Journal Pre-proof

Integrated convolutional and graph neural networks for predicting

mechanical fields in composite microstructures

Marwa Yacouti, Maryam Shakiba

PII: S1359-835X(24)00616-X
DOI: https://doi.org/10.1016/j.compositesa.2024.108618
Reference: JCOMA 108618
To appear in: Composites Part A
Received date : 25 October 2024
Accepted date : 24 November 2024

Please cite this article as: M. Yacouti and M. Shakiba, Integrated convolutional and graph neural
networks for predicting mechanical fields in composite microstructures. Composites Part A (2024),
doi: https://doi.org/10.1016/j.compositesa.2024.108618.

This is a PDF file of an article that has undergone enhancements after acceptance, such as the
addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive
version of record. This version will undergo additional copyediting, typesetting and review before it
is published in its final form, but we are providing this version to give early visibility of the article.
Please note that, during the production process, errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.

© 2024 Published by Elsevier Ltd.

 Journal Pre-proof

Graphical Abstract

of
Integrated Convolutional and Graph Neural Networks for Predict-
ing Mechanical Fields in Composite Microstructures

pro
Marwa Yacouti, Maryam Shakiba
Data-Driven Analysis Results
Composite Linear Nonlinear
microstructures stress field stress field

Deep
learning
framework

re- CompINet
Input: 256 × 256 × 1

Binary
Architecture
128 × 32 × 32

representation
CNN
Encoder
Output: 256 × 256 × 1
128 × 32 × 32

CNN
lP
Decoder
128 × 32 × 1

𝑛" 𝑛#
𝑛$
𝑛! GNN
• Coordinates of the 𝑛' 𝑛%
centers of the fibers 𝑛&
• Connectivity matrix
rna
Jou
 Journal Pre-proof

Highlights

of
Integrated Convolutional and Graph Neural Networks for Predict-
ing Mechanical Fields in Composite Microstructures

pro
Marwa Yacouti, Maryam Shakiba

• The proposed framework, CompINet, employs a novel combination of

convolutional and graph neural networks to predict the mechanical re-

re-
sponses of composites.

• CompINet predicts both the linear and nonlinear stress distribution

within the microstructural representation of composites.
lP
• CompINet requires 20 times less data compared to existing data-driven
approaches.

• CompINet is not only more accurate but also more consistent and re-
rna

liable.
Jou
 Journal Pre-proof

Integrated Convolutional and Graph Neural Networks

of
for Predicting Mechanical Fields in Composite
Microstructures

pro
Marwa Yacoutia , Maryam Shakibaa,∗
a
Smead Department of Aerospace Engineering Sciences, University of Colorado
Boulder, 3775 Discovery Dr, Boulder, 0309-0429, Colorado, USA

Abstract
re-
This paper introduces CompINet, a novel approach that leverages graph
and convolutional neural networks to predict mechanical fields within mi-
crostructural representations of composites. Analyzing local mechanical fields,
lP
such as stress in composites, is crucial for predicting performance and failure,
and planning repair strategies. The critical role of the fiber’s nearest neighbor
distances in shaping linear and nonlinear stress responses within the compos-
rna

ite’s microstructure motivates our approach. CompINet exploits the power

of graph neural networks to capture the microscale intricacies of compos-
ites, particularly the locations of the fibers and the distances between them.
The proposed framework demonstrates remarkable accuracy and consistency
in predicting microscale mechanical fields, requiring 20 times less data than
Jou

existing data-driven methods. CompINet offers significant improvements in

both linear and nonlinear composite analyses.

∗
Corresponding author:
Email address: [email protected] (Maryam Shakiba)

Preprint submitted to Composites Part A October 25, 2024

 Journal Pre-proof

Keywords: machine learning, composites, linear and nonlinear stress,

of
graph neural network

1. Introduction

pro
Complex fiber-reinforced composites are prevalent in natural and engi-
neering domains. For instance, numerous soft tissues within the human body
consist of highly oriented collagen fibers placed in a hydrated matrix, endow-
ing them with exceptional load-bearing and energy absorption capabilities [1].
re-
Similarly, unidirectional fibers embedded in epoxy matrices find applications
across aerospace, automotive, biomedical, soft robotics, and civil engineering
fields. A solid understanding of the mechanics of fiber-reinforced composites
lP
is essential for developing effective strategies for repair and composite design.
The performance and failure of these materials are largely influenced by lo-
cal mechanical fields such as stress, strain, and damage, which vary spatially
depending on the geometry and behavior of the constituent materials [2].
rna

Hence, accurately predicting such local mechanical responses under applied

loads is key to improving the design and durability of composite structures.
Solutions based on laminate plate theories are well-established and com-
monly used to derive the governing equations of composite laminates [3].
Jou

Both analytical and numerical methods were employed to solve such govern-
ing equations and predict the response of composites under various loading
conditions [4, 5, 6, 7, 8, 9, 10, 11, 12, 3]. The Navier and Lévy methods are two
widely used classical approaches that provide exact solutions for the displace-

2
 Journal Pre-proof

ment field in laminated composite rectangular plates [3]. Navier solutions are

of
applied to composite plates with simply supported boundaries [13, 10], and
Lévy’s method is used when two opposing edges are simply supported, while

pro
the other two edges can have arbitrary boundary conditions [14]. Over the
years, numerous analytical models have been developed, but they are limited
in flexibility and applicability, as they are formulated for specific geometries,
boundary conditions, and loading cases [15, 16, 17, 18].
To overcome these limitations, numerical methods like finite element anal-
re-
ysis (FEA) have expanded the range of solvable problems, accommodating
more flexible boundary conditions and complex geometries [7, 19, 20, 21]. Un-
like analytical methods, finite element techniques have been established as
lP
a powerful tool for predicting local crack initiation and propagation in com-
posite laminates [22]. For instance, the Extended Finite Element Method
(XFEM) enhances traditional FEA by allowing the representation of dis-
continuities, such as cracks, without requiring the mesh to conform to the
rna

crack’s geometry [23, 24]. Another promising approach is the phase field
method, which provides a continuous description of crack evolution [25, 26].
Meanwhile, Cohesive Zone Models (CZMs) can simulate delamination and
interface damage by incorporating cohesive elements that model the traction-
Jou

separation laws at material interfaces [27, 28, 29]. However, while XFEM ef-
fectively avoids mesh refinement challenges, it struggles with modeling ran-
dom crack growth without predefined paths. In contrast, the phase field
method excels at capturing cracks without predetermined patterns but is

3
 Journal Pre-proof

computationally expensive due to its fine mesh requirements [30]. Simi-

of
larly, although CZMs are effective for simulating interface damage, the ma-
trix damage model needs to be implemented separately. Therefore, even

pro
with the reduced computational cost of using supercomputers, analyzing the
microstructural responses of composites, especially with high fiber volume
fractions, remains a challenging task and can take several days on supercom-
puting facilities [31]. Given these challenges, FEA remains an active area of
research as new techniques are continuously developed to enhance accuracy
and efficiency [32].
re-
More recently, machine learning (ML) methods have emerged to address
some of the challenges associated with FEA, particularly by providing faster
lP
surrogate models [33, 34]. ML approaches have been employed in two distinct
ways. First, data-driven methods utilize measurement or numerical simula-
tion data from physical systems to train the ML model. This ML model
relates inputs to outputs, without the need to solve physics-based governing
rna

equations. Once trained, the ML model can make predictions within the
“interpolation” range, where model inputs closely resemble data from the
training set. In the second approach, the ML algorithm solves physical laws
as an optimization problem without the need to provide any training data.
Jou

This class of models is commonly known as Physics-Informed Neural Net-

works (PINN) [35, 36]. This paper focuses on the first approach, where data
is used to train the ML algorithm, which will be the subject of the rest of the
literature review and the proposed method. Moreover, we focus on literature

4
 Journal Pre-proof

where ML approaches have been used to predict the distribution of mechan-

of
ical fields within microstructural representations of composites, rather than
their global homogeneous responses.

pro
Image-based ML approaches, where the input and output data are pre-
sented as pixilated information, were used to predict the distribution of me-
chanical field within microstructural representations of composites. Con-
volutional Neural Networks (CNNs) were employed to predict linear stress
distributions [37, 34, 38, 39, 40, 41], as well as nonlinear stress and damage re-
re-
sponses [42, 43, 44, 45, 46] within two-dimensional (2D) microstructural rep-
resentations of fiber-reinforced composites. Moreover, conditional Generative
Adversarial Networks (cGANs) have been applied to predict stress and strain
lP
fields in microstructural representations of composites [47, 48]. Furthermore,
Graph Neural Networks (GNNs), which operate on graph-structured data,
have also been utilized, employing mesh-to-graph mapping to capture the re-
lationship between composite microstructure and mechanical response [49].
rna

These methods were mostly used to predict the microstructural response of

composites at a given applied strain level. To model the temporal evolution
of local stress fields in composites, bidirectional Long Short-Term Memory
(LSTM) networks, which capture temporal dependencies by processing in-
Jou

formation both forward and backward in time, have been adapted [43]. ML
approaches have been shown to provide accurate predictions for the mechan-
ical properties of composites while significantly accelerating their analysis by
several orders of magnitude [34, 50].

5
 Journal Pre-proof

In summary, several ML methods have been utilized to predict the distri-

of
bution of mechanical fields in fiber-reinforced composites, and comparative
studies have evaluated their performance. Sepasdar [48] demonstrated that

pro
a U-Net-based convolutional neural network model achieved similar accu-
racy to a cGAN in predicting both linear and nonlinear von Mises stress
field in 2D microstructures of fiber-reinforced composites. Both models were
trained on 8000 samples, but the U-Net model trained faster, completing in
8 hours compared to 12 hours for the cGAN. In a separate study, Yacouti
re-
and Shakiba [41] compared two CNN architectures: U-Net and Residual Net-
work (ResNet). Using a training data set of 2000 samples, U-Net achieved
a mean R2 value of 0.96 across the test set, outperforming ResNet, which
lP
achieved an R2 of 0.93 for the prediction of the linear von Mises stress field
in fiber-reinforced composites.
While ML approaches provide accurate and faster solutions compared to
finite element analysis [34, 50], they also have certain limitations. A primary
rna

challenge lies in the significant computational cost associated with generating

the required training data. These models often necessitate large training data
sets, typically between 1000 and 2000 samples [47, 34, 41]. These studies rely
on finite element analysis to generate their training data, representing a sig-
Jou

nificant computational cost. Therefore, there is a pressing need for innovative

methods that reduce data requirements while maintaining predictive accu-
racy. Moreover, existing data-driven approaches do not leverage the physical
and geometrical information of a composite’s microstructural representation

6
 Journal Pre-proof

in the training process. It is well established that the location of the fibers

of
is a crucial aspect that governs the distribution of mechanical fields in com-
posites [51, 52, 31]. In traditional pixel-based models, treating composite

pro
microstructures as images or binary representations leads to the loss of key
information regarding the impact of each fiber on its neighboring. These
limitations must be addressed to move closer to the potential replacement of
finite element analysis with machine learning methods.
In this work, we introduce CompINet (Composite Integrated Network), a
re-
novel framework that integrates geometric information from the microstruc-
tural representation of composites into the learning process to address some
of the aforementioned shortcomings. CompINet combines a CNN to pro-
lP
cess the binary representation of the composite with a GNN to capture the
interactions between fibers. By leveraging the inherent microstructural char-
acteristics of fiber-reinforced composites, CompINet reduces the computa-
tional cost of data generation and optimizes the training process, resulting
rna

in a more efficient ML model with enhanced predictive accuracy. In this

study, we utilize CompINet to predict the distribution of von Mises stress
distribution in 2D microstructural representations of fiber-reinforced com-
posites under uniaxial loading. To assess the performance of CompINet, we
Jou

conducted a comparative analysis with a baseline model based on the U-Net

architecture [41]. U-Net was chosen as the baseline model because it provided
one of the best performances in previous studies [41, 48].
This paper is organized as follows: In the first section, we describe the

7
 Journal Pre-proof

proposed integrated deep learning framework. The second section details

of
the generation and pre-processing of data, including a detailed description
of the microstructural representation of the composite, the finite element

pro
framework, and the constitutive equations used to model the responses of
different material constituents. The third section is devoted to the training
process. Finally, in the fourth section, we evaluate the performance of the
proposed deep learning framework, CompINet, in predicting the distribution
of stress fields.
re-
2. Integrated deep network structure

CompINet integrates two different classes of neural network architectures:

lP
CNNs and GNNs. This section first introduces the key concepts behind
CNNs and GNNs, followed by a description of the CompINet algorithm in
Section 2.2.
rna

2.1. Key concepts

Here, we provide a brief overview of the fundamental concepts of CNNs

and GNNs; the key components of the techniques used in our work. It must
be noted that a comprehensive introduction to NNs is beyond the scope of
Jou

this paper, as the field is extensively covered in several textbooks [53, 54, 55].
CNNs are a type of machine learning model designed specifically for pro-
cessing grid-like data, such as images [55]. CNNs use convolutional layers to
learn spatial hierarchies of features from input images. A convolutional layer

8
 Journal Pre-proof

works by applying a set of filters (or kernels) to the input data, which slides

of
over the data to extract local patterns [55]. Each filter is initialized with
random weights, which are continuously updated during the training process

pro
to minimize prediction errors. The convolution operation involves comput-
ing the dot product between the filter and overlapping regions of the input.
Convolutional layers are generally followed by activation functions. The role
of the activation functions is to introduce non-linearity into the model [56].
One of the most commonly used activation functions is the Rectified Linear
re-
Unit (ReLU), which is defined as f (x) = max(x, 0) [55, 57]. The ReLU acti-
vation function sets all negative values to zero while keeping positive values
unchanged. To stabilize the learning process, batch normalization can be ap-
lP
plied after the activation function [58]. Batch normalization normalizes the
output of a layer by scaling it so that it remains within a stable range which
enables the model to train faster and often results in better performance [58].
Additionally, a pooling layer can be used to reduce the spatial dimensions of
rna

the feature maps, typically by applying operations like average pooling [59]
or max pooling [60]. Average pooling divides the feature map into smaller
regions and computes the average value for each region, while max pooling
selects the maximum value from each region. Pooling layers help decrease
Jou

the computational load, control overfitting, and retain essential information

from the features [61]. For more detail on CNNs, the readers are referred
to [54, 55]. The review of research works that used CNNs in the field of solid
mechanics was provided in the Introduction.

9
 Journal Pre-proof

One of the widely used CNN architectures is U-Net [62], which was primarily

of
designed for biomedical image segmentation but has since been adapted for
various image-to-image tasks. U-Net features a symmetric encoder-decoder

pro
structure. The encoder compresses the input into high-level features through
downsampling layers, while the decoder recovers spatial resolution through
upsampling. A key feature of U-Net is the use of skip connections, which
transfer information from corresponding encoder layers directly to the de-
coder, helping retain spatial details lost during downsampling. Another
re-
prominent CNN architecture is ResNet combined with Squeeze-and-Excitation
(SE) blocks [63]. The key feature of ResNet is its shortcut connections, which
allow the output of earlier layers to bypass one or more layers and be added
lP
directly to the output of later layers. This enables the network to learn resid-
ual functions rather than direct mappings, facilitating more efficient learning,
particularly in deep networks. The SE block enhances this architecture by
recalibrating channel-wise feature responses. ResNet-SE architecture won
rna

first place in the ILSVRC (i.e., ImageNet Large Scale Visual Recognition
Challenge) 2017 competition [63], and since then has been used in many
applications including the prediction of stress fields in composites [34, 41].
Graph neural networks, in contrast, are designed to handle data struc-
Jou

tured as graphs. In GNNs, nodes represent entities, and edges define the
relationships between these entities [53]. GNNs can effectively learn and
propagate information based on the connections within the graph. Typi-
cally, GNNs consist of three key components: the encoder, which maps the

10
 Journal Pre-proof

input node features and graph structure onto a latent space; the message-

of
passing module, where nodes aggregate information with their neighbors to
capture complex interactions; and the decoder, which interprets the updated

pro
node representations to produce the final output, such as predictions or clas-
sifications [53]. For more detail on GNN, the readers are referred to [53].
The review of research works that used GNN in the field of solid mechanics
was provided in the Introduction.

2.2. CompINet architecture re-

CompINet combines CNNs and GNNs to build a comprehensive represen-
tation of the relationship between the microstructural geometry of composites
and the distribution of mechanical fields within the microstructure. The CNN
lP
encodes the pixel-based binary image of the composite’s microstructural rep-
resentation, assigning a value of one to the matrix and zero to the fibers, as
shown in Figure 1(A), while the GNN encodes the network of fibers (i.e., the
rna

locations of fibers and the distances between them) as a graph input, with
nodes representing the fibers and edges representing the distances between
them, as shown in Figure 1(B). The incorporation of GNNs for capturing mi-
crostructural geometric features was inspired by the similarity between the
Jou

representation of fibers and the concept of graph networks. In fiber-reinforced

composites, the locations of fibers and the distances between them–often
referred to as nearest neighbor distances–are critical factors that influence
stress distributions within composite microstructures [51, 52, 31]. There-

11
 Journal Pre-proof

A Binary representation Pixel-based representation

of
Composite microstructure
with 46 fibers

pro
Coordinates of 𝑥1 𝑦
𝑥2 𝑦2
the center of
B Graph network the 46 fibers
𝑥3 𝑦3
… …

Connectivity matrix (46x46)

re- 1
2
3
4
5
1
0
𝑑21
2

𝑑31 𝑑32
𝑑41 𝑑42 𝑑43
3 4
𝑑12 𝑑13 𝑑14 𝑑15
0
5

𝑑23 𝑑24 𝑑25

0

𝑑51 𝑑52 𝑑53 𝑑54

𝑑34 𝑑35
0 𝑑45
0
…
…
…
…
…
… … … … … …
lP
Figure 1: The composite microstructure is represented as (A) a pixel-based binary image,
and (B) a graph network. In the pixel-based representation, a 256×256 grid was used,
where pixels are assigned a value of “1” for the matrix and “0” for the fibers. In the graph,
each node corresponds to a fiber, and the edges encode the nearest neighbor distances,
dij , as defined by the connectivity matrix.
rna

fore, in addition to the pixel-based binary representation of the microstruc-

ture used in previously developed CNN-based frameworks (e.g., [38, 47, 34]),
this work also integrates a graph-based representation of the composite mi-
crostructure. The GNN is well-suited for learning the interactions and rela-
Jou

tionships between individual fibers [53], while the CNN excels at extracting
pixel-wise features [53].
Figure 2 illustrates the details of the CompINet algorithm that employs
a CNN encoder, a GNN, and a CNN decoder. A ResNet-SE [63] architec-

12
 Journal Pre-proof

of
Microstructures 3 Conv blocks 5 ResNet blocks

Binary representation
Input: 256 × 256 × 1 Linear von Mises

128 × 32 × 32
128 × 128 × 32

32 × 32 × 128

32 × 32 × 128
32 × 32 × 128

32 × 32 × 128
32 × 32 × 128
64 × 64 × 64
stress fields

Conv block ResNet block 3 DeConv blocks

pro
Skip

Output: 256 × 256 × 1

Skip

128 × 128 × 32
128 × 32 × 32

64 × 64 × 64
Multiply
Graph representation
SE block
SE block Nonlinear von Mises
𝐂
Convolutional + ReLU + Batch Normalization Average Pooling Dense stress fields

DeConv block
Transpose Convolutional
𝑛2 𝑛3 𝑛2 𝑛3 Batch Normalization
𝑛2 𝑛3
𝔼 𝕄ℙ 𝔻

128 × 32 × 1
𝐃𝐅 𝑛4
𝑛4 𝑛4
𝑛1 𝑛1
𝑛1 𝑛7 𝑛7
𝑛7 𝑛5 𝑛5
𝑛5
𝑛6 𝑛6
𝑛6

Encoded nodes Message passing Decoded nodes

Figure 2: Description of CompINet architecture. CompINet takes as input the binary

representation of the composite microstructure (a 256×256 pixel image, where pixels on
the matrix are assigned a value of “1” and pixels on the fibers are assigned a value of “0”)
re-
and its corresponding graph representation (comprising two matrices of node features,
C, and edge features, DF). The first matrix C stores the list of fibers along with the
coordinates of their centers (xi ,yi ), while the second matrix DF is a connectivity matrix
lP
that stores the distances between the centers of the nearest neighbor fibers dij . The binary
microstructure representation is input into the CNN encoder section of the framework,
featuring 3 Conv blocks and 5 ResNet blocks. Simultaneously, the graph representation
undergoes encoding into a graph structure using the module E. These encoded features
are then mapped into a larger latent space and analyzed by a message-passing module,
denoted as MP. Subsequently, the transformed features are processed by the decoder D.
The outputs from both the CNN and GNN are multiplied and further processed by a CNN
rna

decoder consisting of 3 DeConv blocks.

ture is used as the CNN encoder to process the binary representation of the
composite microstructures in the form of an image with 256× 256 pixels. As
illustrated in Figure 2, the CNN encoder comprises 3 Conv blocks followed
Jou

by 5 ResNet blocks. A similar CNN architecture has been used in previ-

ous studies [40, 34, 41]. Each Conv block contains one convolutional layer
followed by ReLU operation, batch normalization, and SE block.
The input to the GNN includes node and edge features matrices (i.e., C

13
 Journal Pre-proof

and DF). In this work, the node feature matrix C contains the coordinates

of
of the center of the N fibers. The edge feature matrix, DF, holds information
about the graph structure (i.e., connectivity matrix) and stores the distances

pro
between the nearest neighbor fibers, as illustrated in Figure 2. The node
and edge features are encoded using two distinct neural networks (i.e., E)
that contain two linear layers, each followed by a ReLU activation function,
and a normalization layer. The encoded node and edge features are then
updated using message-passing module (i.e., MP) which employs two Multi-
re-
Layer Perceptrons (MLPs). The edge MLP processes concatenated node
features, neighbor features, and edge attributes to compute messages for
edge updates. Meanwhile, the node MLP takes aggregated messages and
lP
original node features, passing them through a series of linear layers, ReLU
activation function, and normalization layer to update the node features.
Then, the resulting node features are processed by the decoder (i.e., D).
The decoder, D, is implemented using MLP with linear and ReLU activation
rna

layers. The architecture of the GNN used in this work was derived from a
study conducted by Maurizi et al. [49].
The output of the CNN encoder, portrayed as a 128×32×32 tensor, is
then element-wise multiplied with the GNN output, a tensor with dimensions
Jou

128×32×1. Subsequently, the extracted enhanced features, represented as

a 128 ×32×32 tensor, are processed by a CNN decoder to construct the
desired output dimensions, a 256×256×1 stress field. The CNN decoder,
illustrated in Figure 2, comprises three DeConv blocks. Each block includes

14
 Journal Pre-proof

a deconvolutional layer with batch normalization and ReLU operation.

of
3. Data generation

pro
The data used to train CompINet consists of sets of microstructural repre-
sentations of a fiber-reinforced composite and their corresponding mechanical
field distributions. In this study, we selected carbon fiber-reinforced polymer
composites as the material of interest and von Mises stress distribution as

re-
the mechanical field of focus.
Carbon fiber-reinforced composite was chosen as the literature reports
the distribution of fibers’ nearest neighbor distances (NND) for this class of
composites [31]. Moreover, this composite is characterized by a high volume
lP
fraction of randomly distributed fibers (i.e., typically greater than 55% of
the composite’s volume) within an epoxy matrix, making it more similar to
the structures of soft tissues and other nature-based composites. The high
rna

fiber volume fraction results in adjacent fibers being in close proximity, and
sometimes even touching one another. The matrix cracking portion of a
transverse crack typically occurs in the small spaces between the adjacent
fibers, signifying the importance of the distribution of NNDs [31, 52]. There-
fore, an accurate microstructural representation of composite is of utmost
Jou

importance as the stress distribution and concentration greatly depend on

the microstructural representation and fibers’ locations [52], and the fibers’
dispersion can significantly impact the robustness of the numerical simula-
tions [52, 31]. It is also worth noting that the high fiber volume fraction

15
 Journal Pre-proof

and the close proximity of fibers make FEA meshing of such composites a

of
challenging task, which significantly increases the computational cost of their
analysis. These numerical difficulties make them an attractive candidate for

pro
substitution with ML-based surrogate models.
Moreover, we particularly selected von Mises stress for this work because
it effectively combines normal and shear stresses, providing a single, repre-
sentative measure for material yielding and failure. Other stress measures
could be used to train ML models. For instance, the elastic stress compo-
re-
nents under multiaxial loading can be predicted by training the ML model
on separate stress components corresponding to uniaxial tensile strain and
in-plane shear strain. By applying the principle of superposition, these stress
lP
fields can then be combined to construct the local stress distribution under
multiaxial loading [47, 64]. In this study, we selected one representative stress
measure, the von Mises stress field, as the focus will be on the evaluation of
the performance of the proposed approach.
rna

In the following subsections, first, we explain the random generator algo-

rithm used to construct distinct microstructural representations of the fiber-
reinforced composite. Then, we present the FEA framework used to conduct
numerical simulations to obtain the stress distribution, along with the em-
Jou

ployed constitutive equations and material properties. Finally, we describe

how the data is pre-processed for use within CompINet.

16
 Journal Pre-proof

3.1. Data set description

of
A random fiber generator developed in our previous work [46] was uti-
lized to construct the 2D microstructural representations of the carbon fiber-

pro
reinforced composite. The algorithm generates random fiber locations to
match a target fiber volume fraction and a distribution of NND. The effi-
ciency of the algorithm in accurately capturing the variability of microstruc-
tures is demonstrated in [46].

re-
Two different fiber volume fractions were considered: 60% and 47%. A
typical fiber volume fraction for carbon fiber-reinforced composites is around
60% [65, 31, 66, 67]. Additionally, a second volume fraction of 47% was
chosen, motivated by advancements in additive manufacturing techniques.
lP
These techniques allow for continuous fiber volume fractions between 40%
and 50% [68, 69]. A total of 350 microstructural representations of carbon
fiber-reinforced composite with dimensions 54µm×54µm were generated, for
rna

each volume fraction. The diameter of the circular fibers was equal to 7µm.
These microstructural representations were then simulated using the finite
element framework described below. It is important to note that the param-
eters of the constitutive equations were kept constant across all simulations,
with the only variable being the location of the fibers.
Jou

3.2. High-fidelity numerical framework

An efficient and robust numerical framework along with accurate consti-

tutive equations for different constituents were used to run the high-fidelity

17
 Journal Pre-proof

simulation and generate data [31]. A summary of the FEA and the consti-

of
tutive equations implemented within the framework is presented herein.

3.2.1. Finite element model

pro
This work employed a nonlinear cohesive interface-enriched generalized
finite element method (IGFEM) scheme due to its ability to apply noncon-
forming mesh to complex geometries with discontinuous gradient fields, such
as composites. A nonconforming mesh decreases the computational burden

re-
by enabling the application of a uniform mesh to the geometry. Nonlinear
cohesive IGFEM is an extension of IGFEM that allows an element to include
one to two embedded cohesive interfaces modeled by cohesive zone models.
For more details on the formulation, implementation, and verification of non-
lP
linear cohesive IGFEM, the reader can refer to [70, 71, 31].
In the nonlinear cohesive IGFEM framework, three-node triangular plane-
strain elements were used. A mesh sensitivity analysis determined a uniform
rna

mesh size with 20 elements along the diameters of each fiber. A strain of 1.2%
was incrementally applied to the right edge of the microstructure, while the
horizontal displacement of the left edge was restrained. This boundary con-
dition subjected the composite to uniaxial transverse tension while allowing
Jou

for free vertical deformations. A schematic of the composite microstructure

with the applied boundary conditions and loading is shown in Figure 3-A,
while Figure 3-B illustrates a typical macroscopic stress-strain curve of the
composite’s response. In this study, we used CompINet to predict the stress

18
 Journal Pre-proof

A B 70 (I)

of
(II)

Macroscopic stress (MPa)

60
50
40
30
(II)

pro
20
10
(I)
0
0 0.2 0.4 0.6 0.8 1 1.2
Macroscopic strain (%)

Figure 3: (A) Schematic of the boundary conditions and the applied loading (B) the
macroscopic stress-strain curve of a sample microstructure under the prescribed loading.
The selected linear von Mises stress field (I) corresponds to 0.024% strain and the non-

re-
linear von Mises stress field (II) corresponds to the maximum macroscopic stress. The
macroscopic stress was calculated as the sum of the reactions at the nodes of the left edge
of the microstructure divided by the cross-sectional area.

field distribution within the composite microstructures at two loading stages:

first, during the elastic linear stage at the applied strain of 0.024%, and sec-
lP
ond, at the peak macroscopic stress where damage has already initiated in
the matrix, as illustrated in Figure 3-B. It is important to note that the
peak macroscopic stress did not occur at the same strain level across all
rna

microstructures, which is why it was not associated with a specific strain

level.

3.2.2. Constitutive equations

Robust nonlinear constitutive behaviors for fibers, matrix, and fiber/matrix

Jou

interfaces were implemented within the nonlinear cohesive IGFEM frame-

work. The matrix consisted of an elasto-plastic material with a brittle dam-
age model. The Tschoegl yield criterion specifies the initiation of plastic

19
 Journal Pre-proof

deformation [72] as

of
σ , ϵpeq ) = f (σ
ϕ(σ σ ) − σY (ϵpeq ) (1)

pro
where σ is the stress tensor and ϵpeq is the equivalent plastic strain. In Eq. [1],
σ ) and σY (ϵpeq ) are the yield function and yield threshold, respectively,
f (σ
defined as

f (σ re-
σ ) = 6J2 + 2I1 (σyc − σyt ) & σY (ϵpeq = 0) = 2σyc σyt (2)

where I1 and J2 are the first and second invariants of the stress and deviatoric
stress tensors, respectively, σyt and σyc are the matrix strengths under uniaxial
lP
tension and compression, respectively, and ϵpeq = 0 denotes the initiation of
plastic deformation. The evolution of plastic deformation was modeled by
an isotropic hardening law, defined using a non-associated flow rule, based
rna

on the von Mises criterion. To ensure a smooth elastic-to-plastic transition,

a behavior proposed by Ramberg and Osgood [73] was considered as

 b
σY (ϵpeq = 0)
H=a (3)
σ)
f (σ
Jou

where H is the slope of the tangent line to the plastic branch, a and b are two
parameters that control the shape and smoothness of the elastic-to-plastic
transition. The initiation of failure was determined by the Tschoegl yield

20
 Journal Pre-proof

of
Table 1: Material properties for different constituents of the carbon fiber-reinforced com-
posite. E, G, and ν are Young’s modulus, the shear modulus, and the Poisson’s ratio,
respectively. 1 and 2 represent the out-of-plane and in-plane directions, respectively, and
3 is the direction perpendicular to the 1 and 2 directions.
Matrix

pro
Elastic Properties Plastic Properties Damage Properties
E σc σt
ν a b ϵc ϵt A B µ
(GPa) (MPa) (MPa)
3.9 0.39 79 62 20000 12 0.35 0.04 0.95 2 10

Carbon Fiber Fiber & Matrix

E1 E2 G12 G23 Tc δc Gc
ν12
(GPa)
233
(GPa)
23.1
(GPa) (GPa)
8.96
re-
8.27 0.2
(MPa) (nm)
70 1
(N/m)
8.75

criterion [72] formulated based on the strain tensor, ϵ , as

lP
ϕ′ (ϵϵ) = 6J2′ + 2I1′ (ϵc − ϵt ) − 2ϵc ϵt (4)

where I1′ and J2′ are the first and second invariants of the strain and deviatoric
strain tensors, respectively, and ϵt and ϵc are the failure strains under uniaxial
rna

tension and compression, respectively. After the initiation of failure, the

evolution of damage was simulated based on the Simo and Ju continuum
damage constitutive equation [74, 75], where the damaged stiffness matrix,
Dd , was calculated by the penalization of the elasto-plastic stiffness matrix,
Jou

Dep , via a damage index, d, as Dd = (1 − d)Dep . The damage index d was

calculated and updated via

 
dt
d=d+ (G − Y ) (5)
1 + µdt

21
 Journal Pre-proof

where µ is a viscosity parameter, dt is the pseudo-time, G is the damage

of
parameter, and Y is the damage threshold. G and Y were defined as

1−A Y + µdtG

pro
G = 1 − τ¯0 − Ae[B(τ¯0 −τ̄ )] & Y = (6)
τ̄ 1 + µdt

√
where A and B are two damage parameter constants, τ̄ = 2Ξ is a damage
parameter calculated based on the strain energy Ξ, and τ¯0 is a constant called
the initial damage threshold which is equal to τ̄ at the start of damage. The
re-
damage index d was updated only when G > 0, Y > 0, and G − Y > 0.
The fiber/matrix interfacial debonding was simulated using a bilinear
CZM constitutive model proposed by Ortiz and Pandolfi [76], where the
lP
cohesive behavior was defined based on an effective opening displacement,
p
δ = δn2 + δt2 , taking into account the interaction between normal and tan-
gential debonding displacements (i.e., δn and δt , respectively). The param-
eters of the cohesive law are fracture toughness, Gc , cohesive strength, Tc ,
rna

and critical opening displacement, δc . CZMs can cause convergence issue in

static analyses when modeling brittle debonding failures. The convergence
issue is caused by the Newton-Raphson iterations entering an infinite cy-
cle [77]. Therefore, an artificial viscosity proposed by [78] was added to the
Jou

cohesive law to overcome the difficulty. The parameters of the aforemen-

tioned constitutive equations for different components of a typical carbon
fiber-reinforced composite are presented in Table 1 which were obtained by
fitting to experimental results [31, 79, 4]. The numerical framework, consti-

22
 Journal Pre-proof

tutive equations, and material properties summarized here were previously

of
implemented and verified by the authors [31].

3.3. Data pre-processing

pro
The stress fields provided at nodal points of the finite element mesh were
transformed into a uniform grid. This transformation consisted of convert-
ing the stress fields into regularly spaced grids through interpolation among
neighboring finite element nodal points. To perform this conversion, Par-
re-
aView [80], an open-source data visualization and analysis software, was
used. Therefore, the processed stress contours consisted of 256×256 pixels
containing stress values. Similarly, the microstructure geometry was con-
verted into a 256×256 pixels binary representation. The pixels that lie on
lP
the matrix were assigned a value of “1” and the pixels that lie on the fibers
were assigned a value of “0”. The input for the GNN component of CompINet
was generated by storing the coordinates of the fiber centers. These coordi-
rna

nates were then used to calculate the distances between the fiber centers and
generate the connectivity matrix.

4. Training process
Jou

The training data set used in this study consists of the microstructural
representations of fiber-reinforced composite and the corresponding linear
and nonlinear von Mises stress fields. The random generator algorithm was
used to produce 350 microstructural representations of the fiber-reinforced

23
 Journal Pre-proof

composites for each 60% and 47% fiber volume fraction. The presented finite

of
element framework, constitutive equations, and material properties were used
to run 350 simulations. Out of the 350 simulation results, we used 100

pro
samples as a test set. The remaining 250 samples were flipped vertically,
resulting in a total of 500 samples available for training and validation.
The 500 training samples were used to create subsets of various sizes (i.e.,
50, 100, 250, and 500 samples) for training and validation purposes. Each
data set subset contained both original microstructures and their vertically
re-
flipped versions. For instance, in the 100-sample subset, 50 were original sam-
ples, and 50 were their flipped counterparts. While this data augmentation
increased the data set size, it could introduce correlations between the origi-
lP
nal and flipped images. To reduce any potential bias from these correlations,
we ensured that both original and flipped samples were always included in
each training subset. Each subset was split into 80% for training and 20% for
validation. The 80/20 split, also referred to as the 80-20 rule, is widely used
rna

because it generally strikes a good balance between having sufficient data

for model training and enough data for testing and aligns with the Pareto
principle [81, 82]. Additionally, for each subset, we repeated the training
process 10 times using randomly selected training and validation samples to
Jou

account for variability in the framework’s performance based on the chosen

samples. We used Mean Absolute Error (MAE) as the loss function which

24
 Journal Pre-proof

was calculated as

of
N̄ n
1 X1X
LM AE = |σij − σ̄ij | (7)
N̄ j=1 n i=1

where N̄ is the number of training samples, n is the number of pixels in each

pro
sample (i.e., 256 × 256), σij and σ̄ij are the target (i.e., FEA results) and the
CompINet predicted stress at ith pixel in the j th sample. Please note that
the indices in σij herein contain different meanings from the conventional
components of the stress tensor in solid mechanics.
re-
The training was conducted using Google Colaboratory. Fine-tuning of
the hyperparameters was performed to choose the best optimizer and select
the optimal hyperparameters. We tried two different optimizers, ADAM and
Stochastic Gradient Descent (SGD). We also gradually varied the learning
lP
rate from 0.1 to 0.0001 and examined its impact on the predictive perfor-
mance of the model. Additionally, we varied the learning rate decay coef-
ficient between 0.5 and 1. This tuning process was carried out separately
rna

for CompINet’s training on linear and nonlinear stress fields. Ultimately, we

selected an ADAM optimizer with a learning rate of 0.04 for the linear stress
field prediction and an SGD optimizer with a learning rate of 0.1, for the
nonlinear stress field prediction which has been observed to yield stable con-
Jou

vergence of the loss function. The learning rate decay coefficient was equal
to 0.9. Figure S1 shows the evolution of the loss function during the training
process obtained during the training of CompINet and the baseline model
(i.e., U-Net-based architecture). The training was stopped after 100 epochs

25
 Journal Pre-proof

ensuring the convergence of the loss function as illustrated in Figure S1 in

of
the supplementary material (SM). Figure S1 shows that the training and
validation losses of CompINet converged in less than 40 epochs without any

pro
significant gap between them. CompINet attains a notably lower training
and validation loss compared to the baseline and exhibits a more consistent
and smooth convergence, indicating stable and robust learning.

5. Results and discussion

re-
In this section, we investigate the capabilities of CompINet in predict-
ing the linear and nonlinear von Mises stress distribution within the mi-
crostructural representation of fiber-reinforced composites. As mentioned
lP
earlier, to assess CompINet’s performance, we compared it to a U-Net-based
model, a CNN model that has demonstrated superior accuracy in previous
research [37, 38, 39, 41, 48]. Studies have shown that U-Net outperforms
rna

models like Residual Networks [41] and achieves accuracy comparable to

cGANs [48]. Therefore, the comparison with U-Net provides a robust evalu-
ation of the performance of CompINet. The architecture of the baseline U-
Net-based model is detailed in [41]. To assess the performance of CompINet,
we utilized different evaluation metrics. First, we examined the evolution of
Jou

the root mean squared error (ERM SE ) for different training and validation

26
 Journal Pre-proof

data set sizes. ERM SE was calculated as

of
v
u N n
u1 X 1X
ERM SE =t (σij − σ̄ij )2 (8)
N j=1 n i=1

pro
where N is the number of test samples (i.e., 100 samples). Moreover, corre-
lation plots and stress error maps were utilized to examine stress prediction
patterns in a randomly selected test sample, highlighting trends like stress
overestimation or underestimation, as well as areas with concentrated er-
re-
rors. To comprehensively assess the model’s performance, we calculated R2
for each test sample then plotted the distribution of R2 values across 100
test samples. Investigating the distribution of R2 values across the test set
lP
provides insights into both the predictive accuracy and the consistency of
the model’s predictions throughout the entire test set, which is a crucial
performance indicator.
rna

In the following subsections, the performance of CompINet in predicting

the linear and nonlinear von Mises stress distributions, as shown in Figure 3,
for the composite with a 60% fiber volume fraction is presented. Additional
information about the composite with a 47% fiber volume fraction is briefly
discussed in Section 5.3 and can be found in detail in SM Figure S9.
Jou

5.1. Linear response for 60% fiber volume fraction composite

This section assesses the efficacy of CompINet in predicting the linear

stress field. Figure 4 provides a comparative analysis of the prediction out-

27
 Journal Pre-proof

CompINet Baseline

of
A Predictions B Error maps C Correlation plots D Error maps E Correlation plots

50
samples

pro
100
samples

250
samples

F
re- G
lP
rna

Figure 4: Comparison of CompINet and the baseline performance in predicting the lin-
ear von Mises stress field. For a randomly selected test sample, the figure presents (A)
CompINet linear stress field predictions, (B) CompINet error maps (i.e., the absolute dif-
ference between the CompINet predicted stress and the FEA stress at each pixel), (C)
CompINet versus FEA stress correlation plots, (D) the baseline stress error maps (i.e.,
the absolute difference between the stress predicted by the baseline model at each pixel
and the FEA stress), and (E) the baseline versus FEA stress correlation plots for 50, 100,
and 250 training and validation samples. Stresses are given in MPa. (F) Comparison
Jou

of R2 distributions calculated across 100 test samples using CompINet and the baseline.
Different training and validation set sizes are used (i.e., 50, 100, 250 samples). The opacity
of the histograms increases as the data set size grows. The vertical dashed lines indicate
the mean R2 for each distribution. (G) Comparison of ERM SE calculated for CompINet
and the baseline for different training and validation set sizes. For the baseline, the size
of the training and validation data set varies between 50 and 2000 samples [41], while for
CompINet, the size of the training and validation set varies between 50 and 500 samples.
Training is repeated 10 times for each case.

28
 Journal Pre-proof

comes achieved using CompINet and the baseline model.

of
First, a random test sample was chosen. Figure 4 column A to C presents
CompINet linear stress prediction contour, the stress error maps (i.e., the

pro
absolute difference between the predicted stress and the FEA stress at each
pixel), and the correlation plots (i.e., CompINet predicted stress at each pixel
versus its corresponding FEA value) for the selected test sample obtained
based on surrogate models with 50, 100, and 250 training and validation
sample sizes. Figure 4 columns D and E depict the linear stress error maps
re-
and the correlation plots associated with the baseline model predictions ver-
sus FEA for each training and validation subset. Comparing the stress error
maps in columns B and D clearly shows CompINet’s superior performance
lP
over the baseline model. A close investigation of the error maps indicates
that there are higher levels of error in the baseline predictions in between
fibers. CompINet shows an improved accuracy at these high-stress concen-
tration areas. This notable improvement highlights the GNN’s crucial role
rna

in learning the interactions between fibers and the intricacies of the system.
Moreover, the correlation plots depicted in columns C and E and associated
R2 values reported for each case provide compelling evidence of CompINet’s
improved performance versus the baseline. The R2 values exhibit an increase
Jou

from 0.881, 0.905, and 0.927 for the baseline to 0.942, 0.954, and 0.969 for
CompINet predictions. Figure 4-E shows that the baseline model tends to
underestimate higher stresses in the fibers and overestimate lower stresses
in the matrix. CompINet effectively addresses this issue, offering a more

29
 Journal Pre-proof

accurate prediction of stress distribution across various regions emphasizing

of
again the role of GNN in capturing the complexity of the fibers network.
Additional stress error maps and correlation plots for the worst and best-

pro
performing test samples for linear stress prediction are depicted in Figures
S3-S4 in the supplementary material.
Second, to delve deeper into the comparison, Figure 4-F presents the dis-
tribution of R2 values across all the 100 test samples for CompINet and the
baseline, which were trained using 50, 100, and 250 samples. The vertical
re-
lines indicate the mean R2 value for each distribution. The shades of blue
present the baseline performances whereas the shades of green illustrate the
CompINet cases. Figure 4-F clearly indicates that CompINet consistently
lP
achieves higher mean R2 values compared to the baseline. In addition to
the improved mean R2 , CompINet exhibits narrower distributions of R2 val-
ues with standard deviations of 0.007, 0.006, and 0.005 for 50, 100, and 250
training and validation samples, respectively. In comparison, the baseline
rna

standard deviations are 0.01, 0.008, and 0.006 for the same training and val-
idation set sizes. A less scattered R2 distribution highlights the consistency
and reliability of CompINet compared to the baseline. These findings under-
score the robustness of CompINet for accurate and reliable linear stress field
Jou

predictions.
Finally, Figure 4-G provides a comprehensive comparison by tracking
the evolution of ERM SE for predictions made by CompINet and the baseline
model across 100 test samples. The training and validation set sizes varied

30
 Journal Pre-proof

between 50 and 500 samples for CompINet, while the baseline model, as

of
detailed in [41], was trained using up to 2000 samples. The error bars are
associated with the 10 repetitions of the training process, as explained in

pro
Section 4. It can be seen that for both models, ERM SE decreases with a
diminishing improvement rate when we increase the size of the training and
validation data set. Figure 4-G clearly illustrates that CompINet exhibits
much lower ERM SE compared to the baseline. For instance, a significant
31.6% improvement is observed in the mean value of ERM SE for a training
re-
and validation data size of 250 samples. Moreover, CompINet, trained with
only 250 samples, outperforms the baseline model, which was trained with a
larger set of 2000 samples [41]. Additionally, the cumulative frequency of the
lP
absolute errors between the target and the predicted linear stresses, provided
by CompINet and the baseline, using different training and validation data
set sizes is provided in SM Figure S7.
In summary, CompINet’s remarkable performance is highlighted by achiev-
rna

ing superior accuracy with a training and validation data set 20 times smaller
than that of the baseline. CompINet, trained with only 250 samples, reached
a mean R2 value of 0.96 and an ERM SE of 0.30 MPa, emphasizing its excep-
tional predictive capabilities and showcasing both efficiency in data utiliza-
Jou

tion and effective generalization. This achievement is particularly noteworthy

considering the computational cost associated with data generation which re-
lies mainly on FEA. This considerable reduction in the size of the training
data marks a major step forward in advancing the application of deep learn-

31
 Journal Pre-proof

ing within the field of mechanics.

of
5.2. Nonlinear response for 60% fiber volume fraction composite

This section investigates the accuracy of CompINet in predicting the non-

pro
linear stress field. Figure 5 compares the performance of CompINet and the
baseline in predicting the nonlinear stress distribution. It should be reiter-
ated that the nonlinear stress fields for training, validation, and testing were
chosen at the peak macroscopic stress response as presented in Figure 3.
re-
Similar to the linear stress prediction, first, a random test sample was
chosen. Figure 5 columns A to C presents the CompINet nonlinear stress
prediction contour, the stress error maps, and the correlation plots for the
selected one test sample obtained based on surrogate models with 50, 100,
lP
and 250 training and validation sample sizes. Figure 4 columns D and E de-
pict the nonlinear stress error maps and the correlation plots associated with
the baseline model predictions versus FEA for each training and validation
rna

sample size.
Since the matrix was considered a material with hardening plasticity and
brittle damage behavior, and the interfacial debonding around the inclusions
was taken into account in the FEA simulations, the nonlinear stress fields
Jou

display distinct areas of unloading and local stress drop at the peak macro-
scopic stress. These patterns serve as important indicators, illustrating the
occurrence of damage and therefore its influence on stress distribution before
a total failure. The stress error maps in Figure 5-D demonstrate concentrated

32
 Journal Pre-proof

errors in areas where such unloading exists, and indicate that the baseline

of
model exhibits limitations in capturing these specific behaviors. A direct
comparison of the error concentration regions in Figure 5-D with the dam-

pro
age contour (SM Figure S10) reveals a notable correspondence. The region
of stress error concentration aligns with the area where damage initiation oc-
curs. In contrast, CompINet stands out in more effectively capturing these
nuanced features of local stress drop, as illustrated in Figure 5-B. The more
accurate performance of CompINet can be attributed to the incorporation of
re-
spatial information regarding the nearest fibers’ neighbors that govern stress
concentration and consequently the regions of damage initiation.
Moreover, the correlation plots presented in Figures 5 columns C and E
lP
demonstrate that CompINet exhibits improved performance compared to the
baseline, with enhanced R2 values. The enhancement in stress prediction is
particularly noteworthy within the fibers. Figure 5-E shows that the base-
line model underestimates the higher stresses while overestimating the lower
rna

stress values in fibers. CompINet, on the other hand, addresses and corrects
this discrepancy, showcasing its ability to provide more accurate stress pre-
dictions. Additional stress error maps and correlation plots for the worst and
best-performing test samples for nonlinear stress prediction are presented in
Jou

the SM, Figures S5 and S6.

For a more comprehensive comparison, Figure 5-F illustrates the distri-
bution of R2 values across 100 test samples for CompINet and the baseline
predictions. Both models were trained using 50, 100, and 250 samples. The

33
 Journal Pre-proof

mean R2 values demonstrate an improvement in performance when compar-

of
ing CompINet with the baseline model. However, the distributions of R2
values are more dispersed compared to the linear stress prediction. This

pro
increased variability suggests that the models’ performance exhibits greater
fluctuations, likely influenced by the inherent variability present in the data.
As argued previously, at the peak loading, damage initiation occurred in
some but not in all samples. This variation in the extent of damage con-
tributes to higher variability in the stress distribution, adding complexity to
re-
the prediction of the nonlinear stress field.
Finally, Figure 5-G illustrates the evolution of ERM SE for predictions gen-
erated by CompINet and the baseline model using training and validation
lP
data set sizes ranging from 50 to 500 samples across all 100 test samples.
Similar to the linear stress prediction, the error bars correspond to 10 rep-
etitions of training. As the size of the training and validation data sets
increases, both frameworks exhibit a decrease in ERM SE , though the rate
rna

of improvement diminishes. Notably, CompINet consistently achieves lower

ERM SE compared to the baseline. Additionally, across the 10 training pro-
cesses conducted for each training and validation set, CompINet consistently
demonstrates lower variability in ERM SE compared to the baseline model,
Jou

highlighting its greater robustness in learning. The cumulative frequency of

the absolute errors between the target and predicted nonlinear stresses, pro-
vided by CompINet and the baseline, using different training and validation
data set sizes, is provided in SM Figure S8.

34
 Journal Pre-proof

of
CompINet Baseline
A Predictions B Error maps C Correlation plots D Error maps E Correlation plots

50

pro
samples

100
samples

250
samples
re-
F G
lP
rna

Figure 5: Comparison of CompINet and the baseline model performance in predicting

the nonlinear von Mises stress field. For a randomly selected test sample, the figure
presents (A) CompINet nonlinear stress field predictions, (B) CompINet error maps (i.e.,
the absolute difference between the CompINet predicted stress and the FEA stress at each
pixel), (C) CompINet versus FEA stress correlation plots, (D) the baseline stress error
maps (i.e., the absolute difference between the stress predicted by the baseline model at
each pixel and the FEA stress), and (E) the baseline versus FEA stress correlation for 50,
Jou

100, and 250 training and validation samples. Stresses are given in MPa. (F) Comparison
of R2 distribution calculated across 100 test samples using CompINet and the baseline.
Different training and validation set sizes are used (i.e., 50, 100, 250 samples). The opacity
of the histograms increases as the size of the training and validation data set grows. The
vertical dashed lines indicate the mean R2 for each distribution. (G) Comparison of
ERM SE calculated for CompINet and the baseline for different training and validation set
sizes ranging from 50 to 500 samples. Training is repeated 10 times for each case.

35
 Journal Pre-proof

5.3. Linear response for 47% fiber volume fraction composite

of
In this section, we briefly discuss the outcome of the prediction of the
linear von Mises stress fields in 47% fiber volume fraction composites us-

pro
ing CompINet. The results associated with this fiber volume fraction are
presented in the SM. The general trend and comparison to the baseline are
all similar to the 60% fiber volume fraction results. However, it must be
mentioned that both CompINet and the baseline demonstrate better perfor-

re-
mance at a 60% fiber volume fraction compared to the 47% fraction, which
can be attributed to the greater variability in fiber locations at lower volume
fractions. At 47%, the randomly generated microstructures exhibit more ir-
regular and heterogeneous fiber distributions and higher clustering, leading
lP
to increased variability in the von Mises stress field. This higher variability
makes accurate prediction more challenging for the ML models.
Nonetheless, it is worth mentioning that while FEA excels in predicting
rna

stress fields for lower volume fractions, it encounters challenges in handling

higher volume fractions, where the computational demands escalate due to
the requirement for a finer mesh. Therefore, this is where a faster alternative,
such as CompINet, proves advantageous for higher fiber volume fractions,
where FEA becomes computationally expensive.
Jou

6. Conclusion

This work introduces CompINet, a deep-learning framework that com-

bines two types of neural networks: CNNs and GNNs to predict the distri-

36
 Journal Pre-proof

bution of mechanical fields within composite microstructures. The integra-

of
tion of a CNN and a GNN, allows CompINet to capture intricate details at
the pixel-level while also identifying overall patterns influenced by the lo-

pro
cations and the interactions between fibers. The efficiency and accuracy of
CompINet’s performance were evaluated for predicting the linear and non-
linear von Mises stress distributions in fiber-reinforced composites with two
different fiber volume fractions (i.e., 60% and 47%). CompINet’s perfor-
mance was assessed by comparing its predictions to both the ground truth
re-
FEA data and the baseline ML framework that employs only convolutional
neural networks. The results show that CompINet achieves higher accuracy
and efficiency compared to the baseline model for all considered cases, in-
lP
cluding the prediction of both linear and nonlinear stress fields and the two
different fiber volume fractions of 60% and 47%. Specifically, CompINet sig-
nificantly reduces the training data requirements by a factor of 20 compared
to the baseline model when predicting the linear stress field. Moreover, in
rna

addition to being more accurate than the baseline, CompINet’s performance

is also more reliable and consistent.
Concerning the prediction of the nonlinear stress field, although the ac-
curacy declines compared to the linear stress field prediction, CompINet’s
Jou

prediction of the nonlinear stress field is still more accurate than available
approaches in the literature. Indeed, predicting the nonlinear stress field
is inherently more complex, as it corresponds to the maximum macroscopic
stress reached by the composite, occurring at different stress and strain levels

37
 Journal Pre-proof

across the samples. The increased variability in the training data set (i.e.,

of
stress distribution) for the nonlinear case is attributed to the initiation and
progression of damage in the matrix. Consequently, the presence of unload-

pro
ing regions in certain samples impacts the generalization capability of the
ML frameworks. Therefore, improving CompINet could involve considering
the level of local damage reached by the matrix.
It is worth mentioning that CompINet can still be improved to broaden
its applicability and enhance its predictive capabilities, and this paragraph
re-
addresses some key areas for improvement. First, additional data augmenta-
tion techniques can further reduce the number of required FEA simulations.
In this study, we only applied vertical flipping to the samples; however, hori-
lP
zontal flipping and 180° rotation could also be utilized [64]. Special attention
should be given to the directions of stress components in this case. Moreover,
the current version of CompINet’s GNN input only considers the locations of
the fibers and the distances between them. CompINet’s performance could
rna

be enhanced by incorporating other physical quantities, such as interfacial

properties and distance vectors, to account for various loading directions. Ad-
ditionally, CompINet can be extended to predict the stress field at different
strain levels, allowing for the construction of the entire stress field evolu-
Jou

tion. Combining CompINet with temporal deep learning approaches, such

as Long Short-Term Memory (LSTM), can facilitate the prediction of stress
distribution throughout the loading history. Further considerations, such as
tracking damage evolution, could also be integrated to improve predictions of

38
 Journal Pre-proof

nonlinear response. Lastly, to effectively implement CompINet for practical

of
applications, we aim to integrate it within a multi-scale framework, which can
significantly reduce the computational costs associated with analyzing large

pro
structures. By employing a hierarchical approach where CompINet replaces
high-fidelity models, we can optimize resource utilization while maintaining
accuracy. The aforementioned cases are the focus of the authors’ current and
future study.

re-
CRediT authorship contribution statement

Marwa Yacouti: Conceptualization, Methodology, Software, Writing

– original draft. Maryam Shakiba: Conceptualization, Writing review &
lP
editing, Supervision, Funding acquisition.

Declaration of competing interest

rna

The authors declare that they have no known competing financial inter-
ests or personal relationships that could have appeared to influence the work
reported in this paper.

Acknowledgments
Jou

The authors gratefully acknowledge the support from the Air Force Office
of Scientific Research (AFOSR) Young Investigator Program (YIP) award
#FA9550-20-1-0281. This work utilized the Blanca condo computing re-

39
 Journal Pre-proof

source at the University of Colorado Boulder. Blanca is jointly funded by

of
computing users and the University of Colorado Boulder.

pro
References

[1] M. Zhou, B. Werbner, G. D. O’Connell, Fiber engagement accounts

for geometry-dependent annulus fibrosus mechanics: A multiscale,
structure-based finite element study, Journal of the Mechanical Behavior

re-
of Biomedical Materials 115 (2021) 104292.

[2] M. Shakiba, Detecting transverse cracks initiation in composite lam-

inates via statistical analysis of sensitivity data, Mechanics Research
Communications 115 (2021) 103701.
lP
[3] J. N. Reddy, Mechanics of laminated composite plates and shells: theory
and analysis, second edition, 2nd Edition, CRC Press, Boca Raton, 2003.
rna

[4] M. W. Hyer, S. R. White, Stress Analysis of Fiber-reinforced Composite

Materials, DEStech Publications, Inc, 2009.

[5] Y. Wang, Z. Huang, Analytical micromechanics models for elastoplastic

behavior of long fibrous composites: A critical review and comparative
Jou

study, Materials 11 (10) (2018) 1919, number: 10 Publisher: Multidis-

ciplinary Digital Publishing Institute.

[6] E. Rashidinejad, H. Ahmadi, M. Hajikazemi, W. V. Paepegem, Closed-

form analytical solutions for predicting stress transfers and thermo-

40
 Journal Pre-proof

elastic properties of short fiber composites, Mechanics of Advanced Ma-

of
terials and StructuresPublisher: Taylor & Francis (Dec. 2023).

[7] O. O. Ochoa, J. N. Reddy, Finite element analysis of composite lami-

pro
nates, Vol. 7 of Solid Mechanics and Its Applications, Springer Nether-
lands, Dordrecht, 1992.

[8] M. Dhuria, N. Grover, K. Goyal, Review of solution methodologies

for structural analysis of composites, European Journal of Mechanics
re-
- A/Solids 103 (2024) 105157.

[9] S. Oller, Numerical simulation of mechanical behavior of composite ma-

terials, Lecture Notes on Numerical Methods in Engineering and Sci-
lP
ences, Springer International Publishing, Cham, 2014.

[10] P. Bose, J. Reddy, Analysis of composite plates using various plate theo-
ries -Part 1: Formulation and analytical solutions, Structural Engineer-
rna

ing and Mechanics 6 (6) (1998) 583–612.

[11] P. Bose, J. Reddy, Analysis of composite plates using various plate the-
ories -Part 2: Finite element model and numerical results, Structural
Engineering and Mechanics 6 (7) (1998) 727–746.
Jou

[12] A. Odeh, M. A. Al-Shugaa, H. J. Al-Gahtani, F. Mukhtar, Analysis

of laminated composite plates: A comprehensive bibliometric review,
Buildings 14 (6) (2024) 1574.

41
 Journal Pre-proof

[13] A. Alaimo, C. Orlando, S. Valvano, Analytical frequency response solu-

of
tion for composite plates embedding viscoelastic layers, Aerospace Sci-
ence and Technology 92 (2019) 429–445.

pro
[14] J. N. Reddy, A. A. Khdeir, L. Librescu, Lévy type solutions for symmet-
rically laminated rectangular plates using first-order shear deformation
theory, Journal of Applied Mechanics 54 (3) (1987) 740–742.

[15] M. Bodaghi, A. R. Saidi, Levy-type solution for buckling analysis of

re-
thick functionally graded rectangular plates based on the higher-order
shear deformation plate theory, Applied Mathematical Modelling 34 (11)
(2010) 3659–3673.
lP
[16] M. Mohammadi, A. R. Saidi, E. Jomehzadeh, Levy solution for buckling
analysis of functionally graded rectangular plates, Applied Composite
Materials 17 (2) (2010) 81–93.
rna

[17] M. Dan, A. Pagani, E. Carrera, Free vibration analysis of simply sup-

ported beams with solid and thin-walled cross-sections using higher-
order theories based on displacement variables, Thin-Walled Structures
98 (2016) 478–495.
Jou

[18] R. Li, P. Wang, Z. Yang, J. Yang, L. Tong, On new analytic free vi-
bration solutions of rectangular thin cantilever plates in the symplectic
space, Applied Mathematical Modelling 53 (2018) 310–318.

42
 Journal Pre-proof

[19] L. T. Tenek, J. Argyris, Finite elemente analysis for composite struc-

of
tures, Springer Science & Business Media, 2013.

[20] Y. X. Zhang, C. H. Yang, Recent developments in finite element analysis

pro
for laminated composite plates, Composite Structures 88 (1) (2009) 147–
157.

[21] Q. Guo, W. Yao, W. Li, N. Gupta, Constitutive models for the structural
analysis of composite materials for the finite element analysis: A review
re-
of recent practices, Composite Structures 260 (2021) 113267.

[22] Y. Feng, D. Wu, M. G. Stewart, W. Gao, Past, current and future

trends and challenges in non-deterministic fracture mechanics: A review,
lP
Computer Methods in Applied Mechanics and Engineering 412 (2023)
116102.

[23] T. Belytschko, T. Black, Elastic crack growth in finite elements with

rna

minimal remeshing, International Journal for Numerical Methods in En-

gineering 45 (5) (1999) 601–620.

[24] N. Moës, J. Dolbow, T. Belytschko, A finite element method for crack

growth without remeshing, International Journal for Numerical Methods
Jou

in Engineering 46 (1) (1999) 131–150.

[25] A. Morro, A Phase-Field Approach to Continuum Damage Mechanics,

Materials 15 (21) (2022) 7671, number: 21 Publisher: Multidisciplinary
Digital Publishing Institute.

43
 Journal Pre-proof

[26] M. Hofacker, C. Miehe, A phase field model of dynamic fracture: Robust

of
field updates for the analysis of complex crack patterns, International
Journal for Numerical Methods in Engineering 93 (3) (2013) 276–301,

pro
eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/nme.4387.

[27] F. Zhou, J. F. Molinari, Dynamic crack propagation with cohesive ele-

ments: a methodology to address mesh dependency, International Jour-
nal for Numerical Methods in Engineering 59 (1) (2004) 1–24, eprint:

re-
https://onlinelibrary.wiley.com/doi/pdf/10.1002/nme.857.

[28] M. S. Shephard, N. A. B. Yehia, G. S. Burd, T. J. Weidner, Automatic

crack propagation tracking, Computers & Structures 20 (1) (1985) 211–
lP
223.

[29] G. T. Camacho, M. Ortiz, Computational modelling of impact dam-

age in brittle materials, International Journal of Solids and Structures
rna

33 (20) (1996) 2899–2938.

[30] H. Cui, C. Du, H. Zhang, Applications of Phase Field Methods in Mod-

eling Fatigue Fracture and Performance Improvement Strategies: A Re-
view, Metals 13 (4) (2023) 714, number: 4 Publisher: Multidisciplinary
Jou

Digital Publishing Institute.

[31] R. Sepasdar, M. Shakiba, Micromechanical study of multiple transverse

cracking in cross-ply fiber-reinforced composite laminates, Composite
Structures 281 (2022) 114986.

44
 Journal Pre-proof

[32] W. K. Liu, S. Li, H. S. Park, Eighty Years of the Finite Element Method:

of
Birth, Evolution, and Future, Archives of Computational Methods in
Engineering 29 (6) (2022) 4431–4453.

pro
[33] C.-T. Chen, G. X. Gu, Effect of constituent materials on composite per-
formance: Exploring design strategies via machine learning, Advanced
Theory and Simulations 2 (6) (2019) 1900056.

[34] H. Feng, P. Prabhakar, Difference-based deep learning framework for

re-
stress predictions in heterogeneous media, Composite Structures 269
(2021) 113957.

[35] H. Hu, L. Qi, X. Chao, Physics-informed Neural Networks (PINN) for

lP
computational solid mechanics: Numerical frameworks and applications,
Thin-Walled Structures 205 (2024) 112495.

[36] C.-T. Chen, G. X. Gu, Physics-informed deep-learning for elasticity:

rna

forward, inverse, and mixed problems, Advanced Science 10 (18) (2023)

2300439.

[37] L. Liang, M. Liu, C. Martin, W. Sun, A deep learning approach to

estimate stress distribution: a fast and accurate surrogate of finite-
Jou

element analysis, Journal of The Royal Society Interface 15 (138) (2018)

20170844.

[38] A. Bhaduri, A. Gupta, L. Graham-Brady, Stress field prediction in fiber-

45
 Journal Pre-proof

reinforced composite materials using a deep learning approach, Compos-

of
ites Part B: Engineering 238 (2022) 109879.

[39] B. P. Croom, M. Berkson, R. K. Mueller, M. Presley, S. Storck, Deep

pro
learning prediction of stress fields in additively manufactured met-
als with intricate defect networks, Mechanics of Materials 165 (2022)
104191.

[40] Z. Nie, H. Jiang, L. B. Kara, Stress field prediction in cantilevered struc-

re-
tures using convolutional neural networks, Journal of Computing and
Information Science in Engineering 20 (1) (2019).

[41] M. Yacouti, M. Shakiba, Performance evaluation of deep learning ap-

lP
proaches for predicting mechanical fields in composites, Engineering
with Computers (Mar. 2024).

[42] A. Frankel, K. Tachida, R. Jones, Prediction of the evolution of the stress

rna

field of polycrystals undergoing elastic-plastic deformation with a hybrid

neural network model, Machine Learning: Science and Technology 1 (3)
(2020) 035005.

[43] Y. Wang, D. Oyen, W. G. Guo, A. Mehta, C. B. Scott, N. Panda, M. G.

Jou

Fernández-Godino, G. Srinivasan, X. Yue, Stressnet - deep learning to

predict stress with fracture propagation in brittle materials, npj Mate-
rials Degradation 5 (1) (2021) 6.

46
 Journal Pre-proof

[44] Z. Yang, C.-H. Yu, M. J. Buehler, Deep learning model to predict com-

of
plex stress and strain fields in hierarchical composites, Science Advances
7 (15) (2021) eabd7416.

pro
[45] J. R. Mianroodi, N. H. Siboni, D. Raabe, Teaching solid mechanics
to artificial intelligence—a fast solver for heterogeneous materials, npj
Computational Materials 7 (1) (2021) 99.

[46] R. Sepasdar, A. Karpatne, M. Shakiba, A data-driven approach to full-

re-
field nonlinear stress distribution and failure pattern prediction in com-
posites using deep learning, Computer Methods in Applied Mechanics
and Engineering 397 (2022) 115126.
lP
[47] Z. Yang, C.-H. Yu, K. Guo, M. J. Buehler, End-to-end deep learning
method to predict complete strain and stress tensors for complex hierar-
chical composite microstructures, Journal of the Mechanics and Physics
rna

of Solids 154 (2021) 104506.

[48] R. Sepasdar, A deep learning approach to predict full-field stress distri-

bution in composite materials, Thesis, Virginia Tech (May 2021).

[49] M. Maurizi, C. Gao, F. Berto, Predicting stress, strain and deformation

Jou

fields in materials and structures with graph neural networks, Scientific

Reports 12 (1) (2022) 21834, publisher: Nature Publishing Group.

[50] G. X. Gu, C.-T. Chen, M. J. Buehler, De novo composite design based

47
 Journal Pre-proof

on machine learning algorithm, Extreme Mechanics Letters 18 (2018)

of
19–28.

[51] M. Shakiba, D. R. Brandyberry, S. Zacek, P. H. Geubelle, Transverse

pro
failure of carbon fiber composites: Analytical sensitivity to the distribu-
tion of fiber/matrix interface properties, International Journal for Nu-
merical Methods in Engineering 120 (5) (2019) 650–665.

[52] L. Hernandez, R. Sepasdar, M. Shakiba, Sensitivity of crack formation

re-
in fiber-reinforced composites to microstructural geometry and interfa-
cial properties, in: Proceeding of the American Society for Composites,
Thirty-Fifth Technical Conference, DEStech Publications, Inc., 2020,
lP
pp. 1576–1591.

[53] Z. Liu, J. Zhou, Introduction to Graph Neural Networks, Springer Na-

ture, 2022.
rna

[54] K. Gurney, An Introduction to Neural Networks, CRC Press, London,

2017.

[55] K. O’Shea, R. Nash, An Introduction to Convolutional Neural Networks

(Dec. 2015).
Jou

[56] P. Ramachandran, B. Zoph, Q. V. Le, Searching for Activation Func-

tions, arXiv:1710.05941 [cs] (Oct. 2017).

[57] A. Krizhevsky, I. Sutskever, G. E. Hinton, ImageNet Classification with

48
 Journal Pre-proof

Deep Convolutional Neural Networks, in: Advances in Neural Informa-

of
tion Processing Systems, Vol. 25, Curran Associates, Inc., 2012.

[58] S. Santurkar, D. Tsipras, A. Ilyas, A. Madry, How Does Batch Normal-

pro
ization Help Optimization?, in: Advances in Neural Information Pro-
cessing Systems, Vol. 31, Curran Associates, Inc., 2018.

[59] D. Yu, H. Wang, P. Chen, Z. Wei, Mixed Pooling for Convolutional

Neural Networks, in: D. Miao, W. Pedrycz, D. Ślzak, G. Peters, Q. Hu,
re-
R. Wang (Eds.), Rough Sets and Knowledge Technology, Springer In-
ternational Publishing, Cham, 2014, pp. 364–375.

[60] B. Graham, Fractional Max-Pooling, arXiv:1412.6071 (May 2015).

lP
[61] L. Zhao, Z. Zhang, A improved pooling method for convolutional neu-
ral networks, Scientific Reports 14 (1) (2024) 1589, publisher: Nature
Publishing Group.
rna

[62] O. Ronneberger, P. Fischer, T. Brox, U-Net: Convolutional Networks

for Biomedical Image Segmentation, in: N. Navab, J. Hornegger, W. M.
Wells, A. F. Frangi (Eds.), Medical Image Computing and Computer-
Assisted Intervention – MICCAI 2015, 2015, pp. 234–241.
Jou

[63] K. He, X. Zhang, S. Ren, J. Sun, Deep Residual Learning for Image
Recognition, 2016 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR) (2016) 770–778.

49
 Journal Pre-proof

[64] A. Gupta, A. Bhaduri, L. Graham-Brady, Accelerated multiscale me-

of
chanics modeling in a deep learning framework, Mechanics of Materials
in press (2023).

pro
[65] A. Arteiro, G. Catalanotti, A. R. Melro, P. Linde, P. P. Camanho, Micro-
mechanical analysis of the in situ effect in polymer composite laminates,
Composite Structures 116 (2014) 827–840.

[66] H.-w. He, F. Gao, Effect of Fiber Volume Fraction on the Flexural Prop-
re-
erties of Unidirectional Carbon Fiber/Epoxy Composites, International
Journal of Polymer Analysis and Characterization 20 (2) (2015) 180–
189, publisher: Taylor & Francis.
lP
[67] D. Grund, M. Orlishausen, I. Taha, Determination of fiber volume frac-
tion of carbon fiber-reinforced polymer using thermogravimetric meth-
ods, Polymer Testing 75 (2019) 358–366.
rna

[68] M. A. Caminero, J. M. Chacón, I. Garcı́a-Moreno, G. P. Rodrı́guez, Im-

pact damage resistance of 3D printed continuous fibre reinforced ther-
moplastic composites using fused deposition modelling, Composites Part
B: Engineering 148 (2018) 93–103.
Jou

[69] R. Matsuzaki, M. Ueda, M. Namiki, T.-K. Jeong, H. Asahara,

K. Horiguchi, T. Nakamura, A. Todoroki, Y. Hirano, Three-dimensional
printing of continuous-fiber composites by in-nozzle impregnation, Sci-
entific Reports 6 (1) (2016) 23058, publisher: Nature Publishing Group.

50
 Journal Pre-proof

[70] M. Shakiba, D. R. Brandyberry, S. Zacek, P. H. Geubelle, Transverse

of
failure of carbon fiber composites: Analytical sensitivity to the distribu-
tion of fiber/matrix interface properties, International Journal for Nu-

pro
merical Methods in Engineering 120 (5) (2019) 650–665.

[71] S. Zacek, D. Brandyberry, A. Klepacki, C. Montgomery, M. Shakiba,

M. Rossol, A. Najafi, X. Zhang, N. Sottos, P. Geubelle, C. Przybyla,
G. Jefferson, Transverse failure of unidirectional composites: Sensitivity

re-
to interfacial properties, Springer International Publishing, Cham, 2020,
pp. 329–347.

[72] N. Tschoegl, Failure surfaces in principal stress space, Journal of poly-

lP
mer science Part C: Polymer symposia 32 (1) (1971) 239–267.

[73] W. Ramberg, W. R. Osgood, Description of stress-strain curves by three

parameters (1943).
rna

[74] J. C. Simo, J. Ju, Strain-and stress-based continuum damage models—I.

formulation, International journal of solids and structures 23 (7) (1987)
821–840.

[75] J. Simo, J. Ju, On continuum damage-elastoplasticity at finite strains,

Jou

Computational Mechanics 5 (5) (1989) 375–400.

[76] M. Ortiz, A. Pandolfi, Finite-deformation irreversible cohesive elements

for three-dimensional crack-propagation analysis, International Journal
for Numerical Methods in Engineering 44 (9) (1999) 1267–1282.

51
 Journal Pre-proof

[77] R. Sepasdar, M. Shakiba, Overcoming the convergence difficulty of co-

of
hesive zone models through a Newton-Raphson modification technique,
Engineering Fracture Mechanics (2020) 107046.

pro
[78] Y. Gao, A. Bower, A simple technique for avoiding convergence prob-
lems in finite element simulations of crack nucleation and growth on
cohesive interfaces, Modelling and Simulation in Materials Science and
Engineering 12 (3) (2004) 453.

re-
[79] B. Fiedler, M. Hojo, S. Ochiai, K. Schulte, M. Ando, Failure behavior
of an epoxy matrix under different kinds of static loading, Composites
Science and Technology 61 (11) (2001) 1615–1624.
lP
[80] U. Ayachit, The ParaView Guide: A Parallel Visualization Application,
Kitware, Inc., 2015.

[81] K. K. Dobbin, R. M. Simon, Optimally splitting cases for training and

rna

testing high dimensional classifiers, BMC Medical Genomics 4 (1) (2011)

31.

[82] V. R. Joseph, Optimal Ratio for Data Splitting, Statistical Analysis and
Data Mining: The ASA Data Science Journal 15 (4) (2022) 531–538.
Jou

52
 Journal Pre-proof

Declaration of interests

☒ The authors declare that they have no known competing financial interests or personal relationships
that could have appeared to influence the work reported in this paper.

of
☐The authors declare the following financial interests/personal relationships which may be considered
as potential competing interests:

pro
re-
lP
rna
Jou