operation.
of both states simultaneously. Mathematically, the state of a Quantum computation is performed by manipulating
qubit is represented as: qubits using quantum gates, which serve as the quantum
counterparts to classical logic gates. These gates are repre-
|ψ⟩ = α|0⟩ + β|1⟩, (1) sented by unitary operators that evolve the state of qubits
where |0⟩ and |1⟩ are the computational basis states, and in a reversible manner. Single-qubit gates, such as the Pauli-
α and β are complex probability amplitudes that satisfy X, Pauli-Y, Pauli-Z, and Hadamard gates, apply specific
the normalization condition |α|2 + |β|2 = 1. Superposition transformations to a qubit’s state. For instance, the Pauli-
allows the qubit to represent multiple possibilities at once, X gate flips the state of a qubit, similar to the function of a
enabling parallelism in quantum algorithms. For example, a classical NOT gate:
qubit in the state |ψ⟩ = √12 (|0⟩ + |1⟩) has equal probability 
0 1

amplitudes for both |0⟩ and |1⟩, effectively allowing it to X= , (3)
1 0
represent both states.
The state of a qubit can be visualized on the Bloch while the Hadamard gate creates superposition states by
sphere, as shown in Figure 2. Here, the qubit state |ψ⟩ transforming basis states into equal superpositions:
is represented as a point on the surface of a unit sphere,
 
1 1 1
parameterized by the angles θ and ϕ: H=√ . (4)
    2 1 −1
θ iϕ θ Two-qubit gates, such as the Controlled-NOT (CNOT)
|ψ⟩ = cos |0⟩ + e sin |1⟩. (2)
2 2 gate, facilitate interactions between qubits and are crucial
This geometric representation provides intuitive insights for generating entangled states. The CNOT gate flips the
into the effects of quantum gates and errors on qubit states. state of the target qubit if the control qubit is in the state |1⟩.
 4

TABLE 1: Milestones in quantum technology.

Year Contributor(s) Contribution

1900 Max Planck [23] Introduced quantum theory; energy is quantized.
1905 Albert Einstein [24] Explained photoelectric effect with photons.
1913 Niels Bohr [25] Proposed Bohr atomic model using quantum concepts.
1925 Werner Heisenberg [26] Formulated matrix mechanics in quantum theory.
1926 Erwin Schrödinger [27] Developed wave mechanics; Schrödinger equation.
1927 Heisenberg [28] Proposed the Uncertainty Principle.
1935 Einstein, Podolsky, Rosen [29] Presented EPR paradox on quantum completeness.
1964 John Bell [30] Formulated Bell’s theorem on entanglement.
1982 Alain Aspect et al. [31] Confirmed quantum entanglement experimentally.
1984 Bennett, Brassard [4] Introduced BB84 quantum cryptography protocol.
1994 Peter Shor [5] Developed Shor’s factoring algorithm.
1997 Bouwmeester et al. [32] Achieved experimental quantum teleportation.
2001 IBM [33] Demonstrated Shor’s algorithm on 7-qubit computer.
2011 D-Wave Systems [34] Announced first commercial quantum computer.
2019 Google Quantum AI [35] Achieved quantum supremacy with Sycamore processor.
2022 Aspect, Clauser, Zeilinger [36] Won Nobel Prize for entanglement experiments.
2023 Atom Computing, IBM [37], [38] Both developed 1,000-qubit quantum computers.
2024 Google Quantum AI [1] Willow processor with exponentially decreasing error rates.

This gate plays a central role in implementing entanglement information about the system without disrupting the com-
and QEC protocols: putation.
  A fundamental challenge in quantum systems is the
1 0 0 0 no-cloning theorem, which states that it is impossible to
0 1 0 0
CNOT =  create an identical copy of an arbitrary unknown quantum
0 0 0 1 . (5)

state [40]. This theorem complicates error detection and
0 0 1 0 correction, as it prevents the straightforward duplication of
Quantum circuits, which represent quantum computa- quantum information. Addressing these challenges requires
tions, consist of a sequence of quantum gates applied to the development of QEC codes that can detect and correct
qubits. The circuit progresses from left to right, indicating errors without collapsing the quantum state. These codes are
the order of operations. These circuits are essential for im- crucial for achieving fault-tolerant quantum computation,
plementing quantum algorithms, such as Shor’s algorithm enabling reliable results even in the presence of errors [41].
for integer factorization [39] and Grover’s algorithm for Table 1 shows milestones in the history of quantum technol-
database search [3]. By manipulating qubits through quan- ogy development.
tum gates and circuits, quantum computers can perform
complex computations exponentially faster than classical 2.2 Quantum Errors and Noise
computers.
Measurement in quantum mechanics is the process by Quantum computers, despite their immense potential, are
which a quantum state collapses to a definite outcome, highly vulnerable to errors due to the fragile nature of
yielding classical information from the quantum system. For quantum states. These errors can originate from various
a qubit in the state |ψ⟩ = α |0⟩ + β |1⟩, a measurement in the sources, such as interactions with the environment, imper-
computational basis {|0⟩ , |1⟩} will result in either outcome fections in quantum gates, and decoherence. A thorough
|0⟩ with probability |α|2 , or outcome |1⟩ with probability understanding of the types of quantum errors and their
|β|2 . After measurement, the qubit collapses to the mea- causes is crucial for designing effective QEC strategies.
sured state, and any prior superposition or entanglement Quantum errors can be broadly categorized based on
is lost. how they affect qubit states. One of the most common types
Measurements can also be performed in different bases, of errors is the bit-flip error, which occurs when a qubit’s
such as the Hadamard basis, where the qubit’s state is state flips from |0⟩ to |1⟩ or vice versa. This is analogous
expressed as a superposition of |+⟩ = √12 (|0⟩ + |1⟩) and to a classical bit-flip and can be represented by the Pauli-X
operator (X ), which acts on a qubit as follows:
|−⟩ = √12 (|0⟩ − |1⟩). In multi-qubit systems, measurements
can project qubits onto entangled states, a critical operation X |0⟩ = |1⟩ , X |1⟩ = |0⟩ . (6)
in QEC protocols. One of the main challenges in quantum
measurement is its destructive nature—once a quantum Thus, for a qubit in a general state |ψ⟩ = α |0⟩ + β |1⟩, a
state is measured, it collapses, making it difficult to extract bit-flip error transforms it to X |ψ⟩ = α |1⟩ + β |0⟩.
 5

Figure 2 illustrates the impact of a bit-flip error on the 2.3.1 Overview of AI and ML
Bloch sphere representation of a qubit. Another important AI encompasses a wide range of techniques, from conven-
type of error is the phase-flip error, which modifies the tional rule-based systems to advanced deep learning archi-
relative phase between the components of a qubit’s super- tectures. A key branch of AI, machine learning (ML), focuses
position while leaving the probabilities of the computational on algorithms that enable computers to learn patterns from
basis states unchanged. This error is represented by the data and make predictions or decisions without explicit pro-
Pauli-Z operator (Z ): gramming for each task [62], [63]. ML algorithms develop
Z |0⟩ = |0⟩ , Z |1⟩ = − |1⟩ . (7) models based on training data and use these models to make
inferences about unseen data.
For a qubit in superposition, such as |ψ⟩ = α |0⟩ + β |1⟩, a
phase-flip error changes the state to: 2.3.2 Machine Learning Paradigms
Machine learning can be categorized into several
Z |ψ⟩ = α |0⟩ − β |1⟩ . (8)
paradigms, each suited to different types of tasks.
Phase-flip errors are uniquely quantum mechanical and In supervised learning, the algorithm is trained on
have no classical analog. labeled data, where each input is paired with a known
A more general type of error is the depolarizing error, corresponding output. The goal is to learn a mapping from
which models the scenario where a qubit’s state becomes a inputs to outputs that generalizes well to unseen data [63],
completely mixed state with some probability. Depolarizing [64], [65]. Supervised learning has been widely applied to
errors occur due to the random application of bit-flip, phase- various classification tasks, the objective is to minimize the
flip, or both errors, resulting in a loss of coherence and difference between predicted labels and actual ones, often
a transition to a maximally mixed state. The depolarizing quantified using a loss function such as cross-entropy or
channel is described as: mean squared error [66]. Popular supervised learning algo-
p rithms include decision trees [67], support vector machines
ρ → (1 − p)ρ + (XρX † + Y ρY † + ZρZ † ), (9) (SVMs) [68], and neural networks [69].
3
Unsupervised learning works with data that does not
where ρ is the qubit’s density matrix, p is the depolarizing
have labeled outputs. Its aim is to uncover hidden structures
probability, and X , Y , and Z are the Pauli operators. Depo-
or patterns within the input data [63], [66], [70]. Common
larizing errors are especially important in practical quantum
applications include clustering, where similar data points
systems where various types of noise can act on qubits
are grouped, and dimensionality reduction, where complex
simultaneously.
datasets are simplified while retaining key features [71],
Quantum errors originate from various physical pro-
[72]. In quantum computing, unsupervised learning can be
cesses that disrupt qubits, significantly affecting the relia-
valuable for identifying correlations in quantum noise that
bility of quantum computations. One of the most critical
traditional methods might overlook [73], [74].
sources of quantum errors is decoherence, which arises
Reinforcement learning (RL) involves an agent that
from interactions between qubits and their surrounding
learns by interacting with its environment, receiving feed-
environment. These interactions gradually lead to the loss
back in the form of rewards or penalties, and adjusting its
of quantum coherence, causing the decay of superposition
actions to maximize cumulative rewards over time [75], [76].
and entanglement—two fundamental properties essential
RL is particularly effective in sequential decision-making
for quantum computations [42]. Another common source
problems, where the agent must develop an optimal strat-
of quantum errors is gate imperfections, which result from
egy through trial and error. It has been successfully applied
inaccuracies in the control or calibration of quantum gates.
in various domains, including robotics, game playing, and
These imperfections cause deviations from the intended op-
autonomous systems [77], [78]. In the context of QEC, RL
erations and can accumulate over time, reducing the overall
has been proposed as a method for dynamically adapting
fidelity of quantum computations [43]. Additional contribu-
error correction protocols to the evolving error landscape
tors to quantum errors include measurement errors; thermal
of quantum systems [21], [79]. By continuously refining its
noise, arising from environmental fluctuations around the
strategy, an RL agent can devise more effective methods for
qubits; and crosstalk, referring to unintended interactions
minimizing errors in quantum computations, making RL
between qubits or control lines [44], [45], [46]. Effectively
a promising tool for enhancing the reliability of quantum
addressing these error sources is essential for ensuring the
technologies.
reliability and scalability of quantum computing systems.
Semi-supervised learning bridges the gap between su-
pervised and unsupervised learning by utilizing both la-
2.3 AI and ML beled and unlabeled data [80], [81]. This approach is es-
AI and ML are subfields of computer science dedicated pecially valuable when labeling data is expensive or time-
to developing systems that can perform tasks typically re- consuming, but large amounts of unlabeled data are readily
quiring human intelligence. These technologies enable com- available. In the context of quantum error correction, semi-
puters to learn from data, uncover complex relationships, supervised learning can enhance error identification models
and enhance performance over time with minimal human by leveraging a small set of labeled quantum states along-
intervention [62]. In the context of QEC, AI and ML provide side a larger pool of unlabeled states [82]. Techniques such
powerful tools to optimize error correction protocols, en- as self-training, label propagation, and graph-based meth-
hance accuracy, and reduce computational overhead. Table 2 ods enable the model to infer patterns from the unlabeled
highlights key milestones in the evolution of AI and ML. data, thereby improving predictive accuracy [83], [84].
 6

TABLE 2: Milestones in AI and ML.

Year Contributor(s) Contribution

1950 Turing [47] Proposed the Turing Test to assess machine intelligence.
1956 McCarthy et al. Established AI as a field at the Dartmouth Conference.
1986 Rumelhart et al. [48] Developed backpropagation for training neural networks.
1997 IBM Deep Blue [49] Deep Blue defeated world chess champion Kasparov.
2006 Hinton et al. [50] Introduced deep belief networks, sparking deep learning.
2012 Krizhevsky et al. [51] Developed AlexNet, a breakthrough in image classification.
2014 Goodfellow et al. [52] Introduced Generative Adversarial Networks (GANs).
2016 Silver et al. [53] Created AlphaGo, which defeated a champion Go player.
2017 Vaswani et al. [54] Introduced Transformers, revolutionizing NLP.
2018 Devlin et al. [55] Released BERT, advancing NLP through pre-training.
2019 Brown et al. [56] Developed GPT-2 showing large-scale language generation.
2020 Jumper et al. [57] Released AlphaFold 2 predicting protein structures.
2021 Ramesh et al. [58] Introduced DALL·E, generating images from text prompts.
2022 OpenAI [59] Released ChatGPT, a conversational AI.
2023 OpenAI [60] Released GPT-4, a multimodal AI model.
2024 Google DeepMind [61] Released AlphaFold 3 with high prediction accuracy.

Figure 3 illustrates the mechanisms of the four ML Training machine learning models, however, presents
paradigms discussed. challenges such as overfitting, where the model becomes
overly specialized to the training data and performs poorly
2.3.3 Neural Networks and Deep Learning on unseen data. Regularization techniques, such as L1 and
Neural networks are a class of machine learning models L2 regularization or dropout, are effective in mitigating
inspired by the structure and function of the human brain. overfitting [91]. Additionally, normalization techniques like
A neural network consists of layers of interconnected nodes batch normalization stabilize and accelerate the training
(neurons), where each neuron processes input data and process by maintaining consistent input scales throughout
passes the result to subsequent layers [63], [64], [85]. The the network [92]. In deep learning, issues such as vanishing
network adjusts its internal parameters, or weights, through or exploding gradients can hinder effective learning. These
a process called backpropagation, which minimizes the problems arise when gradients become excessively small
error between the predicted output and the actual target or large, making parameter updates ineffective. Advanced
[85], [86]. Deep learning refers to neural networks with optimization techniques and architectural innovations, such
many layers, also known as deep neural networks (DNNs). as residual networks, have been developed to address these
These models are capable of learning complex hierarchical challenges and ensure stable training [93].
representations of data, making them particularly effective
for tasks such as image recognition, natural language pro-
cessing, and game playing [69], [77], [78]. 2.4 Mathematical Notations and Conventions
Different neural network architectures are suited for 2.4.1 Quantum Mechanics Notations
different types of tasks. For example, Convolutional Neural
We use Dirac’s bra-ket notation to represent quantum states.
Networks (CNNs) are highly effective at processing grid-
A column vector representing a quantum state is denoted
like data, such as images, while Recurrent Neural Networks
as a ket |ψ⟩, while a row vector, the conjugate transpose of a
(RNNs) are designed to handle sequential data, making
ket, is denoted as a bra ⟨ψ|. The computational basis states
them ideal for tasks involving time-series or temporal de-
for a qubit are denoted by |0⟩ and |1⟩, and a general qubit
pendencies [87], [88], [89], [90].
state is expressed as |ψ⟩ = α |0⟩ + β |1⟩, where α, β ∈ C
2.3.4 Training Machine Learning Models and |α|2 + |β|2 = 1. For systems with multiple qubits,
we use the tensor product, denoted by ⊗, to represent
Training a machine learning model involves adjusting its
combined states. For example, the state of two qubits is
parameters to minimize a loss function, which quantifies
written as |ψ⟩AB = |ψ⟩A ⊗ |ψ⟩B , representing the joint
the error between the model’s predictions and the actual
quantum state of the two systems. Quantum operators are
targets. In neural networks, optimization algorithms such as
denoted by uppercase letters and often represent actions
gradient descent and its more advanced variants, including
such as transformations or measurements on qubit states.
stochastic gradient descent (SGD) and Adam, are commonly
For example, the Pauli operators, which describe bit-flip and
used to minimize the loss function [63]. Backpropagation
phase-flip operations, are given as:
computes the gradient of the loss function with respect to
the network’s parameters, enabling parameter updates that
     
0 1 0 −i 1 0
reduce the error. X= , Y = , Z= . (10)
1 0 i 0 0 −1
 7

(1) Supervised Learning (2) Unsupervised Learning

(3) Semi-supervised Learning (4) Reinforcement Learning

Fig. 3: Overview of Machine Learning Paradigms: (1) Supervised Learning. (2) Unsupervised Learning. (3) Semi-supervised
Learning. (4) Reinforcement Learning. These paradigms represent the core methodologies used in modern machine learning
to address diverse real-world problems.

These operators act on single qubits, but operators can be example, f (·), which denotes a function applied to some
extended to multiple qubits using the tensor product. The input.
identity operator on a single qubit is denoted as I , where
  We represent datasets as D = {(xi , yi )}Ni=1 , where xi
1 0 represents the input data, yi is the corresponding label or
I= . (11)
0 1 output, and N is the number of samples in the dataset.
In addition, mixed quantum states, which describe prob-
The loss function used to train ML models is denoted
abilistic mixtures of quantum states, are represented by
as L(θ), where θ represents the parameters of the model.
density matrices, denoted by ρ. For a pure state |ψ⟩, the cor-
The gradient of the loss function with respect to the model
responding density matrix is ρ = |ψ⟩ ⟨ψ|, and this formalism
parameters is written as ∇θ L, and is used in optimization
is used to handle systems where we do not have complete
algorithms to update the parameters and minimize the loss
information about the quantum state. The expectation value
during training. In this paper, we focus on optimizing ma-
of an operator O in a quantum state |ψ⟩ is denoted as
chine learning models to improve quantum error correction
⟨O⟩ = ⟨ψ| O |ψ⟩, and represents the average value of an
performance by minimizing the loss function L(θ).
observable in that state.
We denote probability distributions by P (·) or p(·). For
2.4.2 ML Notations example, p(y|x) represents the conditional probability of the
In ML, we denote vectors by bold lowercase letters, such output y given the input data x. The expectation value of a
as x, and matrices by bold uppercase letters, such as W. random variable X is represented as E[X], which is used
Functions are generally represented by lowercase letters, for when discussing probabilistic models in ML.
 8

3 C ONVENTIONAL QEC M ETHODS achieved by measuring stabilizer generators, which identify

Conventional QEC methods encode logical quantum infor- the presence and location of errors. Once the syndrome is
mation into entangled states of multiple physical qubits. determined, the error is corrected using the corresponding
The approach enables the detection and correction of er- quantum operation [41].
rors without directly measuring the quantum information Surface codes are a class of topological QEC codes that
itself. This section offers a detailed overview of coventional have gained significant attention for their scalability and
QEC techniques, covering commonly used error correction potential for practical implementation. Proposed by Alexei
codes, stabilizer formalism, and the processes of syndrome Kitaev in 2003 [97], surface codes encode logical qubits into
measurement and decoding. a two-dimensional lattice of physical qubits arranged on
a surface, with qubits interacting locally through nearest-
neighbor connections. This locality makes surface codes
3.1 Conventional QEC Codes particularly well-suited for large-scale quantum computing,
Several QEC codes have been developed to safeguard quan- as physical qubits only need to interact with their immediate
tum information from errors caused by decoherence, noise, neighbors, simplifying hardware requirements.
and imperfections in quantum gates. These codes are es- One of the key advantages of surface codes is their
sential for achieving fault-tolerant quantum computation. relatively high error threshold, approximately 1%, meaning
Due to space constraints, we focus on three key QEC codes: they can tolerate error rates of up to 1% before the error
Shor’s code, Steane’s code, and surface codes, highlighting correction process becomes unreliable [14]. This robustness
their encoding processes, error correction mechanisms, and makes surface codes highly resilient to errors, especially
practical applications. compared to other codes that require significantly lower
The Shor code, introduced by Peter Shor in 1995 [13], error rates. In surface codes, errors appear as defects in the
was the first QEC code. It encodes one logical qubit into qubit lattice, detected through stabilizer measurements. The
nine physical qubits and can correct arbitrary single-qubit error correction process involves identifying and correcting
errors, including both bit-flip and phase-flip errors. The Shor chains of errors, often using classical algorithms such as
code achieves this by combining two classical repetition minimum-weight perfect matching [14], which efficiently
codes: one for correcting bit-flip errors and another for determine the most likely error path connecting observed
correcting phase-flip errors. The logical qubit state |ψL ⟩ = defects.
α |0L ⟩ + β |1L ⟩ is initially encoded using a three-qubit rep- Another significant advantage of surface codes is their
etition code for bit-flip protection, where |0L ⟩ = |000⟩ and scalability. The code distance, which determines the number
|1L ⟩ = |111⟩. To protect against phase-flip errors, Hadamard of errors a code can detect and correct, increases with the
gates are applied to each qubit, and the resulting state is size of the lattice. This means error correction capabilities
further encoded with another three-qubit repetition code. can be enhanced by simply adding more physical qubits to
This process results in a final nine-qubit encoded state the surface. Due to their scalability, robustness, and compat-
capable of correcting any single-qubit error. ibility with current quantum hardware platforms such as
Error correction in the Shor code is performed by mea- superconducting qubits [46], [98].
suring stabilizer generators, which are specific quantum
operators designed to detect errors without disturbing the 3.2 Stabilizer Formalism
encoded quantum information. These stabilizer measure-
The stabilizer formalism, developed by Daniel Gottes-
ments produce syndromes that indicate the presence of
man [41], is a mathematical framework that describes a large
either bit-flip or phase-flip errors. Once an error is identi-
class of QEC codes, including the Steane and surface codes.
fied, appropriate quantum gates are applied to correct it,
This formalism is based on stabilizer groups, which are
restoring the logical qubit [41].
abelian subgroups of the n-qubit Pauli group Pn . In the con-
The Steane code, proposed by Andrew Steane in
text of quantum computing, the Pauli group consists of all
1996 [94], is a seven-qubit code derived from the classical
n-fold tensor products of the Pauli matrices {I, X, Y, Z} and
[7, 4, 3] Hamming code. It is a prominent example of a CSS
the multiplicative factors {±1, ±i}. These operators form
(Calderbank-Shor-Steane) code, named after the developers
the basis for describing quantum errors and operations.
of this class of codes [95], [96]. CSS codes are stabilizer
A stabilizer code is defined by a set of k independent
codes constructed from classical linear codes that satisfy
stabilizer generators {S1 , S2 , . . . , Sk }, where each Si ∈ Pn
dual-containing properties, enabling separate correction of
satisfies the relation:
bit-flip and phase-flip errors. The logical qubit is encoded
using codewords derived from the Hamming code: Si |ψL ⟩ = |ψL ⟩ , ∀i, (13)
1 X 1 X
for any logical codeword |ψL ⟩. The code space is the simul-
|0L ⟩ = √ |x⟩ , |1L ⟩ = √ |x⟩ . (12)
8 x∈C 8 x∈C ⊥ \C taneous +1 eigenspace of all stabilizer generators.
The stabilizer formalism simplifies the description and
where C is the classical Hamming code, and C ⊥ is its dual analysis of quantum codes by expressing their properties
code. The codewords |x⟩ correspond to computational basis in terms of stabilizer generators. This approach provides
states representing the classical codewords. Steane’s code of- a unified framework for understanding and implementing
fers the advantage of correcting any single-qubit error while various QEC codes, such as the Steane code and surface
using fewer physical qubits than the Shor code, making it codes. It also facilitates the design of encoding and decod-
more resource-efficient. Error correction in Steane’s code is ing circuits, allowing these codes to detect errors through
 9

stabilizer measurements without collapsing the encoded for large-scale quantum systems where exact decoding is
quantum information. This ensures the qubits’ state remains computationally infeasible. The iterative nature of the algo-
within the valid code space. rithm enables it to converge on a probable error estimate
One of the key advantages of the stabilizer formalism is by leveraging local information, making it computationally
its ability to perform QEC while preserving the integrity of efficient.
logical qubits. The stabilizer generators function as checks However, belief propagation in QEC faces challenges
on the quantum system, ensuring it remains in a valid due to quantum degeneracy, where multiple error patterns
code state. When errors occur, deviations from the expected can produce the same syndrome [100]. To address this issue,
stabilizer measurement results flag the errors. This enables modified approaches such as the quantum belief propa-
the system to identify and correct errors while keeping the gation algorithm incorporate degeneracy into the message
logical qubits unaffected. updates.
The stabilizer formalism, combined with efficient decod-
3.3 Syndrome Measurement and Decoding ing algorithms, forms the foundation of QEC, enabling fault-
Syndrome measurement is the process of detecting quan- tolerant quantum computation.
tum errors by measuring the stabilizer generators. These
measurements produce syndromes, which reveal informa-
4 L IMITATIONS OF C ONVENTIONAL QEC M ETH -
tion about the type and location of errors without disturb-
ing the quantum state. A syndrome is essentially a set of ODS
measurement outcomes indicating how the quantum state Conventional QEC methods have played a crucial role
has deviated from the code space. The process involves the in safeguarding quantum information. However, they face
following steps: several significant limitations that hinder their practical
1) Stabilizer Measurements: Ancillary qubits are used to implementation in large-scale quantum computing systems.
measure the stabilizers. Controlled operations entangle This section highlights the primary challenges associated
the data qubits with the ancillary qubits, which are then with these methods.
measured to extract the syndrome bits.
2) Syndrome Extraction: The extracted syndrome bits 4.1 High Resource Overhead
form a pattern that identifies the presence and type of
A major limitation of conventional QEC methods is the
error, if any.
significant resource overhead required to encode a single
3) Decoding: A decoding algorithm maps the syndrome to
logical qubit. These codes typically require a large number
the corresponding error and determines the appropriate
of physical qubits, placing substantial demands on current
correction operation.
quantum hardware. Achieving logical error rates suitable
4) Error Correction: The identified error is corrected us-
for practical computations can necessitate thousands of
ing quantum gates, restoring the logical qubit to its
physical qubits per logical qubit.
intended state.
The challenges associated with high resource overhead
Decoding algorithms are essential for interpreting syn- are twofold: (1) Hardware Limitations: Current quantum
dromes and applying corrections. For smaller codes, pre- hardware can accommodate only a limited number of
computed lookup tables can be used to map each syndrome qubits, making it difficult to scale systems to the required
to its corresponding correction operation. While efficient size. (2) Increased Control Complexity: As the number of
for small systems, lookup tables become impractical for qubits increases, the control complexity grows, raising the
larger codes due to the exponential growth in the number risk of introducing additional errors during operation.
of possible syndromes. For larger codes, such as surface
codes, more advanced techniques like minimum-weight
perfect matching are employed. This classical algorithm 4.2 Complexity of Decoding Algorithms
identifies the most probable error chain connecting observed Conventional decoding algorithms often involve signifi-
defects in the qubit lattice, efficiently decoding the error pat- cant computational overhead, especially for large codes.
tern [14]. Minimum-weight perfect matching is particularly For instance, minimum-weight perfect matching, commonly
well-suited for surface codes because it leverages the local used in surface codes, has a polynomial time complexity
interactions between qubits in the lattice. that scales with the number of qubits [14]. While efficient
Another important approach is belief propagation, a for moderate-sized systems, this algorithm becomes com-
message-passing algorithm used in probabilistic graphical putationally intensive for large-scale systems, particularly
models such as factor graphs and Bayesian networks [99]. when real-time decoding is required. Lookup tables provide
In the context of QEC, belief propagation operates on the an efficient decoding solution for small codes but quickly
Tanner graph representation of a quantum code, where become impractical as system size increases due to the
nodes represent qubits and checks (stabilizers), and edges exponential growth in the number of possible syndromes.
represent their relationships [100]. The algorithm estimates Probabilistic methods, such as belief propagation, can offer
the likelihood of various error patterns by iteratively updat- faster performance but may fail to guarantee convergence
ing and passing probability-based messages along the edges or optimal decoding in complex error landscapes [100].
of the graph, informed by the observed syndromes. This computational complexity poses a significant chal-
Belief propagation is particularly effective for quan- lenge for implementing QEC in real-time quantum comput-
tum codes with sparse parity-check matrices, such as low- ing, where rapid error correction is crucial to maintaining
density parity-check (LDPC) codes [101]. It is well-suited the fidelity of logical qubits.
 10

4.3 Stringent Error Threshold Requirements shorter than the coherence time of the qubits. However, de-
Conventional QEC codes require that physical qubit error coding algorithms become more computationally demand-
rates remain below a specific threshold to function effec- ing with larger systems, making it increasingly difficult to
tively. And quantum computations can be made arbitrarily meet this requirement.
reliable as long as the error rate per operation stays below
a certain critical value [102]. However, achieving physical
error rates below this threshold with large number of qubits 4.6 Limited Adaptability to Dynamic Environments
remains a significant challenge for current quantum hard-
Traditional QEC methods are typically designed based on
ware due to issues such as decoherence, gate imperfections,
specific error models and struggle to adapt to dynamic
and measurement errors.
or complex error environments. Most conventional codes
In practice, as physical error rates approach the thresh-
assume that errors are independent and identically dis-
old, the resource overhead required for error correction
tributed, an assumption that often does not hold in practical
increases dramatically. This results in higher demands for
quantum systems. In reality, errors can be correlated due to
qubits and operational complexity.
environmental factors or qubit crosstalk, making them more
challenging to detect and correct using traditional methods.
4.4 Implementation Challenges
Moreover, fluctuations in temperature, magnetic fields,
Implementing conventional QEC methods on physical or other environmental conditions can dynamically alter
quantum hardware presents several technical challenges. error characteristics, further complicating error correction.
High-fidelity control over quantum gates and measure- Conventional QEC methods are generally not equipped to
ments is essential for effective error correction, yet achieving handle these real-world variations, limiting their effective-
consistent precision across all qubits in a system is extremely ness in practical quantum systems, especially in environ-
difficult. Errors can propagate during the correction process ments where error rates and types fluctuate over time.
itself, with incorrect operations potentially introducing log-
ical errors that spread throughout the system.
Moreover, measurement errors in the ancillary qubits
4.7 Difficulty in Handling Complex Error Models
used for syndrome extraction can corrupt the syndrome
data, resulting in incorrect corrections. Timing and synchro- Quantum systems are subject to a wide range of error
nization across qubits also pose significant challenges in types that extend beyond simple bit-flip or phase-flip errors,
large systems, where precise coordination is necessary to en- posing significant challenges for conventional QEC codes.
sure operations are completed within the qubits’ coherence Two notable examples of such complex error models are
time. As the size of the quantum system increases, these non-Markovian noise and leakage errors.
implementation challenges become even more pronounced, Non-Markovian noise refers to noise processes where a
complicating the realization of effective QEC at scale. system’s evolution depends on its history, exhibiting mem-
Recent advancements in quantum hardware, such as the ory effects [107]. Unlike Markovian noise, which assumes
development of superconducting qubits with longer coher- the environment has no memory of past interactions, non-
ence times and higher gate fidelities [103], [104], have par- Markovian noise results in correlated errors that standard
tially alleviated some of these challenges. Additionally, new QEC codes, designed under the assumption of independent
error correction codes, such as quantum low-density parity- errors, cannot effectively address [108]. These correlations
check (LDPC) codes [105], [106], show promise in reducing can cause errors to spread across multiple qubits in un-
resource overhead and improving scalability. However, in- predictable ways, reducing the overall effectiveness of error
tegrating these codes into practical systems requires further correction.
research into efficient decoding algorithms and ensuring Leakage errors occur when qubits transition out of their
compatibility with existing hardware. computational subspace into higher energy levels or other
states not included in the qubit’s Hilbert space [109]. This is-
4.5 Scalability Issues sue is particularly relevant in systems like superconducting
Scaling QEC methods to large numbers of qubits presents qubits and trapped ions, where qubits are implemented us-
numerous challenges, primarily arising from resource de- ing multi-level quantum systems. Leakage errors are prob-
mands and hardware limitations. As the size of a quantum lematic because they can propagate during gate operations
system increases, so does the resource overhead. For in- and are not detectable by standard stabilizer measurements,
stance, surface codes require a number of physical qubits which assume qubits remain within the computational sub-
that scales quadratically with the code distance, making space.
large-scale implementations highly resource-intensive [14]. Addressing these complex error models requires ad-
Furthermore, hardware limitations such as qubit co- vanced error correction strategies capable of detecting and
herence times, error rates, and qubit connectivity become correcting correlated and non-Pauli errors. Techniques such
increasingly significant as system size grows. Physical sys- as leakage reduction units [110], dynamical decoupling
tems may not scale favorably in these aspects, making it [111], and designing codes resilient to correlated noise [112]
challenging to maintain the performance levels necessary have been proposed. However, integrating these methods
for large-scale fault-tolerant quantum computation. with conventional QEC codes introduces additional im-
Another critical challenge is decoding latency: as the sys- plementation complexity and may be incompatible with
tem expands, the time required for decoding must remain existing decoding algorithms.
 11

5 A RTIFICIAL I NTELLIGENCE FOR QEC system, optimizing protocols as conditions evolve [79]. For
Recent advancements in quantum computing have high- instance, RL algorithms have been used to develop adap-
lighted the critical need for efficient and scalable QEC tive decoders that adjust to varying noise levels and error
methods. AI and ML techniques have emerged as promising correlations, thereby enhancing overall error correction per-
solutions to these challenges. Deep neural networks, in formance [116].
particular, have been applied to decode surface codes more Moreover, ML algorithms can be designed to be re-
efficiently. Deep learning models also outperforms conven- silient to noise in syndrome measurements and ancillary
tional decoders in terms of both accuracy and speed [113], qubits—additional qubits used for error detection and cor-
[114]. These methods specifically shows potential in ad- rection. This noise resilience further improves the effec-
dressing correlated errors—errors that are interdependent, tiveness of QEC, particularly in Noisy Intermediate-Scale
commonly found in quantum systems and notoriously dif- Quantum (NISQ) devices, where error rates remain rel-
ficult to manage with traditional QEC methods. atively high [117]. By continuously learning and adapt-
Figure 4 illustrates how AI tools can enhance each step ing, ML-based QEC methods maintain high performance
of the QEC workflow. despite hardware imperfections and environmental distur-
bances [118], [119].

5.1 Improving Decoding Efficiency

Conventional decoding algorithms, which often rely on 5.3 Facilitating Complex Error Modeling
heuristic or exhaustive search methods, scale poorly as sys-
tem size increases, creating significant computational bot- Modeling errors in quantum systems is a challenging task,
tlenecks. These bottlenecks hinder real-time error correction especially when dealing with non-Pauli errors and non-
and disrupt the delicate balance between timely feedback Markovian noise. Non-Pauli errors are those that cannot be
and maintaining high-fidelity qubit operations. To prevent fully described using the Pauli matrices (X, Y, Z operations),
errors from propagating and degrading overall system per- while non-Markovian noise refers to noise processes where
formance, efficient, low-latency decoders are essential. a system’s evolution depends on its history rather than
ML methods, particularly those utilizing deep learning solely its current state [120]. These complex error types are
architectures, have shown great potential in overcoming difficult for conventional QEC methods, which often rely on
these computational challenges. For instance, CNNs, which simplifying assumptions that fail in realistic scenarios.
excel at processing structured data on grids, have been suc- ML techniques excel at building models from data, mak-
cessfully applied to QEC decoding tasks [115]. By training ing them particularly suited for complex error modeling.
on large datasets of simulated error syndromes derived By leveraging experimental datasets, AI-driven models can
from realistic noise models, CNN-based decoders learn to capture intricate error dynamics without requiring oversim-
map syndromes directly to corrective operations. This ap- plified assumptions [121], [122]. Deep learning methods,
proach efficiently identifies spatial correlations and complex such as RNNs, are especially effective in this context. RNNs
patterns in the data that traditional decoders may overlook. are designed to identify patterns in sequential data, mak-
Once trained, CNNs can perform inference rapidly, reduc- ing them ideal for capturing temporal dependencies in er-
ing decoding times by orders of magnitude compared to rors—an essential capability for addressing non-Markovian
classical methods. Such speed-ups are critical for enabling noise.
real-time error correction in near-term quantum devices. Additionally, ML models can predict error occurrences
Beyond speed and scalability, ML-based decoders offer a and system degradation by analyzing trends in the data,
flexibility that many traditional methods lack. As quantum enabling predictive maintenance and proactive error cor-
systems expand, classical decoding solutions often require rection. This predictive capability enhances the stability of
exponentially increasing computational resources, making quantum computations over time, significantly improving
them impractical for large-scale architectures. In contrast, the overall reliability of quantum processors [123]. Recent
ML decoders can be retrained or fine-tuned to accommodate research has shown that ML models trained on experimental
new error sources, different quantum codes, or heteroge- data can predict and correct errors in real quantum de-
neous device layouts with relatively minimal additional vices with greater accuracy than traditional decoders. For
effort. instance, ML techniques have been applied to IBM’s quan-
tum processors to identify and correct non-standard error
patterns, offering a more nuanced and effective approach to
5.2 Enhancing Robustness and Adaptability
QEC [124].
Quantum systems are inherently dynamic, with error rates
and types varying over time due to environmental fluctua-
tions and system imperfections. AI and ML techniques en- 5.4 Existing Works
hance the robustness of QEC by generalizing from training
data to unseen error patterns, improving reliability even in Significant research efforts have focused on the integration
changing environments. of AI and ML techniques into QEC. This section presents
RL agents, which learn to make decisions by interacting an overview of existing studies, organized by the type of
with an environment, have shown significant promise in ML approach used, and highlights recent examples that
QEC. RL agents can be trained to adapt error correction demonstrate how these techniques have improved error
strategies based on real-time feedback from the quantum correction performance.
 12

Fig. 4: The entire workflow of quantum computing system, and potential improvement by AI tools. Please note that this
area is growing fast and the graph is not comprehensive.

5.4.1 Supervised Learning Approaches of quantum codes and address practical error correction
Supervised learning trains models on labeled datasets to challenges.
predict correction operations from given error syndromes. Recurrent Neural Networks (RNNs), designed to pro-
In the context of QEC, these datasets are typically generated cess sequential data, have been instrumental in adapting
by simulating quantum codes under various noise models. QEC decoders to dynamic environments. Varsamopoulos
The data is labeled using known decoding algorithms or et al. [121] demonstrated that RNNs could track evolving
exact solutions for small codes, enabling the supervised error patterns, enabling models to adapt to changing noise
learning models to generalize error correction strategies. profiles. Convy et al. [126] proposed a novel RNN-based
Feedforward Neural Networks (FNNs) were among the continuous quantum error correction (CQEC) protocol for
earliest supervised learning approaches in QEC. Torlai and superconducting qubits, addressing real-world imperfec-
Melko [20] demonstrated an FNN-based decoder for toric tions such as auto-correlated noise, transient syndrome dy-
code, a topological code defined on a torus. Their work high- namics, and steady-state drift. Their results revealed that
lighted the FNN’s ability to accurately map syndromes to the RNN-based CQEC protocol outperformed conventional
corrections, showcasing its potential in topological QEC ap- double-threshold schemes, achieving fidelity comparable to
plications. Locher et al. [125] extended these ideas through Bayesian classifiers. Recently, the AlphaQubit model [127], a
the application of quantum autoencoders for autonomous recurrent transformer-based neural network, set new bench-
QEC. Their framework effectively handled computational marks in decoding surface codes by leveraging both simu-
errors, qubit losses, and noise-specific adaptations, offering lated and experimental data under realistic noise conditions,
a robust solution for dynamic error correction in quantum highlighting its robustness and accuracy.
systems. Transformer-based architectures have also been ex-
Convolutional Neural Networks (CNNs), well-known plored by other researchers, Wang et al. [128] introduced a
for their capability to exploit spatial correlations, have transformer-based decoder leveraging self-attention mech-
proven to be particularly effective for topological quantum anisms for global contextual awareness. By incorporating
codes such as the surface code. Baireuther et al. [114] and a mixed loss function addressing both local physical er-
Maskara et al. [115] demonstrated that CNN architectures, rors and global parity constraints, and employing transfer
which encode the local structure of qubit layouts, achieve learning to reduce training overhead across different code
high accuracy and scalability. These models benefit from distances, Transformer-QEC achieved superior logical error
data augmentation techniques and hyperparameter opti- rates and scalability. Similarly, the Deep Quantum Error Cor-
mization, which further enhance generalization across var- rection framework proposed by Choukroun et al. [129] in-
ious noise scenarios. The success of CNNs in this domain troduced the Quantum Error Correction Code Transformer
underscores their ability to adapt to the lattice structure (QECCT). This architecture addressed key QEC challenges,
 13

such as overcoming measurement collapse through noise By grouping similar syndromes, these methods enable the
prediction, optimizing the non-differentiable Logical Error targeted refinement of QEC strategies, reducing overhead
Rate (LER) metric using a differentiable approximation, in classical processing and qubit resources. Clustering is
and employing a pooling mechanism to handle faulty syn- especially effective in handling noise patterns with strong
drome measurements. Applied to surface and toric codes, correlations or non-local dependencies, providing insights
QECCT demonstrated robustness and scalability, outper- that traditional decoders may fail to capture.
forming state-of-the-art decoders like Minimum Weight Per- Dimensionality reduction techniques simplify and in-
fect Matching (MWPM) under independent and depolariza- terpret high-dimensional syndrome data, aiding the visual-
tion noise models. ization and understanding of complex error models. Meth-
Neural Ensemble Decoding, which combines multiple ods like Principal Component Analysis (PCA) and t-distributed
neural networks, has been proposed to enhance decoding Stochastic Neighbor Embedding (t-SNE) highlight dominant
robustness and accuracy for topological quantum codes. error modes and clusters of errors with shared underlying
This method significantly outperforms conventional de- causes. These insights guide practitioners in developing
coders such as MWPM, particularly in noisy environments more intuitive and efficient decoding strategies. Dimension-
with correlated errors. The work by Convy et al. [130] ality reduction also supports hardware noise characteriza-
demonstrated the neural ensemble decoder’s superior per- tion, identifying systematic errors that can be addressed
formance on various topological codes, showcasing its po- with tailored QEC protocols, making it a powerful diagnos-
tential for high error-correction fidelity with reduced com- tic tool.
putational complexity in large-scale quantum systems. More recent work incorporates generative modeling
Graph Neural Networks (GNNs), which process data and representation learning into unsupervised QEC work-
structured as graphs, are particularly suited to leveraging flows. Variational Autoencoders (VAEs) and other genera-
the lattice structure of quantum codes. GNNs map qubits tive models learn compact representations of complex error
and stabilizer measurements to nodes and edges, respec- landscapes, enabling hybrid decoders that integrate unsu-
tively, enabling them to learn directly from the geometric pervised pattern recognition with downstream supervised
properties of the code. Lange et al. [131] proposed a GNN- or reinforcement learning steps. VAEs transform raw syn-
based framework that achieved competitive performance drome data into lower-dimensional latent spaces, simplify-
under realistic noise models, highlighting the promise of ing clustering and pattern recognition tasks. These latent
graph-based methods for decoding. spaces also inform parameter tuning in supervised decoders
Supervised learning in QEC is not limited to neural or guide error mitigation strategies. Generative models
networks. Classical machine learning methods, such as further address data scarcity by creating synthetic error
Support Vector Machines (SVMs), offer interpretable de- syndromes, improving the robustness of training datasets
cision boundaries and competitive accuracy for smaller for experimental quantum systems.
codes [115]. However, methods such decision trees and Cao et al. [134] introduce a groundbreaking framework
random forests face challenges in handling the high- for decoding QEC using unsupervised generative modeling
dimensional error spaces associated with large-scale codes. with autoregressive neural networks, specifically transform-
Chen et al. [132] showcased a high-fidelity readout scheme ers. The method models the joint probability distribution
for transmon qubits, integrating machine learning tech- of stabilizers, logical operators, and syndromes, facilitat-
niques to achieve QEC thresholds. This work underscored ing efficient and scalable maximum likelihood decoding
the importance of supervised learning in state discrimina- without labeled data. By capturing intricate dependencies
tion and error correction. Davaasuren et al. [133] proposed between these elements, the framework enhances decoding
a general framework for constructing fast and near-optimal of complex error syndromes. It achieves superior logical
machine-learning-based decoders for topological stabilizer error rates compared to traditional decoders like Minimum
codes. By leveraging a linear prediction framework and uni- Weight Perfect Matching (MWPM) and Belief Propagation
form data construction, their approach demonstrated supe- Augmented Ordered Statistics Decoding (BPOSD). Addi-
rior decoding efficiency and accuracy for surface and color tionally, it demonstrates versatility across QEC topologies,
codes, further advancing the field of machine-learning- such as surface codes and quantum LDPC codes. By re-
driven QEC decoders. ducing computational complexity from O(4k ) to O(2k ),
qecGPT provides an efficient, parallelizable solution for
5.4.2 Unsupervised Learning Approaches fault-tolerant quantum computing, significantly advancing
Unsupervised learning techniques aim to discover patterns, scalability in QEC.
structures, or correlations within unlabeled data. In QEC,
this is particularly valuable, as not all error configurations 5.4.3 Semi-Supervised Learning Approaches
are easily labeled, and the complexity of noise models Semi-supervised learning bridges the gap between fully
often outpaces the ability to generate comprehensive la- supervised and unsupervised methods by leveraging both
beled datasets. By uncovering hidden structures in error labeled and unlabeled data. In the context of QEC, the
syndromes, unsupervised methods can inspire the design generation of accurately labeled training data (i.e., matched
of more adaptive and resource-efficient error correction syndrome-correction pairs) can be computationally expen-
protocols. sive, as it often requires simulating quantum circuits un-
Clustering algorithms, such as K-means and hierar- der realistic noise conditions. At the same time, abundant
chical clustering, have been applied to identify groups of unlabeled syndrome data may be available from continu-
correlated errors that conventional decoders may overlook. ous quantum device operation. Semi-supervised techniques
 14

help reduce the reliance on large, fully labeled datasets, the toric code under uncorrelated noise. Their method
thereby cutting down on the classical computational cost of achieved near-optimal performance, approaching the theo-
data generation and improving the scalability of ML-based retical threshold of 11%, by rewarding the agent based solely
decoders. on successful logical state recovery.
Recent studies have explored pseudo-labeling and con- Sweke et al. [137] further explored RL-based decod-
sistency regularization techniques to enhance decoder ing in the context of fault-tolerant quantum computation.
performance while using fewer labeled samples. Pseudo- They reformulated the decoding problem as a reinforcement
labeling generates artificial labels for unlabeled data using learning task, leveraging deep Q-learning to train decoding
a pre-trained model, which are then incorporated into the agents for surface codes under simplified phenomenological
training process to refine the model further. Consistency noise models. Their approach demonstrated the flexibility of
regularization enforces the model to produce stable predic- RL techniques in handling faulty syndrome measurements
tions across augmented versions of the same input, thereby and set the foundation for integrating advanced RL algo-
improving robustness. For instance, Ji et al. [135] integrated rithms in realistic quantum noise scenarios.
semi-supervised graph-based methods into their QEC de- Fitzek et al. [138] introduced a Deep Q-Learning decoder
coder pipeline, enabling effective inference of underlying for the toric code under depolarizing noise, pushing the
error structures despite limited labeling resources. These boundaries of RL applications in QEC. Their approach used
techniques significantly improve performance by efficiently a deep Q-network to approximate Q-values for selecting
utilizing the vast amounts of unlabeled data available dur- Pauli corrections, allowing the decoder to handle correlated
ing quantum device operation. bit-flip and phase-flip errors. The RL decoder outperformed
Another promising direction is to incorporate transfer minimum-weight perfect matching (MWPM) for depolar-
learning and domain adaptation strategies. When transi- izing noise with error rates below the threshold ( 16.5%),
tioning between noise regimes or quantum hardware plat- demonstrating superior handling of complex noise models.
forms, semi-supervised approaches can adapt previously This work highlights the potential of RL to address chal-
learned representations to new conditions with minimal ad- lenges posed by noise correlations and scalability in QEC.
ditional labeling. Transfer learning repurposes knowledge Subsequent research has extended these ideas to more
from a pre-trained model to new tasks or environments, complex noise models and larger code sizes, yielding im-
while domain adaptation adjusts the model to align distri- proved performance and generalization capabilities. Hus-
butions between source and target domains. This adaptabil- sain et al. [139] integrated actor-critic methods to develop
ity is particularly beneficial for real-world quantum devices, adaptive decoders that continually refine their strategies as
where error landscapes evolve over time due to drift in error statistics shift. This adaptive approach helps maintain
control parameters, varying qubit coherence times, or hard- high-fidelity logical qubits in fluctuating or time-varying
ware upgrades. Semi-supervised learning thus provides a noise environments.
practical approach to maintain decoder performance under Research has also explored combining RL with other
dynamic noise conditions. learning paradigms. Sweke et al. [140] investigated hybrid
By mitigating the data bottleneck and improving gener- RL-semi-supervised decoders that utilize unlabeled syn-
alization to new conditions, semi-supervised methods are drome data to guide exploration, reducing the need for
poised to make QEC decoders more efficient and robust. As extensive training sets. Additionally, incorporating domain-
quantum processors scale, the ability to harness unlabeled specific knowledge (e.g., known symmetries or stabilizer
data and quickly adapt to changing noise environments will constraints) into RL reward functions has helped accelerate
be critical for achieving long-term, fault-tolerant quantum training and improve interpretability. More recent work has
computation. Semi-supervised learning stands out as a scal- focused on enhancing scalability and robustness. Wagner
able solution for enabling efficient QEC protocols in large- and Devitt [141] proposed domain-informed reinforcement
scale quantum systems. learning strategies that embed stabilizer constraints and
symmetry principles directly into the RL framework. These
5.4.4 Reinforcement Learning Approaches methods significantly reduce exploration complexity and
Reinforcement Learning (RL) techniques frame the decod- training time while maintaining high decoding accuracy.
ing problem as an agent interacting with a stochastic envi- As quantum devices scale and the complexity of realistic
ronment—in this case, a quantum code subject to noise. The noise models increases, RL’s adaptive and data-driven na-
agent receives rewards for successful logical qubit recovery ture offers a flexible path toward more autonomous QEC
and penalties for incorrect or delayed corrective actions. decoders. Ongoing efforts aim to integrate RL-based de-
Through iterative exploration, the RL agent learns policies coders directly into quantum control stacks, enabling near
that adaptively select correction operations, potentially out- real-time decoding and dynamically optimized error correc-
performing static or hand-crafted decoding rules. tion protocols suitable for emerging fault-tolerant quantum
Early explorations used value-based methods and pol- processors.
icy gradient techniques to decode topological codes. For
instance, Andreasson et al. [21] demonstrated that a deep 5.4.5 Deep Learning Models
RL agent could learn to decode the toric code effectively Deep learning models, characterized by neural networks
without explicit knowledge of underlying error distribu- with multiple hidden layers, have introduced new avenues
tions. Domingo Colomer et al. [136] extended these ideas, for handling the complexity and variety of errors in QEC.
applying deep reinforcement learning techniques with con- Unlike simpler machine learning methods, these architec-
volutional networks to design high-threshold decoders for tures excel at learning hierarchical representations directly
 15

from raw data, reducing the need for extensive feature A low logical error rate is essential for executing deep quan-
engineering. tum circuits reliably. Conventional decoding methods, such
Early applications explored Deep Belief Networks as MWPM [143], have historically provided near-threshold
(DBNs) and Autoencoders, showing that these architectures performance for well-studied codes like the surface code.
could efficiently reconstruct or decode quantum states un- However, recent AI-driven decoders (e.g., graph neural
der noisy conditions. Building on this, autoencoders have network-based approaches [144] have shown comparable or
gained popularity due to their ability to learn compressed sometimes even improved logical error rates under realistic
latent representations. Locher et al. [125] demonstrated noise models, while reducing classical computational over-
that autoencoders can successfully map corrupted syn- head.
drome data back to original states, providing efficient er- The threshold error rate (pth ) is the physical error rate
ror recovery for high-dimensional quantum systems. When below which logical error rates can be arbitrarily suppressed
used alongside previously discussed supervised or semi- by increasing the code distance. The Threshold Theorem [102]
supervised pipelines, these deep architectures can reduce ensures that quantum computation can scale if p < pth .
the label burden by distilling essential features of the error For a given physical error rate p, the logical error rate
landscape into more manageable representations. approximately scales as:
Generative Adversarial Networks (GANs), which con-   d+1
sist of competing generator and discriminator models, ex- p 2
PL ≈ A , (15)
tend these capabilities by learning to model complex, high- pth
dimensional error distributions. GANs have been employed where A is a constant and d is the code distance. Advanced
to generate synthetic training samples that complement ML-based decoders have demonstrated thresholds close to
limited experimental data, easing the data scarcity prob- or on par with those of conventional methods for topological
lem. Wang et al. [142] showed that GANs can reproduce codes [115], [145], even under more complex and correlated
intricate error statistics, enabling the development of more error models.
robust decoders when combined with supervised or RL-
based training phases. Such generative modeling supports 6.1.2 Resource Overhead
the hybridization of learning strategies, where unsuper- Resource overhead encompasses the additional physical
vised GAN pretraining provides better initializations for qubits, gates, and measurements required to implement
supervised decoders or informs RL agents exploring new QEC:
parameter regimes. • Qubit Overhead: Ratio of physical to logical qubits.
Deep learning’s hierarchical and adaptive learning ca- • Gate Overhead: Number of gates needed for encoding
pabilities make it a promising framework for addressing and syndrome extraction.
the increasing complexity of QEC in scalable quantum sys- • Measurement Overhead: Frequency and complexity of
tems. As quantum hardware continues to advance, lever- measurements for identifying error syndromes.
aging these models to autonomously adapt to diverse and While conventional surface code decoders often demand
evolving noise conditions will be critical for enabling fault- significant overhead, recent demonstrations have shown
tolerant quantum computation. that AI-based decoders can help optimize syndrome ex-
traction and code configuration, potentially reducing the
6 E VALUATION number of measurements or simplifying the decoding pro-
In this section, we introduce key evaluation metrics for cess [121]. In some cases, ML models can exploit latent
QEC and benchmark AI and ML methods using these structures in the error data to propose lower-overhead QEC
metrics, focusing on critical indicators such as logical error schemes, thus alleviating pressure on hardware-limited de-
rate, threshold error rate, resource overhead, decoding time, vices.
computational complexity, and fidelity.
6.1.3 Decoding Time and Computational Complexity
6.1 Evaluation Metrics in QEC The decoding time and computational complexity of classi-
Evaluating the effectiveness of QEC methods requires a set cal processing steps are critical for achieving real-time QEC.
of reliable, quantitative metrics that capture their ability to Conventional decoders like MWPM scale as O(N 3 ) in the
protect quantum information under realistic noise condi- worst case (N being the number of defects) [143], [146],
tions. These metrics also serve as benchmarks for compar- which may become prohibitive for large-scale quantum
ing conventional decoding strategies with more recent, AI- processors.
driven approaches. In contrast, ML-based approaches often require signifi-
Commonly used evaluation criteria include the logical cant upfront training but offer rapid inference once trained.
error rate, threshold error rate, resource overhead, decoding Convolutional neural network (CNN) or GNN-based de-
time, computational complexity, and fidelity. Together, these coders can decode syndromes in effectively constant or
metrics provide a holistic picture of QEC performance. polynomial time, independent of code size, after train-
ing [114]. Reinforcement learning decoders can adapt to
6.1.1 Logical Error Rate and Threshold Error Rate time-varying noise models, maintaining decoding accuracy
The logical error rate (PL ) quantifies the probability that a with minimal additional computational overhead. This shift
logical qubit experiences an error after a round of QEC: from runtime complexity to a training-inference paradigm is
Number of logical errors key to enabling efficient and scalable QEC in large quantum
PL = . (14) devices.
Total number of logical qubits processed
 16

6.1.4 Fidelity ML decoders have demonstrated thresholds compara-

Fidelity measures how closely the corrected quantum state ble to MWPM or lookup-based decoders for topological
matches the intended logical state. It provides a direct codes [144], [145].
measure of how well QEC preserves quantum information: • Resource Overhead: Quantifies the additional qubits,
gates, or classical computation required. Some ML de-
2
√ √ coders can exploit learned features to reduce overhead,
 q
F = Tr ρσ ρ , (16) suggesting a pathway to more efficient architectures.
• Decoding Time and Computational Complexity: Un-
where ρ is the ideal logical state and σ is the corrected state. like classical decoders with polynomial or cubic scaling,
Fidelity is especially useful for assessing the absolute per- ML decoders often shift complexity from runtime de-
formance of a given QEC strategy. Recent experiments and coding to a training phase. This enables near real-time
simulations indicate that advanced ML decoders can main- inference on large codes once training is complete [114].
tain or improve fidelity compared to traditional decoders, Benchmarking must carefully account for both upfront
especially in regimes with correlated or non-Markovian training costs and per-instance decoding latency.
noise [115], [135]. • Fidelity and State Overlap: Evaluates how closely the
corrected state matches the ideal logical state. This
metric is especially relevant in hardware-based bench-
6.2 Benchmarking AI and ML Methods in QEC
marks, where fidelity directly reflects the quality of
Benchmarking AI and ML-based decoders against conven- quantum operations post-correction [135].
tional methods is essential for quantifying their advantages,
identifying best practices, and guiding future research. By
employing standardized datasets, realistic error models, and 7 F UTURE W ORKS
well-defined performance metrics, researchers can fairly Integration of AI and ML into QEC presents many op-
compare these tools and evaluate their readiness for deploy- portunities, but there are still several challenges and open
ment in large-scale quantum computing systems. questions that need to be addressed. This section outlines
potential future research directions to advance the field.
6.2.1 Datasets and Error Models Several promising research directions must be explored.
These directions aim to improve the efficiency, robustness,
The reliability of benchmarking hinges on representative and scalability of ML models, as well as their integration
datasets that capture the complexity and diversity of quan- with quantum hardware.
tum noise. These datasets typically consist of large sets of
syndrome patterns paired with corresponding known error
configurations. Various quantum codes (e.g., surface codes, 7.1 Development of Efficient Training Methods
toric codes, color codes) are simulated under a range of error A major challenge in applying AI/ML techniques to QEC is
models, including: the requirement for large datasets to train the models. This
• Depolarizing Noise: Each qubit is equally likely to issue can be mitigated by exploring data-efficient learning
experience any one of the three Pauli errors (X , Y , or techniques, such as few-shot learning, which allows models
Z ). This model serves as a baseline for many studies. to generalize from a small number of examples. This is
• Bit-Flip and Phase-Flip Errors: Noise that predomi- especially important for quantum systems where generating
nantly manifests as X or Z errors, reflecting simpler large datasets from real hardware is costly or impractical. By
but practically relevant error channels. learning from limited data, few-shot learning can reduce the
• Correlated Errors: Errors that affect multiple qubits need for extensive dataset generation [149].
simultaneously, often due to crosstalk or shared con- Another key approach is the use of synthetic data gener-
trol lines. Modeling correlated noise is critical, as it ation via generative models, which can create representative
presents a more realistic and challenging scenario for training data based on simulations of quantum noise and
decoders [147]. errors. This can significantly reduce reliance on real-world
• Biased Noise: Noise favoring certain error types (e.g., quantum hardware for dataset collection, as proposed in
dephasing in superconducting qubits), which can be studies [118].
leveraged to reduce overhead in specialized decoding Transfer learning and domain adaptation can also be
strategies [148]. leveraged to adapt models trained on one system to new
error distributions or different hardware platforms. This
6.2.2 Performance Metrics and Quantitative Comparisons cross-platform adaptability would allow models to retain
knowledge gained from one environment while efficiently
A fair comparison between AI/ML-based decoders and con-
generalizing to others, thereby reducing the need for com-
ventional methods requires consistent use of key metrics:
plete retraining when shifting to different quantum archi-
• Logical Error Rate: Measures how effectively the de-
tectures [150].
coder reduces physical errors to preserve logical qubits.
AI-based methods have matched or surpassed tradi-
tional decoders in achieving low logical error rates 7.2 Enhancing Model Robustness and Generalization
under certain conditions [115]. Ensuring that ML models generalize well to unseen error
• Threshold Error Rate: Determines whether the decoder patterns is critical for their real-world deployment in QEC.
can push logical error rates below a scalable threshold. Regularization techniques, such as dropout, weight decay,
 17

and early stopping, can help prevent overfitting to specific fectively with the increasing demands of quantum hard-
datasets, ensuring that the models remain effective across ware.
diverse and changing error landscapes [63]. In addition, en- As quantum hardware evolves, integrating AI/ML de-
semble learning, which combines the predictions of multiple coders with these systems will require ongoing advance-
models, can improve robustness by leveraging the strengths ments in both hardware design and software optimization.
of each model to create a more accurate and generalizable Future efforts will prioritize reducing latency, enhancing
system. modularity, and leveraging existing co-design methodolo-
Adversarial training is another promising technique to gies to support fault-tolerant quantum computing.
improve the robustness of AI/ML decoders. By exposing
models to adversarially generated error patterns—errors
specifically designed to challenge the model—AI/ML sys- R EFERENCES
tems can be trained to handle unexpected and extreme error [1] Google Quantum AI and Collaborators, “Quantum error correc-
conditions, increasing their resilience in real-world quantum tion below the surface code threshold,” Nature, 2024.
environments [151]. This method can be applied to simulate [2] P. W. Shor, “Algorithms for quantum computation: discrete loga-
rithms and factoring,” in Proceedings of the 35th Annual Symposium
worst-case scenarios, ensuring that the models perform well on Foundations of Computer Science, ser. SFCS ’94. USA: IEEE
under a variety of error distributions. Computer Society, 1994, p. 124–134.
[3] L. K. Grover, “A fast quantum mechanical algorithm for database
search,” in Proceedings of the Twenty-Eighth Annual ACM Sympo-
7.3 Integration with Quantum Hardware sium on Theory of Computing, ser. STOC ’96. New York, NY, USA:
Association for Computing Machinery, 1996, p. 212–219.
For AI/ML decoders to be effective in practical quantum [4] C. H. Bennett and G. Brassard, “Quantum cryptography: Public
systems, they must be efficiently integrated with quantum key distribution and coin tossing,” in Proceedings of IEEE Interna-
hardware. This requires a co-design of quantum-classical tional Conference on Computers, Systems and Signal Processing, 1984.
[5] P. Shor, “Algorithms for quantum computation: discrete loga-
systems, where the AI/ML models and quantum hardware rithms and factoring,” in Proceedings 35th Annual Symposium on
are developed in tandem. Hardware-friendly models that Foundations of Computer Science, 1994, pp. 124–134.
are optimized for low-latency operation and minimal re- [6] E. Farhi, J. Goldstone, and S. Gutmann, “A quantum
source consumption will be crucial in reducing the overhead approximate optimization algorithm,” 2014. [Online]. Available:
https://arxiv.org/abs/1411.4028
of real-time error correction. Implementing these models [7] R. P. Feynman, “Simulating physics with computers,” Interna-
on field-programmable gate arrays (FPGAs) or application- tional Journal of Theoretical Physics, vol. 21, no. 6-7, pp. 467–488,
specific integrated circuits (ASICs) is a promising direction, 1982.
[8] I. M. Georgescu, S. Ashhab, and F. Nori, “Quantum simulation,”
as such hardware offers fast, efficient, and scalable decoding Rev. Mod. Phys., vol. 86, pp. 153–185, Mar 2014.
capabilities [152]. [9] W. H. Zurek, “Decoherence, einselection, and the quantum ori-
Another important area is the development of real-time gins of the classical,” Rev. Mod. Phys., vol. 75, pp. 715–775, May
implementation strategies. By deploying AI/ML models 2003.
[10] J. Preskill, “Reliable quantum computers,” Proceedings of the Royal
on edge devices located close to the quantum hardware, Society of London. Series A: Mathematical, Physical and Engineering
communication delays can be minimized, improving the Sciences, vol. 454, pp. 385–410, 1998.
speed and efficiency of error correction. In addition, parallel [11] W. H. Zurek, “Decoherence and the transition from quantum to
processing techniques can be employed to speed up the classical,” Physics Today, vol. 44, no. 10, pp. 36–44, 10 1991.
[12] J. Preskill, “Quantum Computing in the NISQ era and beyond,”
decoding process, allowing AI/ML decoders to keep pace Quantum, vol. 2, p. 79, Aug. 2018.
with the fast error rates observed in large-scale quantum [13] P. W. Shor, “Scheme for reducing decoherence in quantum com-
systems. puter memory,” Phys. Rev. A, vol. 52, pp. R2493–R2496, Oct 1995.
[14] A. G. Fowler, M. Mariantoni, J. M. Martinis, and A. N. Cleland,
Barnes et al. [153] proposed a framework for improv- “Surface codes: Towards practical large-scale quantum computa-
ing efficiency and portability in quantum systems through tion,” Phys. Rev. A, vol. 86, p. 032324, Sep 2012.
a multi-level hardware abstraction layer. This abstraction [15] R. Raussendorf, J. Harrington, and K. Goyal, “Topological fault-
layer facilitates the integration of AI/ML-based decoders tolerance in cluster state quantum computation,” New Journal of
Physics, vol. 9, no. 6, p. 199, jun 2007.
by modularizing the interface between quantum and classi- [16] D. Gao, D. Fan, C. Zha, J. Bei, G. Cai, J. Cai et al., “Establishing a
cal hardware, thereby enhancing portability and scalability new benchmark in quantum computational advantage with 105-
across diverse quantum architectures. Such approaches are qubit zuchongzhi 3.0 processor,” arXiv preprint, 2024. [Online].
Available: https://arxiv.org/abs/2412.11924
critical for enabling efficient co-design of quantum-classical
[17] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol.
systems. Wang et al. [154] explored field-programmable 521, pp. 436–444, 2015.
qubit arrays (FPQAs) and introduced techniques to enhance [18] M. I. Jordan and T. M. Mitchell, “Machine learning: Trends,
the scalability of quantum systems. Their approach focuses perspectives, and prospects,” Science, vol. 349, no. 6245, pp. 255–
260, 2015.
on improving the interplay between quantum hardware and [19] R. Sutton and A. Barto, “Reinforcement learning: An introduc-
classical control systems, demonstrating the feasibility of tion,” IEEE Transactions on Neural Networks, vol. 9, no. 5, pp. 1054–
implementing high-performance AI/ML-based decoders in 1054, 1998.
practical settings. [20] G. Torlai and R. G. Melko, “Neural decoder for topological
codes.” Physical review letters, vol. 119 3, p. 030501, 2016.
Parallel processing strategies also play a key role in [21] P. Andreasson, J. Johansson, S. Liljestrand, and M. Granath,
improving the efficiency of AI/ML-based decoders. Payares “Quantum error correction for the toric code using deep rein-
and Martinez-Santos [155] investigated parallel quantum- forcement learning,” Quantum, vol. 3, p. 183, Sep. 2019.
[22] R. P. Feynman, “Simulating physics with computers,” Interna-
classical computation techniques that enable real-time pro-
tional Journal of Theoretical Physics, vol. 21, pp. 467–488, 1982.
cessing of large datasets and complex error patterns. These [23] M. Planck, “Ueber das gesetz der energieverteilung im normal-
strategies ensure that AI/ML-based decoders can scale ef- spectrum,” Annalen der Physik, vol. 309, no. 3, pp. 553–563, 1901.
 18

[24] A. Einstein, “Über einen die erzeugung und verwandlung des [49] M. Campbell, A. J. Hoane, and F. Hsu, “Deep blue,” Elsevier
lichtes betreffenden heuristischen gesichtspunkt,” Annalen der Artificial Intelligence, vol. 134, no. 1-2, pp. 57–83, 2002.
Physik, vol. 322, no. 6, pp. 132–148, 1905. [50] G. E. Hinton, S. Osindero, and Y.-W. Teh, “A fast learning
[25] N. Bohr, “On the constitution of atoms and molecules,” Philosoph- algorithm for deep belief nets,” Neural Computation, vol. 18, no. 7,
ical Magazine, vol. 26, no. 151, pp. 1–25, 1913. pp. 1527–1554, 2006.
[26] W. Heisenberg, “Über quantentheoretische umdeutung kinema- [51] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classi-
tischer und mechanischer beziehungen,” Zeitschrift für Physik, fication with deep convolutional neural networks,” in NeurIPS,
vol. 33, pp. 879–893, 1925. 2012.
[27] E. Schrödinger, “Quantisierung als eigenwertproblem (erste mit- [52] I. Goodfellow et al., “Generative adversarial nets,” in NeurIPS,
teilung),” Annalen der Physik, vol. 384, no. 4, pp. 361–376, 1926. 2014.
[28] W. Heisenberg, “Über den anschaulichen inhalt der quantentheo- [53] D. Silver et al., “Mastering the game of go with deep neural
retischen kinematik und mechanik,” Zeitschrift für Physik, vol. 43, networks and tree search,” Nature, vol. 529, no. 7587, pp. 484–
no. 3–4, pp. 172–198, 1927. 489, 2016.
[29] A. Einstein, B. Podolsky, and N. Rosen, “Can quantum- [54] A. Vaswani et al., “Attention is all you need,” in NeurIPS, 2017.
mechanical description of physical reality be considered com- [55] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-
plete?” Physical Review, vol. 47, no. 10, pp. 777–780, 1935. training of deep bidirectional transformers for language un-
[30] J. S. Bell, “On the einstein podolsky rosen paradox,” Physics derstanding,” in Proceedings of the 2019 Conference of the North
Physique Fizika, vol. 1, no. 3, pp. 195–200, 1964. American Chapter of the Association for Computational Linguistics,
[31] A. Aspect, P. Grangier, and G. Roger, “Experimental realization 2019.
of einstein-podolsky-rosen-bohm gedankenexperiment: A new [56] A. Radford et al., “Language models are unsupervised multitask
violation of bell’s inequalities,” Physical Review Letters, vol. 49, learners,” OpenAI Blog, vol. 1, no. 8, p. 9, 2019.
no. 2, pp. 91–94, 1982. [57] J. Jumper et al., “Highly accurate protein structure prediction
[32] D. Bouwmeester, J.-W. Pan, K. Mattle, M. Eibl, H. Weinfurter, and with alphafold,” Nature, vol. 596, no. 7873, pp. 583–589, 2021.
A. Zeilinger, “Experimental quantum teleportation,” Nature, vol. [58] A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. Voss, A. Radford,
390, no. 6660, pp. 575–579, 1997. M. Chen, and I. Sutskever, “Zero-shot text-to-image generation,”
[33] L. M. K. Vandersypen, M. Steffen, G. Breyta, C. S. Yannoni, M. H. in Proceedings of the 38th International Conference on Machine Learn-
Sherwood, and I. L. Chuang, “Experimental realization of shor’s ing, ser. Proceedings of Machine Learning Research, M. Meila and
quantum factoring algorithm using nuclear magnetic resonance,” T. Zhang, Eds., vol. 139. PMLR, 18–24 Jul 2021, pp. 8821–8831.
Nature, vol. 414, no. 6866, pp. 883–887, 2001. [59] OpenAI, “Chatgpt: Optimizing language models for dialogue,”
[34] “D-wave one quantum computer,” https://www.dwavesys.com, https://openai.com/blog/chatgpt, 2022.
2011. [60] J. A. OpenAI, S. Adler, S. Agarwal et al., “Gpt-4 technical
[35] F. Arute, K. Arya, R. Babbush, D. Bacon, J. C. Bardin, R. Barends, report,” 2023. [Online]. Available: https://api.semanticscholar.
R. Biswas, S. Boixo, F. G. Brandao, D. A. Buell et al., “Quantum org/CorpusID:257532815
supremacy using a programmable superconducting processor,” [61] J. M. Jumper et al., “Accurate structure prediction of biomolecular
Nature, vol. 574, no. 7779, pp. 505–510, 2019. interactions with alphafold 3,” Nature, vol. 630, pp. 493–500, 2024.
[36] “The nobel prize in physics 2022,” https://www.nobelprize.org/ [62] S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach,
prizes/physics/2022/summary/, 2022. 3rd ed. Pearson, 2016.
[37] IBM Quantum, “Quantum roadmap to 2033,” 2023. [63] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT
[Online]. Available: https://www.ibm.com/quantum/blog/ Press, 2016.
quantum-roadmap-2033
[64] C. M. Bishop, Pattern Recognition and Machine Learning. Springer,
[38] Atom Computing, “Quantum computing technology,” 2006.
2023. [Online]. Available: https://atom-computing.com/
[65] E. Alpaydin, Introduction to Machine Learning, 4th ed. MIT Press,
quantum-computing-technology/
2020.
[39] P. W. Shor, “Polynomial-time algorithms for prime factorization
[66] K. P. Murphy, Machine Learning: A Probabilistic Perspective. MIT
and discrete logarithms on a quantum computer,” SIAM Journal
Press, 2012.
on Computing, vol. 26, no. 5, pp. 1484–1509, 1997.
[67] L. Breiman, J. Friedman, R. Olshen, and C. Stone, Classification and
[40] W. K. Wootters and W. H. Zurek, “A single quantum cannot be
Regression Trees. Wadsworth & Brooks/Cole Advanced Books &
cloned,” Nature, vol. 299, pp. 802–803, 1982.
Software, 1984.
[41] D. Gottesman, “Stabilizer codes and quantum error correction,”
[68] C. Cortes and V. Vapnik, “Support-vector networks,” Machine
arXiv: Quantum Physics, 1997. [Online]. Available: https:
Learning, vol. 20, no. 3, pp. 273–297, 1995.
//api.semanticscholar.org/CorpusID:118824832
[42] W. H. Zurek, “Decoherence, einselection, and the quantum ori- [69] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol.
gins of the classical,” Reviews of Modern Physics, vol. 75, no. 3, pp. 521, pp. 436–444, 2015.
715–775, 2003. [70] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical
[43] J. Gambetta, A. Blais, D. I. Schuster, A. Wallraff, L. Frunzio, Learning: Data Mining, Inference, and Prediction. Springer, 2009.
J. Majer, M. H. Devoret, S. M. Girvin, and R. J. Schoelkopf, [71] I. T. Jolliffe, Principal Component Analysis, 2nd ed. Springer, 2002.
“Qubit-photon interactions in a cavity: Measurement-induced [72] J. MacQueen, “Some methods for classification and analysis of
dephasing and number splitting,” Phys. Rev. A, vol. 74, p. 042318, multivariate observations,” Proceedings of the Fifth Berkeley Sympo-
Oct 2006. sium on Mathematical Statistics and Probability, vol. 1, pp. 281–297,
[44] A. Wallraff, D. I. Schuster, A. Blais, L. Frunzio, R.-S. Huang, 1967.
J. Majer, S. Kumar, S. M. Girvin, and R. J. Schoelkopf, “Approach- [73] J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and
ing unit visibility for control of a superconducting qubit with S. Lloyd, “Quantum machine learning,” Nature, vol. 549, no. 7671,
dispersive readout,” Physical Review Letters, vol. 95, p. 060501, pp. 195–202, 2017.
2005. [74] M. Schuld and F. Petruccione, Supervised Learning with Quantum
[45] M. H. Devoret and R. J. Schoelkopf, “Superconducting circuits for Computers. Springer, 2018.
quantum information: An outlook,” Science, vol. 339, pp. 1169– [75] R. Sutton and A. Barto, Reinforcement Learning: An Introduction.
1174, 2013. MIT Press, 2018.
[46] R. Barends, J. Kelly, A. Megrant, A. Veitia, D. Sank, E. Jeffrey, [76] L. P. Kaelbling, M. L. Littman, and A. W. Moore, “Reinforcement
T. C. White, J. Mutus, A. G. Fowler, B. Campbell et al., “Super- learning: A survey,” Journal of Artificial Intelligence Research, vol. 4,
conducting quantum circuits at the surface code threshold for pp. 237–285, 1996.
fault tolerance,” Nature, vol. 508, no. 7497, pp. 500–503, 2014. [77] V. Mnih, K. Kavukcuoglu, D. Silver et al., “Human-level control
[47] A. M. Turing, “Computing machinery and intelligence,” Mind, through deep reinforcement learning,” Nature, vol. 518, no. 7540,
vol. LIX, no. 236, pp. 433–460, 1950. pp. 529–533, 2015.
[48] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning [78] D. Silver, J. Schrittwieser, K. Simonyan et al., “Mastering the game
representations by back-propagating errors,” Nature, vol. 323, no. of go without human knowledge,” Nature, vol. 550, no. 7676, pp.
6088, pp. 533–536, 1986. 354–359, 2017.
 19

[79] T. Fösel, P. Tighineanu, T. Weiss, and F. Marquardt, “Reinforce- [105] P. Panteleev and G. Kalachev, “Quantum ldpc codes with almost
ment learning with neural networks for quantum feedback,” linear minimum distance,” IEEE Transactions on Information The-
Physical Review X, vol. 8, no. 3, p. 031084, 2018. ory, 2021.
[80] X. Zhu, “Semi-supervised learning literature survey,” Technical [106] N. P. Breuckmann and E. T. Campbell, “Quantum low-density
Report 1530, University of Wisconsin-Madison, 2005. parity-check codes,” PRX Quantum, vol. 2, no. 4, p. 040101, 2021.
[81] O. Chapelle, B. Scholkopf, and A. Zien, Semi-Supervised Learning. [107] I. de Vega and D. Alonso, “Dynamics of non-markovian open
MIT Press, 2006. quantum systems,” Reviews of Modern Physics, vol. 89, no. 1, p.
[82] M. Benedetti, S. Lloyd, S. Sack, and M. Fiorentini, “Semi- 015001, 2017.
supervised learning for quantum classifiers,” New Journal of [108] B. M. Terhal, “Quantum error correction for quantum memories,”
Physics, vol. 21, no. 4, p. 043023, 2019. Reviews of Modern Physics, vol. 87, no. 2, p. 307, 2015.
[83] A. Blum and T. Mitchell, “Combining labeled and unlabeled data [109] P. Aliferis, D. Gottesman, and J. Preskill, “Fault-tolerant quantum
with co-training,” Proceedings of the Eleventh Annual Conference on computation against biased noise,” in Quantum Computing: Back
Computational Learning Theory, pp. 92–100, 1998. Action 2007. World Scientific, 2007, pp. 100–115.
[84] D. Zhou, O. Bousquet, T. N. Lal, J. Weston, and B. Scholkopf, [110] J. Ghosh, A. G. Fowler, J. M. Martinis, and M. R. Geller, “Minimiz-
“Learning with local and global consistency,” Advances in Neural ing leakage errors in a surface code quantum computer,” Physical
Information Processing Systems, vol. 16, pp. 321–328, 2004. Review A, vol. 88, no. 6, p. 062329, 2013.
[85] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning [111] L. Viola and S. Lloyd, “Dynamical suppression of decoherence in
representations by back-propagating errors,” Nature, vol. 323, no. two-state quantum systems,” Physical Review A, vol. 58, no. 4, p.
6088, pp. 533–536, 1986. 2733, 1998.
[86] F. Rosenblatt, “The perceptron: A probabilistic model for in- [112] Y. Fujiwara, “Quantum error correction via codes over finite
formation storage and organization in the brain,” Psychological fields with correlated noise,” Physical Review A, vol. 90, no. 6,
Review, vol. 65, no. 6, pp. 386–408, 1958. p. 062304, 2014.
[87] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based [113] S. Varsamopoulos, B. Criger, and K. Bertels, “Decoding small sur-
learning applied to document recognition,” Proceedings of the face codes with feedforward neural networks,” Quantum Science
IEEE, vol. 86, no. 11, pp. 2278–2324, 1998. and Technology, vol. 3, no. 1, p. 015004, 2017.
[88] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” [114] P. Baireuther, M. D. Caio, B. Criger, C. W. J. Beenakker, and T. E.
Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997. O’Brien, “Neural network decoder for topological color codes
[89] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classifi- with circuit level noise,” New Journal of Physics, vol. 21, no. 1, p.
cation with deep convolutional neural networks,” in Advances in 013003, jan 2019.
Neural Information Processing Systems, vol. 25, 2012, pp. 1097–1105.
[115] N. Maskara, A. Kubica, and T. Jochym-O’Connor, “Advantages of
[90] Z. C. Lipton, “A critical review of recurrent neural networks for versatile neural-network decoding for topological codes,” Physi-
sequence learning,” ArXiv, vol. abs/1506.00019, 2015. [Online]. cal Review A, vol. 99, no. 5, p. 052351, 2019.
Available: https://api.semanticscholar.org/CorpusID:5837849
[116] R. Sweke, I. Roth, J. Izaac, J. Eisert, and C. Granade, “Rein-
[91] G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and forcement learning decoders for fault-tolerant quantum compu-
R. R. Salakhutdinov, “Improving neural networks by preventing tation,” Quantum, vol. 4, p. 257, 2020.
co-adaptation of feature detectors,” 2012. [Online]. Available:
[117] J. Preskill, “Quantum computing in the nisq era and beyond,”
https://arxiv.org/abs/1207.0580
Quantum, vol. 2, p. 79, 2018.
[92] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
network training by reducing internal covariate shift,” ICML, pp. [118] L. Cincio, K. Rudinger, M. Sarovar, and P. J. Coles, “Machine
448–456, 2015. learning of noise-resilient quantum circuits,” PRX Quantum,
vol. 2, p. 010324, Feb 2021.
[93] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for
image recognition,” in CVPR, 2016. [119] M. Hibat-Allah, R. G. Melko, and J. Carrasquilla, “Investigating
topological order using recurrent neural networks,” Phys. Rev. B,
[94] A. M. Steane, “Error correcting codes in quantum theory,” Physi-
vol. 108, p. 075152, Aug 2023.
cal Review Letters, vol. 77, no. 5, pp. 793–797, 1996.
[95] A. R. Calderbank and P. W. Shor, “Good quantum error- [120] H.-P. Breuer, E.-M. Laine, J. Piilo, and B. Vacchini, “Colloquium:
correcting codes exist,” Physical Review A, vol. 54, no. 2, p. 1098, Non-markovian dynamics in open quantum systems,” Reviews of
1996. Modern Physics, vol. 88, no. 2, p. 021002, 2016.
[96] A. M. Steane, “Multiple-particle interference and quantum error [121] S. Varsamopoulos, K. Bertels, and C. G. Almudever, “Comparing
correction,” Proceedings of the Royal Society of London. Series A: neural network based decoders for the surface code,” IEEE
Mathematical, Physical and Engineering Sciences, vol. 452, no. 1954, Transactions on Computers, vol. 69, no. 2, pp. 300–311, 2020.
pp. 2551–2577, 1996. [122] R. Harper, S. T. Flammia, and J. J. Wallman, “Efficient learning
[97] A. Y. Kitaev, “Fault-tolerant quantum computation by anyons,” of quantum noise,” Nature Physics, vol. 16, no. 12, pp. 1184–1188,
Annals of Physics, vol. 303, no. 1, pp. 2–30, 2003. 2020.
[98] J. Kelly et al., “State preservation by repetitive error detection in [123] E. Magesan and J. M. Gambetta, “Effective hamiltonian models
a superconducting quantum circuit,” Nature, vol. 519, pp. 66–69, of the cross-resonance gate,” Physical Review A, vol. 101, no. 5, p.
2015. 052308, 2020.
[99] J. Pearl, “Fusion, propagation, and structuring in belief net- [124] P. Czarnik, A. Arrasmith, P. J. Coles, and L. Cincio, “Error
works,” Artificial Intelligence, vol. 29, no. 3, pp. 241–288, 1986. mitigation with clifford quantum-circuit data,” Quantum, vol. 5,
[100] D. Poulin and Y. Chung, “Belief propagation and quantum error p. 592, 2021.
correction,” Physical Review A, vol. 77, no. 5, p. 052333, 2008. [125] D. F. Locher, L. Cardarelli, and M. Müller, “Quantum Error
[101] A. A. Kovalev and L. P. Pryadko, “Quantum ldpc codes with Correction with Quantum Autoencoders,” Quantum, vol. 7, p.
almost linear minimum distance,” Physical Review A, vol. 88, 942, mar 2023.
no. 1, p. 012311, 2013. [126] I. Convy, H. Liao, S. Zhang, S. Patel, W. P. Livingston, H. N.
[102] D. Aharonov and M. Ben-Or, “Fault-tolerant quantum compu- Nguyen, I. Siddiqi, and K. B. Whaley, “Machine learning for con-
tation with constant error,” in Proceedings of the Twenty-Ninth tinuous quantum error correction on superconducting qubits,”
Annual ACM Symposium on Theory of Computing, ser. STOC ’97. New Journal of Physics, vol. 24, no. 6, p. 063019, jun 2022.
New York, NY, USA: Association for Computing Machinery, [127] J. Bausch, A. W. Senior, F. J. H. Heras, and et al., “Learning high-
1997, p. 176–188. accuracy error decoding for quantum processors,” Nature, vol.
[103] M. Kjaergaard, M. E. Schwartz, J. Braumüller, P. Krantz, J. I.-J. 635, pp. 834–840, 2024.
Wang, S. Gustavsson, and W. D. Oliver, “Superconducting qubits: [128] H. Wang, P. Liu, K. Shao, D. Li, J. Gu, D. Z. Pan, Y. Ding,
Current state of play,” Annual Review of Condensed Matter Physics, and S. Han, “Transformer-qec: Quantum error correction
vol. 11, pp. 369–395, 2020. code decoding with transferable transformers,” 2023. [Online].
[104] F. Arute, K. Arya, R. Babbush, D. Bacon, J. C. Bardin, R. Barends, Available: https://arxiv.org/abs/2311.16082
R. Biswas, S. Boixo, F. G. Brandao, D. A. Buell et al., “Quantum [129] Y. Choukroun and L. Wolf, “Deep quantum error correction,” in
supremacy using a programmable superconducting processor,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38,
Nature, vol. 574, no. 7779, pp. 505–510, 2019. no. 1, 2024, pp. 64–72.
 20

[130] M. Sheth, S. Z. Jafarzadeh, and V. Gheorghiu, “Neural ensemble [152] K. M. Barnes, A. Buyskikh, N. Y. Chen et al., “Optimising the
decoding for topological quantum error-correcting codes,” Phys. quantum/classical interface for efficiency and portability with a
Rev. A, vol. 101, p. 032338, Mar 2020. multi-level hardware abstraction layer for quantum computers,”
[131] M. Lange, P. Havström, B. Srivastava, V. Bergentall, K. Hammar, EPJ Quantum Technology, vol. 10, p. 36, 2023.
O. Heuts, E. van Nieuwenburg, and M. Granath, “Data- [153] K. M. Barnes, T. Bialas, O. Buğdayci, E. T. Campbell, N. I.
driven decoding of quantum error correcting codes using Gillespie, K. Johar, R. Rajan, A. W. Richardson, L. Skoric, C. Topal,
graph neural networks,” 2023. [Online]. Available: https: M. L. Turner, and A. B. Ziad, “A highly efficient qec decoder im-
//arxiv.org/abs/2307.01241 plemented on fpga and asic,” in 2023 IEEE International Conference
[132] L. Chen, H. Li, Y. Lu et al., “Transmon qubit readout fidelity at on Quantum Computing and Engineering (QCE), vol. 02, 2023, pp.
the threshold for quantum error correction without a quantum- 375–376.
limited amplifier,” npj Quantum Information, vol. 9, p. 26, 2023. [154] H. Wang, D. B. Tan, P. Liu, Y. Liu, J. Gu, J. Cong, and S. Han, “Q-
[133] A. Davaasuren, Y. Suzuki, K. Fujii, and M. Koashi, “General pilot: Field programmable qubit array compilation with flying
framework for constructing fast and near-optimal machine- ancillas,” in Proceedings of the 61st ACM/IEEE Design Automation
learning-based decoder of the topological stabilizer codes,” Phys. Conference, ser. DAC ’24. New York, NY, USA: Association for
Rev. Res., vol. 2, p. 033399, Sep 2020. Computing Machinery, 2024.
[134] H. Cao, F. Pan, Y. Wang, and P. Zhang, “qecgpt: decoding [155] E. Payares and J. Martinez-Santos, “Parallel quantum computa-
quantum error-correcting codes with generative pre-trained tion approach for quantum deep learning and classical-quantum
transformers,” 2023. [Online]. Available: https://arxiv.org/abs/ models,” Journal of Physics: Conference Series, vol. 2090, no. 1, p.
2307.09025 012171, nov 2021.
[135] N.-H. Ji, H.-Q. Sun, B. Xiao, P.-L. Song, and H.-Y. Ma, “Quantum
decoder design for subsystem surface code based on multi-head
graph attention and edge weighting,” Chinese Physics B, 2024.
[Online]. Available: http://iopscience.iop.org/article/10.1088/
1674-1056/ada1c7
[136] L. Domingo Colomer, M. Skotiniotis, and R. Muñoz-Tapia, “Re-
inforcement learning for optimal error correction of toric codes,”
Physics Letters A, vol. 384, no. 17, p. 126353, 2020.
[137] R. Sweke, M. S. Kesselring, E. P. L. van Nieuwenburg, and J. Eis-
ert, “Reinforcement learning decoders for fault-tolerant quantum
computation,” Machine Learning: Science and Technology, vol. 2,
no. 2, p. 025005, dec 2020.
[138] D. Fitzek, M. Eliasson, A. F. Kockum, and M. Granath, “Deep q-
learning decoder for depolarizing noise on the toric code,” Phys.
Rev. Res., vol. 2, p. 023230, May 2020.
[139] A. Hussain, D. E. Browne, and E. T. Campbell, “Adaptive quan-
tum error correction via reinforcement learning,” arXiv preprint
arXiv:2212.12345, 2022.
[140] V. K, D. S, and S. T, “Rl-qec: Harnessing reinforcement learning
for quantum error correction advancements,” in 2024 Interna-
tional Conference on Trends in Quantum Computing and Emerging
Business Technologies, 2024, pp. 1–5.
[141] M. Wagner and S. J. Devitt, “Domain-informed reinforcement
learning for fault-tolerant quantum computing,” NPJ Quantum
Inf., vol. 9, p. 123, 2023.
[142] H. Wang, Z. Song, Y. Wang et al., “Target-generating quantum er-
ror correction coding scheme based on generative confrontation
network,” Quantum Information Processing, vol. 21, p. 280, 2022.
[143] A. G. Fowler, “Minimum weight perfect matching of fault-
tolerant topological quantum error correction in average o(1)
parallel time,” Quantum Information and Computation, vol. 15, no.
1-2, pp. 145–158, 2015.
[144] M. Lange, P. Havström, B. Srivastava, V. Bergentall, K. Hammar,
O. Heuts, E. van Nieuwenburg, and M. Granath, “Data-
driven decoding of quantum error correcting codes using
graph neural networks,” 2023. [Online]. Available: https:
//arxiv.org/abs/2307.01241
[145] J. Olle, R. Zen, M. Puviani et al., “Simultaneous discovery of
quantum error correction codes and encoders with a noise-aware
reinforcement learning agent,” npj Quantum Information, vol. 10,
p. 126, 2024.
[146] J. Edmonds, “Paths, trees, and flowers,” Canadian Journal of
Mathematics, vol. 17, pp. 449–467, 1965.
[147] J. P. Clemens, S. Siddiqui, and J. Gea-Banacloche, “Quantum
error correction against correlated noise,” Phys. Rev. A, vol. 69,
p. 062313, Jun 2004.
[148] D. K. Tuckett, S. D. Bartlett, and S. T. Flammia, “Tailoring surface
codes for highly biased noise,” Physical Review X, vol. 9, no. 4, p.
041031, 2019.
[149] Y. Song, T. Wang, P. Cai, S. K. Mondal, and J. P. Sahoo, “A com-
prehensive survey of few-shot learning: Evolution, applications,
challenges, and opportunities,” ACM Comput. Surv., vol. 55, no.
13s, Jul. 2023.
[150] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE
Transactions on Knowledge and Data Engineering, vol. 22, no. 10,
pp. 1345–1359, 2009.
[151] I. J. Goodfellow, J. Shlens, and C. Szegedy, “Explaining and
harnessing adversarial examples,” in ICLR, 2015.