Thanks to visit codestin.com
Credit goes to www.scribd.com

0% found this document useful (0 votes)
6 views18 pages

Data-Driven Machine Learning Approach

This paper presents a novel data-driven machine learning approach using physics-informed neural networks (PINN) to solve population balance models (PBM), which are essential for understanding particulate system dynamics in process engineering. The proposed method offers a flexible and efficient solution framework that avoids the need for diffusion terms typically required for stability in numerical methods. Numerical comparisons demonstrate that the PINN approach yields accurate results in predicting particle size distributions, validating its effectiveness against traditional methods.

Uploaded by

Amiya Das
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
6 views18 pages

Data-Driven Machine Learning Approach

This paper presents a novel data-driven machine learning approach using physics-informed neural networks (PINN) to solve population balance models (PBM), which are essential for understanding particulate system dynamics in process engineering. The proposed method offers a flexible and efficient solution framework that avoids the need for diffusion terms typically required for stability in numerical methods. Numerical comparisons demonstrate that the PINN approach yields accurate results in predicting particle size distributions, validating its effectiveness against traditional methods.

Uploaded by

Amiya Das
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 18

Ali Advances in Continuous and Discrete Models (2025) 2025:12 Advances in Continuous

https://doi.org/10.1186/s13662-025-03876-1
and Discrete Models

RESEARCH Open Access

Data-driven machine learning approach


based on physics-informed neural network
for population balance model
Ishtiaq Ali1*
*
Correspondence:
[email protected] Abstract
1
Department of Mathematics and
Statistics, College of Science, King Population balance models (PBM) are a fundamental tool in the field of process
Faisal University, P.O. Box 400, engineering and materials science for understanding and forecasting particulate
Al-Ahsa, 31982, Saudi Arabia system dynamics. These models are essential for the design and optimization of a
wide range of industrial processes because they capture the evolution of particle size
distribution in response to different phenomena such as nucleation, growth,
aggregation, and breakage. The complex and non-linear nature of these models
makes it difficult to find the analytical solution and even sometimes the approximate
solution. This paper introduces a novel machine learning approach based on
physics-informed neural network (PINN) for the approximate solution of PBM. Our
strategy utilizes the use of a customized neural network framework that has been
developed and trained on data generated from simulated PBM by using a finite
difference method for the differential operators to identify the underlying dynamics
and patterns controlling the evolution of particle distribution. In general, PINN uses
automatic differentiation for the computation of differential operators, which is based
on the chain rule and needs several matrix operations for computing, which reduces
the processing efficiency during the training. The PINN approach provides a flexible,
effective and highly adaptable solution framework by utilizing neural networks to
approximate the solution and defining the loss function as the sum of the differential
equation’s residuals at specific random or regular points inside the domain, as well as
the residuals of the initial and boundary conditions. It is shown numerically that this
approach does not need a diffusion term for a stable solution, which is often needed
in most numerical methods for solving these models. For further validation, we
compare the results with the exact solution and find them with a very good
agreement with each other.
Keywords: Population balance model; Data-driven machine learning approach;
Physics-informed neural network; Numerical simulations

1 Introduction
Population balances (PB) are crucial for industrial crystallization since they simulate the
formation and features of crystals within a crystallizer, eventually generating the final
solid product. These balances offer a mathematical framework for monitoring changes,

© The Author(s) 2025. Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives
4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format,
as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and
indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived
from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons
licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons
licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain
permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/
4.0/.
Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 2 of 18

especially particle characteristics, such as size, shape, and density, through the crystal-
lization process, as well as for tracking the preservation and transformation of particle
numbers. PB are structured analogous to mass and energy balances, highlighting the non-
conservation of specific particle characteristics and aiming to figure out the relative con-
tributions of multiple factors to these fluctuations. These PB can combine other variables
from the particle phase space, a phrase referring to the multidimensional array of particle
attributes, to describe other significant aspects, although they typically focus on track-
ing particle numbers based on linear size. This procedure uses multiple separate variables
to model the changing properties of particles, allowing for a more detailed analysis that
involves modeling changes in particle form in addition to size. In order to study the dy-
namics of particle populations, Hulburt and Katz developed the population equilibrium
modeling technique [1]. Since then, it has been widely used in a wide range of scientific
fields, such as engineering, physics, chemistry, biology, and meteorology. This method-
ology is especially common in the field of crystallization, where it is the primary method
used to describe the size, shape, and other properties of particles. The approach has been
thoroughly investigated through a combination of theoretical and empirical research. To
enable real-time optimization and control, process engineers need a model that can pre-
dict outcomes much faster than the process occurs. Currently, in the field of particulate
science, controlling characteristics like particle size distribution (PSD) and form is essen-
tial because of their significant influence on the final product and further processing steps.
The PBE is recognized as a key tool for modeling the dynamics of crystallization pro-
cesses and mathematically can be expressed as a partial differential equation, which may
include integral components to accommodate for secondary processes such as agglomer-
ation and particle breaking. These PBE make a substantial contribution to practical appli-
cations by making it possible to optimize industrial processes and enhance the efficiency
of dynamic particle system operations. For processes like crystallization, drug manufac-
turing, and material engineering, it estimates particle size distributions under different
circumstances, supporting strong design strategies. Solving PBE accurately to obtain the
crystal size distribution (CSD) is crucial. Numerous numerical studies have explored PBE,
and each has advantages and disadvantages of its own [2–4]. Among these, the most com-
mon numerical approaches to solving PBE are the method of moments, the method of
characteristics, the weight residuals, the orthogonal collocation method, the Monte Carlo
method, finite-difference schemes, and discrete population balances, respectively [5–8].
Ramkrishna’s review offers valuable insights into these methods [9]. The method of mo-
ments estimates CSD through its moments, facing limitations under complex conditions.
The method of characteristics simplifies PDEs to ODEs along characteristic lines, suitable
for simpler physics, but falters with complexity. Weight residuals and orthogonal colloca-
tion are based on basis functions, with their effectiveness contingent on the choice of these
functions. The Monte Carlo method, though versatile, incurs high computational costs.
Finite-difference and finite-volume methods, adaptable to complex scenarios, demand ex-
tensive grid points for precision. For other notable methods for approximate solutions of
these models, we refer the reader to [10–23].
Gunawan et al.’s development of the high-resolution finite-volume methods (HR-FVM)
is notable for its capacity to compute the particle size distribution (PSD) with accuracy
and to solve PB equations involving agglomeration and breakage processes numerically,
all without the problems of numerical diffusion and dispersion [24]. From a modeling
Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 3 of 18

perspective, this makes HR-FVM the perfect option for optimization and control appli-
cations, while it’s important to keep in mind that the computational requirements can
be very high. A thorough discussion of these techniques may be found in the work by
Qamar et al. in [25]. To improve the high-resolution finite volume methods’ (HR-FVM)
computing efficiency, a number of initiatives have been implemented. To reduce mesh
size, Qamar et al. devised an adaptive mesh approach [26]. A master/slave CPU cluster ar-
chitecture was proposed by Gunawan and his team as a parallel computing strategy [27].
The Fast HR-FVM, created by Majumder and colleagues, uses coordinate transformation
to speed up simulations without sacrificing accuracy. Prakash and associates utilized the
Matlab Parallel Computing Toolbox and Distributed Computing Server to execute HR-
FVM algorithms in parallel on CPUs [28]. Recent advancements have introduced highly
accurate numerical methods for solving the PBE, such as the HR-van method with a flux
limiter, the Lattice–Boltzmann method, the weighted essentially non-oscillatory (WENO)
method, and spectral methods. Notably, spectral methods, particularly the spectral collo-
cation method, stand out for their efficiency and lower grid point requirements compared
to other approaches [29–31].
Given the PBE’s nature as a convection–reaction equation with pronounced hyperbolic
characteristics, these features pose challenges to stability in numerical solutions. To avoid
this, a diffusive term is introduced, enhancing stability at the expense of some deviation
from the exact solution. This adjustment, while affecting accuracy, ensures the reliability
of the numerical outcomes. Population balances are a typical engineering process used
in drug crystal formation, pollutant production in fires, and the growth of microbial and
cell populations [32]. The aim of this work is to introduce a novel and stable technique
that does not need the addition or deletion of diffusion terms to be stable. To this end, we
apply a data-driven technique based on PINN for the approximate solution of PBE. The
main advantage of the PINN method over other traditional approaches is that it produces
a prediction function over the entire computational domain rather than a discrete solu-
tion on the meshes as in mesh-based methods; they are applicable to high-dimensional
problems; and they are mesh-free, avoiding the problems of mesh generation on complex
regions and the construction of high-precision discrete schemes on meshes of poor geo-
metric quality. The primary goal of this work is to optimize the loss function made up of
PBM residuals by training neural networks to approximate the equation’s solution.
The general form of the Population Balance Equation (PBE) can be expressed as:

∂ F (L, t) ∂ ∂ 2F
+ [F (L, t)G (L, t)] +  2 (L, t) = H(L, t, F ). (1)
∂t ∂L ∂L

The function F (L, t) represent the population density, which describes how the distribu-
tion of a given characteristic L (which may be age, size, or some other measurable prop-
erty) changes with time t. The expression ∂∂tF indicates the rate of variation of this density
with respect to time, giving insight on the population’s dynamic behavior. A growth or
velocity function is included via the inclusion of G (L, t) in the equation, specifically in
the term ∂∂L [F (L, t)G (L, t)], which accounts for changes in the characteristic L brought
about by growth or other processes and  > 0 is a positive constant. For the purpose of
simulating events like cell growth, particle aggregation, or the dissemination of features
within a population, this section of the equation is crucial. In addition, when taking into
Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 4 of 18

consideration the variability and distribution of features, the diffusion term ∂∂ LF2 repre-
2

sents the dispersion or spread in the characteristic L throughout the population. It also
takes into consideration the dispersion or fluctuation of particle properties, including size,
within a system, help in modeling the inherent unpredictability in particle systems, and
depicts dispersion effects resulting from stochastic or random processes. Furthermore,
especially in situations where convection predominates, it stabilizes numerical solutions
by reducing abrupt gradients or discontinuities. In order to preserve computational sta-
bility and appropriately represent physical dispersion, this term is essential. In the final
analysis, depending on the particular application, the function H(L, t, F ) acts as a source
or sink term, such as external variables that influence the population density, for example,
migration, births and deaths, or chemical reactions. This expression makes it possible to
include complex connections between outside variables, which significantly improves the
model’s adaptability to a wide range of scientific investigations.

2 Physics-informed neural network technique for PBE


Artificial neural networks have been utilized to solve issues in many different applica-
tion domains over the past few decades, including computer vision and natural language
processing, among many others. The scientific machine learning (ML) community has re-
cently seen the emergence of another extremely promising application: the use of artificial
neural networks, commonly known as PINN, to solve partial differential equations (PDE).
Since its initial introduction in the landmark study in [33], PINN have been gaining sig-
nificant importance in both academia and industry. In PINN, in addition to training the
model with data-driven supervised neural networks, the model is taught physics equations
to promote consistency with the system’s understood physics. They have the advantage of
being able to reliably extrapolate beyond the existing data and being data-driven in learn-
ing a model, while still ensuring consistency with physics. Because of this, PINN can pro-
duce stronger models with fewer data points. It combines the flexibility of artificial neural
networks with the benefits of classical numerical methods to provide a potential tool for
solving PDE. By adding the physical aspects of a PDE into the training process, PINN are
designed to learn the solution to a PDE. This method can greatly increase training speed
and accuracy by integrating known data points with a physics-based loss function.
PINN has shown great promise as a viable substitute for more conventional numerical
techniques like Finite Element Methods (FEMs). It uses the physical principles inherent
in the PDE to direct the learning process. This makes it possible to include known data
points to the training process, which can improve the solution’s accuracy and effectiveness
even further. The capacity of PINN to integrate the physical characteristics of the PDE into
the learning process is its main benefit. By doing this, the network is guaranteed to under-
stand the fundamental physical linkages and restrictions, which results in more precise
and dependable solutions. PINN is an invaluable tool for resolving a variety of issues in
a variety of disciplines, including engineering, physics, and finance. They can also handle
complex and nonlinear PDE. To futher explore PINN, we refer the reader to [34–40].
To apply PINN to the given model in equation (1), a neural network architecture that
can approximate the solution was designed, as shown in Fig. 1.
This network takes inputs of spatial and temporal coordinates and outputs an approxi-
mation of the function F (L, t), which represents the exact population density. Let us de-
note the neural network as a function F (x, t; θ ), where x and t are the spatial and temporal
Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 5 of 18

Figure 1 Schematic of the proposed physics-informed neural network for solving the model equation (1)

coordinates, respectively, and θ represents the network approximation of F (L, t). The net-
work approximate the solution to the model equation of the form:

F (x, t; θ ) ≈ F (L, t). (2)

Define the loss function to ensure the network satisfies the physical laws, boundary and
initial conditions, and observational data of the form:

L(θ ) = LPDE (θ ) + λ1 LBC (θ ) + λ2 LIC (θ ) + λ3 Ldata (θ ). (3)

Here the term LPDE enforces the satisfaction of the PBE:

NPDE  2
1   ∂ Fθ ∂   ∂ 2 Fθ 
LPDE (θ ) =  
NPDE  ∂t + ∂ L G (L, t)Fθ +  ∂ L2 – H(L, t, Fθ ) , (4)
i=1

The term LBC enforces boundary conditions and is given by:

1 
NBC
 
LBC (θ ) = Fθ (LBC , tBC ) – g(LBC , tBC )2 , (5)
NBC i=1

where g(LBC , tBC ) specifies the boundary values, while the term LIC enforces adherence
to the initial condition and is defined by:

1 
NIC
LIC (θ ) = |Fθ (LIC , 0) – h(LIC )|2 , (6)
NIC i=1

where h(LIC ) is the initial distribution. The last term Ldata matches observational or sim-
ulated data and is given by:

1 
Ndata
Ldata (θ ) = |Fθ (Ldata , tdata ) – FFDM (Ldata , tdata )|2 . (7)
Ndata i=1
Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 6 of 18

Here FFDM is the approximate solution using FDM. For our PDE, we discretize the domain
into a grid of points over L and t, and approximate the derivatives as follows:
For time derivative approximation using a forward difference scheme, the time deriva-
tive can be approximated by:

∂F F (L, t + t) – F (L, t)


≈ . (8)
∂t t

Similarly, the first spatial derivative, using a central difference scheme, is approximated
by:

∂ F (L + L, t)G (L + L, t) – F (L – L, t)G (L – L, t)


[F (L, t)G (L, t)] ≈ . (9)
∂L 2L

The second spatial derivative is given by:


 
∂ 2F F (L + L, t) – 2F (L, t) + F (L – L, t)
 ( L , t) ≈  . (10)
∂ L2 (L)2

The fully discretized form of the model equation (1) using FDM is given by:
 n n
Fin+1 – Fin F G + Fin+1 Gin+1 Fi+1 – 2Fi + Fi–1
+ i i + – Hin = 0. (11)
t 2L L2

The terms λ1 , λ2 and λ3 are the weights for BC, IC, and data loss, respectively. The sim-
ple approach is to choose λi = 1, for all i, that is, all loss components are treated equally
initially, and adjustments can be made based on validation errors using normalization, if
needed. The training dataset is generated using a FDM to solve the PBE over a grid of
spatial and temporal coordinates, incorporating predefined boundary and initial condi-
tions. The validation dataset is derived similarly but uses different resolutions or parame-
ter variations to evaluate the generalization of the model during training. Testing datasets
are constructed to assess the model’s performance on unseen scenarios, often involving
altered parameters, broader domains, or challenging edge cases. By comparing PINN pre-
dictions against exact solutions or independent numerical methods, we ensure the model’s
robustness and accuracy beyond the training conditions.

3 Numerical examples
In this section, we present two numerical examples for demonstrating the strength and
effectiveness of PINN. From these examples, we are able to see directly how PINN use the
underlying physical rules to direct the learning process, ensuring both accuracy and gener-
alizability in different situation. This practical strategy demonstrates how the models can
solve complicated differential equations and accurately anticipate physical processes, fre-
quently outperforming traditional numerical methods in terms of speed and scalability.
Furthermore, by comparing the results of these examples with the exact solution, it be-
comes clear how efficient and flexible PINN are to a range of difficult-to-solve problems.
Both below examples are chosen for [29].

3.1 Example 1
Consider the classical example of crystal growth in a batch process under specific assump-
tions. We consider a system where the crystal growth rate G is constant and independent
Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 7 of 18

of crystal size L. Processes like aggregation, nucleation, and breakage are ignored, sim-
plifying the system to totally focus on crystal growth. The crystal growth rate is given as
G = 1.0 mm/s, and the CSD follows a Gaussian distribution defined by:
 
1010 (L – μ)2
F (L, 0) = √ exp – , (12)
2πσ 2σ 2

This Gaussian distribution is chosen because it accurately reflects symmetric changes


around a mean size, which is common in crystallization processes. It is simpler to compute
and evaluate. The parameters μ and σ are the mean and standard deviation of the crystal
size distribution, respectively, with μ = 20μm and σ = 3. The simulation parameters in-
clude a time step dt = 0.001 s, a parameter  = 0.005, and the end time of the simulation
tend = 60 s. Given that the crystal growth rate G is constant, the exact solution for the CSD
at any time t can be found by shifting the initial distribution by G t, that is:

F (L, t) = F (L – G t, 0). (13)

This equation states that the entire distribution of crystal sizes shifts uniformly as time
progresses, with no change in the shape of the distribution, due to the constant growth rate
applied equally to all crystals. To illustrate this, we simulate the process and plot the initial
CSD and the CSD at t = 30 s and at t = 60 s to show the shift in the crystal growth process
as shown in Fig. 2 and Fig. 3 for N = 200 and N = 600, respectively. A complete horizontal
shift without any change in shape or spread is shown by comparison with the exact solu-
tion for growth that is independent of size. The exact solution and PINN are compared to
indicate how well the PINN trained on FDM data follows this ideal behavior. This com-
parison may reveal small differences, such as a tiny numerical diffusion or a larger spread,
inherited from the FDM training data, as shown in Figs. 2–3. These images illustrate the

Figure 2 Example 1: Crystal size distribution over different time with N = 200 grid points
Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 8 of 18

Figure 3 Example 1: Crystal size distribution over different time with N = 600 grid points

ability of PINNs to handle hyperbolic dynamics typical of convection-dominated systems


in size-independent growth situations, where the crystal growth rate G is constant. As can
be seen from the model’s stability in these figures, dispersion and oversmoothing—two
prominent problems in numerical simulations of hyperbolic equations—are not captured
by the neural network when simulating the linear advection of crystal sizes. How well the
PINN handles edge cases is shown by closely examining figures near borders (e.g., small
or big crystal sizes). Given how difficult it can be to enforce boundary criteria in neural
networks, effective border behavior free of oscillations or instabilities indicates that the
network has learned the required boundary conditions implicitly. A reliable and safe crys-
tallization process depends on precisely modeling the whole range of crystal sizes, which
is ensured by good boundary behavior. This is especially true when scaling up from lab to
industrial scales.
We observe that while increasing the grid points, one can get closer and closer to the
exact solution. For the clear understanding the comparison between exact and PINN so-
lution, we also provide subplot at different time, as shown in Fig. 4. In order to validate our
PINN solution, we also train our neural network using the exact initial conditions and the
result is shown in Fig. 5. The distribution of errors between expected and actual values is
visually represented by a neural network training error histogram, which aids in identify-
ing underfitting or overfitting tendencies during the model training process. This is shown
in Fig. 6. How well a neural network model generalizes from its training data to new data
depends on how well it performs during training. This is usually evaluated utilizing met-
rics like accuracy and precision. In case of Example 1, the best validation performance
is epoch 10, as shown in Fig. 7. When a neural network is trained for regression tasks,
it learns to predict continuous output values. The mean squared error or mean absolute
error between the predicted and actual values is often used to evaluate the network’s per-
formance, as shown in Fig. 8. Training states of neural networks, including weights and
Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 9 of 18

Figure 4 Example 1: Subplot of crystal size distribution over different time with N = 600 grid points

Figure 5 Example 1: Comparison of exact initial condition and predicted by PINN

biases, change as a result of techniques like optimization algorithms and backpropagation,


which seek to effectively decrease the loss function, as shown in Fig. 9.
Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 10 of 18

Figure 6 Example 1: Neural network training error histogram

Figure 7 Example 1: Neural network training performance

3.2 Example 2
In the second example, crystallization takes place in a batch process, where the crystal size
L determines the crystal growth rate G . The crystal growth rate G (L, t) = G0 L is a linear
function of crytal size. Again, H(L, t, F ) = 0 as in Example 1 and the following equation is
satified by CSD:
 
N0 L
F (L, 0) = exp – ,
L L̄
Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 11 of 18

Figure 8 Example 1: Neural network training regression

Figure 9 Example 1: Neural network training states


Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 12 of 18

Figure 10 Example 2: Crystal size distribution over different time for N = 50 grid points

where N0 and L̄ are constants. The exact solution to this equation is given by:
 
N0 L
F (L, t) = exp – exp(–G0 t),
L L̄e–G0 t

with parameters values, L̄ = 0.01 μm3 , N0 = 1, G0 = 0.1 (μm3 )/s,  = 0.002, and dt =
0.00001 s. This example deals with size-dependent growth in a batch process, where the
difficulty rises as the size-dependent growth rate changes. Over time, the size distribution
becomes distorted and broader due to the size-dependent growth rate, which makes larger
particles grow more quickly. In addition to making numerical approaches more difficult,
this non-linear behavior creates steeper gradients and shifts in the distribution, proving
that the PINN can properly capture intricate, size-dependent dynamics. With the help of
FDM data, PINN is trained to reproduce how the FDM handles these fluctuations; this is
often demonstrated by a shift and distortion (broadening or stretching) of the CSD. Larger
crystals develop more quickly, as shown in Fig. 10–11. This also shows that evolution of
the distribution and any numerical distortions common to FDM, such as small oscillations
or inconsistencies at locations where the growth rate varies quickly. Comparing the exact
solution, a more theoretically ideal transformation of the CSD with a clear representa-
tion of the non-linear growth effects can be seen in the exact solution for size-dependent
growth. The neural network’s ability to capture the complex dynamics trained on FDM
data is demonstrated by comparing this to the PINN predictions, which highlights devi-
ations such as over-smoothing and underestimating steep gradients. In order to see the
comparison between exact and PINN predicted solution, a more clear view is given in the
subplot consist of Fig. 12. The validation of the PINN performace is evaluated against the
exact initials condition and the output is given in the form of Fig. 13.
These results show how accurate predictions are for size-dependent growth, where the
growth rate G (L) varies with crystal size, which further confirm that the PINN can man-
Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 13 of 18

Figure 11 Example 2: Crystal size distribution over different time for N = 100 grid points

Figure 12 Example 2: subplot of crystal size distribution over different time for N = 100 grid points

age non-linear dependencies and interactions within the system. This demonstrates the
network’s capacity to discover intricate linkages and patterns from the data and ingrained
physical rules rather than from explicit programming. The robustness of PINNs against
parameter uncertainty can be demonstrated by looking at how different factors (such as
growth rates or initial circumstances) affect the model’s output. Figures 10–13 also illus-
trate that long-term forecasts over protracted times or until steady state is reached can
Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 14 of 18

Figure 13 Example 2: Comparison of exact initial condition and predicted by PINN

show how the PINN can continue to be accurate and stable over time without straying
from or drifting from physical expectations. Understanding the efficacy of neural networks
in practical applications requires evaluating their training performance. Performance met-
rics reveal how well the network predicts new data. The architecture of the network, which
includes the quantity and size of layers as well as the activation functions employed, has
an impact on training performance as well. Furthermore, the quantity and caliber of train-
ing data are critical factors in determining how well a neural network learns. One can use
sophisticated methods such as cross-validation to assess how resilient the network is to
variations in data subsets. The best performance result is at epoch 7, which is shown in
Fig. 14. An error histogram for diagnosing model behavior, especially when it comes to
determining bias and variance problems is given in Fig. 15. Figure 16 indicates the train-
ing states for Example 2 examine how well the model reflects the data trends. Techniques
such as feature scaling, correct network parameter setup, and regularization to avoid over-
fitting improve the performance of the regression. Plotting anticipated vs. actual values is
a common method of evaluating regression models in order to visually examine how well
the model reflects the data trends is shown in Fig. 17.

4 Results and discussion


First, we approximate the solutions of model equation (1) over a discrete grid using the
FDM. By solving the equation step-by-step through the domain, FDM produces data
points that show the state of the system at different moments in time. Despite being an
approximation with underlying numerical faults, such as discretization error, this data of-
fers a comprehensive dataset reflecting the numerical features of the FDM as well as the
underlying physics. The neural network learn from this training dataset, which comprises
both the ideal physical rules and the real-world numerical behaviors frequently observed
in classical simulations. The training dataset is then used as training data for PINN. This
dual feature of the data aids PINN in creating a solid model sensitive to both computational
Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 15 of 18

Figure 14 Example 2: Neural network training performance

Figure 15 Example 2: Neural network training error histogram

and theoretical sensitivities. These PINN are designed so that their architecture immedi-
ately benefits from an understanding of the governing differential equations by including
the differential equation into the loss function. Consequently, the network is trained to
satisfy the physical principles governing the process in addition to fitting the training data
(from FDM). The neural network weights must be adjusted during the training phase to
guarantee that the predictions are consistent with the physical equations and to reduce
the discrepancy between the network’s predictions and the FDM data. In addition to im-
Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 16 of 18

Figure 16 Example 2: Neural network training states

Figure 17 Example 2: Neural network training regression


Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 17 of 18

proving generalization, this dual purpose also help PINN address some of the numerical
errors in the FDM data.

5 Conclusion
In this study, the FDM is employed in order to generate training data, which is then used
to train a PINN and predict how the particle distribution will change over time. To verify
their accuracy and efficiency, the outcomes from the PINN are compared with those from
the exact solution. It has been demonstrated that more complex and optimal crystalliza-
tion techniques, which may dynamically modify parameters in response to changes in the
crystal size distribution, require accurate modeling of size-dependent growth. In indus-
trial applications where process conditions may alter or be unpredictable, this robustness
is very valuable. Models that function well under a variety of circumstances lower the pos-
sibility of process failures and boost overall effectiveness. For continuous crystallization
techniques to work and for the development of strategies that predict how process modi-
fications will impact the end product over time, long-term stability is essential. In conclu-
sion, the integration of numerical simulation and machine learning to PBMs through the
use of finite difference methods to produce training data for PINNs enables the creation
of reliable models that can produce precise predictions under a variety of circumstances.
If the model configuration, training procedure, and data quality are carefully considered,
it presents an achievable path toward addressing complex systems efficiently and dynam-
ically.

Acknowledgements
This work was supported by the Deanship of Scientific Research, Vice Presidency for Graduate Studies and Scientific
Research, King Faisal University, Saudi Arabia [Grant No. KFU250109].

Author contributions
The author read and approved the final manuscript.

Data availability
Not applicable.

Declarations
Competing interests
The author declares no competing interests.

Received: 31 October 2024 Accepted: 6 January 2025

References
1. Hulburt, H.M., Katz, S.: Some problems in particle technology: a statistical mechanical formulation. Chem. Eng. Sci. 19,
555–574 (1964)
2. Rasmuson, Å.C.: Crystallization process analysis by population balance modeling. In: Myerson, A.S., Erdemir, D., Lee,
A.Y. (eds.) Handbook of Industrial Crystallization, pp. 172–196. Cambridge University Press, Cambridge (2019)
3. Lin, F., Yang, Y., Yang, X.: Exact solutions of population balance equation with aggregation, nucleation, growth and
breakage processes, using scaling group analysis. Symmetry 16, 65 (2024)
4. Le Minh, T., Phan Thanh, T., Nguyen Thi Hong, N., Phan Minh, V.: A simple population balance model for crystallization
of L-lactide in a mixture of n-hexane and tetrahydrofuran. Crystals 12, 221 (2022)
5. Inguva, P.K., Schickel, K.C., Braatz, R.D.: Efficient numerical schemes for population balance models. Comput. Chem.
Eng. 162, 107808 (2022)
6. Su, J., Le, W., Gu, Z., Chen, C.: Local fixed pivot quadrature method of moments for solution of population balance
equation. Processes 6, 209 (2018)
7. Wang, K., Yu, S., Peng, W.: A new method for solving population balance equations using a radial basis function
network. Aerosol Sci. Technol. 54, 644–655 (2020)
8. Röhl, S., Hohl, L., Stock, S., Zhan, M., Kopf, T., von Klitzing, R., Kraume, M.: Application of population balance models in
particle-stabilized dispersions. Nanomaterials 13, 698 (2023)
9. Ramkrishna, D., Singh, M.R.: Population balance modeling: current status and future prospects. Annu. Rev. Chem.
Biomol. Eng. 5, 123–146 (2014)
Ali Advances in Continuous and Discrete Models (2025) 2025:12 Page 18 of 18

10. Immanuel, C.D., Doyle, F.J. III: Computationally efficient solution of population balance models incorporating
nucleation, growth and coagulation: application to emulsion polymerization. Chem. Eng. Sci. 58, 3681–3698 (2003)
11. Tiong, S.I.X., Ahamed, F., Sitaraman, H., Leong, S.L., Ho, Y.K.: Modeling simultaneous particle shrinkage, dissolution and
breakage using the modified moving grid technique. Powder Technol. 421, 118439 (2023)
12. Huang, Y., Zhu, H., Yang, X., Tu, J., Jiang, S.: Population balance modeling for air–water bubbly flow in a vertical
U-bend. J. Comput. Multiph. Flows 10, 170–177 (2018)
13. Li, D., Li, Z., Gao, Z.: Quadrature-based moment methods for the population balance equation: an algorithm review.
Chin. J. Chem. Eng. 27, 483–500 (2019)
14. Qamar, S., Angelov, I., Elsner, M.P., Ashfaq, A., Seidel-Morgenstern, A., Warnecke, G.: Numerical approximations of a
population balance model for coupled batch preferential crystallizers. Appl. Numer. Math. 59(3–4), 739–753 (2009)
15. Qamar, S., Ashfaq, A., Angelov, I., Elsner, M.P., Warnecke, G., Seidel-Morgenstern, A.: Numerical solutions of population
balance models in preferential crystallization. Chem. Eng. Sci. 63, 1342–1352 (2008)
16. Solsvik, J., Jakobsen, H.A.: Evaluation of weighted residual methods for the solution of a population balance model
describing bubbly flows: the least-squares, Galerkin, tau, and orthogonal collocation methods. Ind. Eng. Chem. Res.
52, 15988–16013 (2013)
17. Sewerin, F., Rigopoulos, S.: An explicit adaptive grid approach for the numerical solution of the population balance
equation. Chem. Eng. Sci. 168, 250–270 (2017)
18. Shah, B.H., Ramkrishna, D., Borwanker, J.D.: Simulation of particulate systems using the concept of the interval of
quiescence. AIChE J. 23, 897–904 (1977)
19. Gooch, J.R.P., Hounslow, M.J.: Monte Carlo simulation of size-enlargement mechanisms in crystallization. AIChE J. 42,
1864–1874 (1996)
20. Kumar, S., Ramkrishna, D.: On the solution of population balance equations by discretization—I. A fixed pivot
technique. Chem. Eng. Sci. 51, 1311–1332 (1996)
21. Ma, D.L., Tafti, D.K., Braatz, R.D.: Ind. Eng. Chem. Res. 41, 6217–6223 (2002)
22. Majumder, A., Kariwala, V., Ansumali, S., Rajendran, A.: Lattice Boltzmann method for multi-dimensional population
balance models in crystallization. Chem. Eng. Sci. 70, 121–134 (2012)
23. Ruan, C.L., Liang, K.F., Chang, X.J., Zhang, L.: Weighted essentially nonoscillatory method for two-dimensional
population balance equations in crystallization. Math. Probl. Eng. 2013, 125128 (2013)
24. Gunawan, R., Fusman, I., Braatz, R.D.: High resolution algorithms for multidimensional population balance equations.
AIChE J. 50, 2738–2749 (2004)
25. Qamar, S., Elsner, M.P., Angelov, I.A., Warnecke, G., Seidel-Morgenstern, A.: A comparative study of high resolution
schemes for solving population balances in crystallization. Comput. Chem. Eng. 30, 1119–1131 (2006)
26. Qamar, S., Warnecke, G.: Solving population balance equations for two-component aggregation by a finite volume
scheme. Chem. Eng. Sci. 62, 679–693 (2007)
27. Gunawan, R., Fusman, I., Braatz, R.D.: Parallel high-resolution finite volume simulation of particulate processes. AIChE J.
54, 1449–1458 (2008)
28. Prakash, A.V., Chaudhury, A., Barrasso, D., Ramachandran, R.: Simulation of population balance model-based
particulate processes via parallel and distributed computing. Chem. Eng. Res. Des. 91, 1259–1271 (2013)
29. Ruan, C.: Chebyshev spectral collocation method for population balance equation in crystallization. Mathematics 7,
317 (2019)
30. Mantzaris, N.V., Daoutidis, P., Srienc, F.: Numerical solution of multi-variable cell population balance models. II. Spectral
methods. Comput. Chem. Eng. 25, 1441–1462 (2001)
31. Buck, A., Klaunick, G., Kumar, J., Peglow, M., Tsotsas, E.: Numerical simulation of particulate processes for control and
estimation by spectral methods. AIChE J. 58, 2309–2319 (2012)
32. Ramkrishna, D.: Population Balances: Theory and Applications to Particulate Systems in Engineering. Academic Press,
San Diego (2000)
33. Raissi, M., Perdikaris, P., Karniadakis, G.E.: Physics-informed neural networks: a deep learning framework for solving
forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys. 378, 686–707 (2019)
34. Karniadakis, G.E., Kevrekidis, I.G., Lu, L., Perdikaris, P., Wang, S., Yang, L.: Physics informed machine learning. Nat. Rev.
Phys. 3, 422–440 (2021)
35. Zhu, Y., Zabaras, N., Koutsourelakis, P.S., Perdikaris, P.: Physics-constrained deep learning for high-dimensional
surrogate modeling and uncertainty quantification without labeled data. J. Comput. Phys. 394, 56–81 (2019)
36. Faroughi, S.A., Pawar, N.M., Fernandes, C., Raissi, M., Das, S., Kalantari, N.K., Mahjour, K.: Physics-guided,
physics-informed, and physics-encoded neural networks and operators in scientific computing: fluid and solid
mechanics. ASME J. Comput. Inf. Sci. Eng. 24, 040802 (2024)
37. Mattey, R., Ghosh, S.: A novel sequential method to train physics informed neural networks for Allen–Cahn and
Cahn–Hilliard equations. Comput. Methods Appl. Mech. Eng. 390, 114474 (2022)
38. Cuomo, S., Di Cola, V.S., Giampaolo, F., et al.: Scientific machine learning through physics–informed neural networks:
where we are and what’s next. J. Sci. Comput. 92, 88 (2022)
39. Raissi, M., Karniadakis, G.E.: Hidden physics models: machine learning of nonlinear partial differential equations. J.
Comput. Phys. 357, 125–141 (2018)
40. Yang, L., Meng, X., Karniadakis, G.E.: B-PINNs: Bayesian physics-informed neural networks for forward and inverse PDE
problems with noisy data. J. Comput. Phys. 425, 109913 (2021)

Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

You might also like