Machine Learning for Composite Material Analysis

and its Optimization

Mamta Jaiswal*, Ruchita Kadam*, Madhura Kshirsagar*, Shravani Mahadik*, Priya Shah*, Shweta
Thorat*, Sakshi Tonde*, Dnyaneshwari Bangar*

∗
MKSSS’s Cummins College of Engineering for Women, Pune 411052

Abstract
Composite materials play a vital role in various industries due to their exceptional mechanical properties
and versatility. However, accurately analyzing and optimizing composite material properties pose
significant challenges. Traditional methods such as experimental testing and finite element analysis have
limitations in handling the complex nature of composite materials. In recent years, machine learning
(ML) has emerged as a promising approach to complement traditional techniques, offering new avenues
for composite material analysis and optimization. This paper provides an overview of the applications of
ML in composite material analysis, including prediction of mechanical properties, defect detection,
microstructure characterization, and optimization of material properties. We discuss the challenges and
considerations associated with ML-based approaches, such as data availability, model interpretability,
and generalization across material systems. Additionally, we explore future directions and emerging
trends in ML techniques for composite material science, including the integration of multi-scale
modeling and automation in materials discovery. Overall, this paper highlights the importance of ML in
advancing composite material analysis and optimization and identifies areas for further research in the
field.
Keywords: Composite materials, Machine learning, Material analysis, Material optimization,
multi-scale modeling.

1. Introduction

Composite materials have gained significant attention in various industries due to their unique
combination of properties, including high strength-to-weight ratio, corrosion resistance, and design
flexibility. These materials are extensively utilized in aerospace, automotive, renewable energy, and
infrastructure sectors, among others. However, the effective analysis and optimization of composite
material properties remain key challenges in the field of materials science and engineering. Traditional
methods for analyzing composite materials, such as experimental testing and finite element analysis,
have inherent limitations in capturing the complex behavior of these materials under different loading
conditions and environmental factors. Furthermore, the optimization of composite material properties
for specific applications requires a comprehensive understanding of the material's structure-property
relationships and the ability to efficiently explore the vast design space. In recent years, machine
learning (ML) has emerged as a promising approach to address these challenges and unlock new
opportunities in composite material analysis and optimization. By leveraging ML techniques,
researchers can enhance the accuracy and efficiency of composite material analysis, leading to improved
performance, durability, and cost-effectiveness of composite structures.
In this prospective paper, we discuss the comprehensive landscape of machine learning applications in
composite material analysis and optimization, aiming to provide a thorough understanding of the
potential and challenges in this emerging field. By reviewing relevant literature and case studies, we
elucidate the diverse applications of ML, ranging from predicting mechanical properties to optimizing
material structures for enhanced performance. Furthermore, we delve into the intricacies of ML model
interpretation, data challenges, and ethical considerations, shedding light on the broader implications of
adopting ML in material science research. Ultimately, this paper aims to foster dialogue, collaboration,
and innovation at the intersection of machine learning and composite material science, paving the way
for transformative advancements in materials design, manufacturing, and utilization.

2. Literature Review

Chun-The Chen et al [1] emphasizes on utilizing machine learning algorithms for design and analysis of
new materials/ configurations. It provides an overview of simple popular supervised machine learning
regression tools and neural networks and their application in composite materials modelling and design.
It provides a conclusion of various studies showcasing applications of machine learning algorithms in
composite material research.

Faiksaw Kibrete et al [2] study examines the application of artificial intelligence, machine learning and
deep learning techniques used for predicting the mechanical properties of composites. The study
provides a brief overview of the machine learning and the machine learning algorithms along with
citations of other studies where the algorithms have been applied and results are obtained and used for
the same purpose, same with deep learning. It also allows one to understand strengths and weaknesses of
each algorithm mentioned, allowing the reader to be able to judge the algorithms and make primary
decisions. The paper alo emphasizes on the challenges faced as well as areas that need further research
and focus for the development and utilization of the techniques for analysis of composite materials.

Cheuk Hei Chan et al [3] This paper covers the entire machine learning arena and its application in
advanced material prediction and design. The study explains the fundamental principle of ML and
workflow of a supervised machine learning model and addresses the problem of overfitting and
underfitting in ML algorithms. It also mentions the common ML algorithms and their applications in
material properties’ prediction following the citation showing application in prediction.

C. Velmurugan et al [4] The study focuses on fabricating hybrid Metal Matrix Composites (MMCs) via
stir casting, employing a central composite rotatable design matrix for experimentation. It utilizes
Artificial Neural Networks (ANN) to predict mass loss in dry sliding wear tests, with parameters
including SiC and graphite particle percentages, applied load, and sliding distance. The ANN, trained
using the backpropagation algorithm, demonstrates satisfactory agreement with experimental results,
suggesting its efficiency in composite material characterization. Scanning electron microscopy reveals
abrasion and delamination as predominant wear mechanisms, with delamination severity increasing
with sliding distance, correlating with both experimental and ANN-predicted mass loss values.

Jayanta Bhusan Deb et al [5] The study focuses on introducing a novel framework for predicting the
ablation performance of ceramic matrix composites under harsh environmental conditions using
multiple machine learning regression models. It is the first instance in literature where machine
learning techniques have been applied to forecast the thermal behavior of
boron-nitride-nanoparticle-containing continuous-fiber-reinforced silicon oxy-carbide ceramic matrix
composites in oxy-acetylene torch tests. Results indicate that the XGBoost model performs better with
70% of the torch test training data, while the GBR model shows better accuracy with 80% of the data.
Although the machine learning models lack the ability to explain underlying physical principles, they
offer a promising tool for optimizing design parameters and predicting the lifecycle of thermally
resistive composite materials, particularly in demanding applications such as aero engines and gas
turbines.

America Califano et al [7] The study focuses on implementing a method for structural health
monitoring (SHM) of composite structures using artificial neural networks (ANNs). The algorithm
correlates strains measured in a specific area of the structure to detect changes indicative of damage or
defect propagation. It utilizes data collected from a distributed optical fiber sensor embedded in a
composite plate subjected to four-point bending tests. Unlike other machine learning-based SHM
techniques, this approach is independent of applied loads and material type, eliminating the need for
data reduction procedures and damage pattern association. The method proves less time-consuming and
computationally expensive, making it suitable for real-life SHM applications. It exhibits higher
sensitivity to strain relationship changes compared to traditional strain maps and can identify anomalies
even in less severe damage situations. Future work aims to test the method on more noticeable damage
configurations to further explore its capabilities.

Sergei Evsevleev et al [8] The study focuses on performing a comprehensive 3D microstructural

characterization of a five-phase metal matrix composite (MMC) using synchrotron X-ray computed
tomography (SXCT). It introduces a workflow employing advanced deep learning-based segmentation
techniques to accurately delineate all individual phases present in SXCT data. The segmentation
process, facilitated by a fully convolutional neural network with U-net architecture, achieves high
accuracy even with limited training data. This enables the extraction of precise microstructural
parameters such as volume fractions and particle shapes, which can be utilized in micromechanical
models. The results obtained open up new avenues for quantitative 3D microstructural characterization
of complex materials using CT. Future work aims to utilize microstructural features extracted from
segmented CT data to refine existing analytical models and develop new finite element models. This
advancement holds significant promise for improving the predictive capability of micromechanical
behavior in multiphase metal matrix composites, thereby enhancing their practical utility.

Su Tian et al [9] explores how artificial neural network (ANN) models can be used to solve
fundamental practical problems in the design of composite materials and structures. It investigates their
potential application to train nonlinear constitutive models, increase speed of multiform modeling, and
optimize design. The motivation for this review comes from the increasing application of machine
learning in engineering. Despite promising benefits, there are still obstacles associated with ANN
models hindering their adoption. This also supports data-based advanced material designs such as
Materials Genome Initiative (MGI), positioning Machine Learning approaches more so ANN models at
the forefront of overcoming these challenges. By discussing both advantages and limitations of ANN
models, the article is a roadmap for future research that would contribute to better data-based
approaches for designing and analyzing composite materials and structures.

S. Nikbakt et al [10] This paper focuses on the research on the optimization of composite structures and
has predominantly focused on enhancing performance and minimizing weight. This includes
maximizing critical properties such as buckling load and fundamental frequency while reducing the
overall structural mass. These objectives aim to improve the structural efficiency and durability of
composite materials under various loadings. Studies have explored diverse strategies for achieving
these goals, including optimizing the stacking sequence of layers and other design variables.
Meta-heuristic algorithms such as genetic algorithm (GA) and particle swarm optimization (PSO) have
been widely utilized for their ability to find optimal solutions efficiently. As a result, these optimization
approaches contribute significantly to advancing the application of composite materials in engineering
fields.

Karam Maalawi [11] This paper focuses on how the optimization models help engineers maximize
stability and natural frequencies while minimizing mass and maximizing stiffness within design
constraints, enabling efficient and effective designs.

Jihong Zhu et al [12] This study highlights how topology optimization is a powerful structural design
method that generates innovative lightweight and high-performance configurations, particularly when
combined with additive manufacturing (AM). This integration leverages the strengths of both
techniques to produce complex, high-performing structures tailored for specific applications. Topology
optimization for AM involves designing complex configurations such as lattice and hierarchical
structures while considering process-specific constraints like overhangs, layer orientation, and
connectivity. Additionally, it addresses challenges such as material anisotropy, fatigue performance,
and scale effects on lattice structures to achieve optimal designs. By integrating material, structure, and
performance, this approach offers opportunities to create multifunctional, lightweight, and efficient
structures with broad applications in aerospace and modern manufacturing. This combination paves the
way for next-generation engineering solutions with enhanced performance and extensive engineering
applications.

R.H. Lopez et al [13] The study investigates how different first ply failure criteria affect the
optimization of laminated composite structures using genetic algorithms (GAs). The objectives include
minimizing weight and material cost of composite plates under in-plane loads. By using GAs, the study
efficiently addresses non-convex and multimodal optimization problems typical in composite design.
The analysis focuses on three failure criteria: maximum stress, Tsai-Wu, and Puck failure criterion
(PFC), emphasizing PFC for its practical alignment with real-world observations. The results reveal
that optimal designs vary significantly depending on the chosen failure criterion and loading
conditions. Therefore, careful selection of failure criteria is crucial in GA-based optimization to ensure
accurate and reliable composite designs.

Bofeng Shi et al [14] This study focuses on multi-objective optimization (MOO) methods, such as
evolutionary algorithms and Bayesian optimization, enabling materials scientists to simultaneously
optimize multiple material properties, balancing trade-offs between different objectives. These methods
excel at high-dimensional problems and can guide efficient exploration of Pareto fronts for optimal
materials solutions. Challenges include managing prediction uncertainties, computational complexity,
and incorporating domain-specific knowledge. Innovative strategies like uncertainty-aware
optimization and the integration of machine learning can enhance the efficiency of MOO in materials
science. Multi-scale and multi-physics models offer further opportunities for advancing the field.

J. Maniraj et al [15] This study focuses on how polymer composites are widely utilized in various
applications due to their unique characteristics, yet designing and optimizing these materials remains a
complex and resource-intensive process. Machine learning (ML) presents a promising avenue to
streamline this process by providing predictions of composite properties based on their microstructures.
This literature review aims to explore the potential of ML techniques in enhancing the design and
optimization of polymer composites while addressing challenges and suggesting future research
directions.ML techniques offer the capability to predict composite properties by analyzing
microstructural features. Supervised learning methods, such as regression and classification algorithms,
enable the prediction of mechanical, thermal, and other properties based on input features.
Unsupervised learning approaches, such as clustering, aid in identifying patterns within composite
microstructures, facilitating material characterization and optimization.

Diab W. Abueidda et al [16] The paper presents a convolutional neural network (CNN) model to
accurately predict mechanical properties of two-dimensional checkerboard composites. Trained on
finite element analysis data, the CNN exhibits high accuracy in capturing stiffness, strength, and
toughness. Integration with a genetic algorithm optimizer yields optimized composite configurations.
The study underscores the potential of CNN models in structural and materials analysis, suggesting
they could revolutionize optimization techniques. Future work may explore interface effects and
involve fabricating and testing optimized composites.

Christian Emeka Okafora et al [17] This paper explores the integration of machine learning (ML)
techniques for the design of reinforced composite materials, aiming to overcome the constraints of
traditional approaches. It offers a detailed protocol for implementing ML in composite materials
technology, emphasizing the importance of data hygiene. The review covers various aspects, including
the application of ML algorithms in material and process selection, data sourcing methods, and
emerging digital tools for ML implementation. Additionally, it identifies research gaps and suggests
areas for future exploration, aiming to guide further advancements in utilizing ML for composite
material design.

A. Sharma et al [18] This paper explores how machine learning (ML) can enhance the understanding
and application of polymer composites by addressing various aspects such as prediction, optimization,
feature identification, uncertainty quantification, reliability, and sensitivity analysis. Despite challenges
like overfitting and noise, ML techniques like neural networks, regression, and ensemble methods offer
significant potential for advancing polymer composite research. The paper highlights promising trends
such as hybrid ML, adaptive learning, and multi-scale analysis, providing valuable insights for
researchers and practitioners and offering recommendations for addressing challenges and guiding
future research directions.

3. Composite material analysis technique

3.1. Machine Learning as a complementary approach for composite material analysis

As the composites are structural materials, there is a need to design and develop more complex and
advanced materials. With the application of additive manufacturing, it is now possible to develop vast
designs using composites. This design space also increases analysis complexity for composites [1].
While the above-mentioned conventional processes offer effectiveness, to overcome their limitation,
Machine Learning is utilized for the analysis of composite materials enabling complex and multiscale
analysis by predicting properties of the composite materials, identifying patterns, generating new
configurations and evaluating their performance and reliability. Thus, ML models possess importance
not only in analysis and prediction of material properties, but also in designing composites with desired
properties [1] [2].

Machine learning is a subset of artificial intelligence that utilizes algorithms typically known as
machine learning models for predictive analysis, decision making, classification of data without
explicit coding. It has its applications widespread in various domains, one of which is in composite
materials analysis and optimization. They possess the ability to learn from the input data, identify
patterns and make informed decisions. These models are imported and utilized through in-built libraries
offered by various coding languages [2].

Figure 1. Classification of Machine Learning Algorithms [2]

As shown in the Fig 1, machine learning algorithms are classified as [2]:

1. Supervised Learning: In this type of learning, the model is trained with labelled data, that is, the
data is tagged with informative labels so that the model can learn from it. It is used for both
classifications, where input is categorized as per the attributes and regression to predict values. It
can be used to predict and classify composite material properties and types.
2. Unsupervised Learning: In this type of learning, the input data is not labelled and the model
recognizes patterns in the input data and forms clusters. It can be used to group the composite
materials based on their patterns or structures.
3. Semi-Supervised Learning: In this learning approach, the model is fed with both labelled and
unlabeled data.
4. Reinforced Learning: In reinforced learning, the learning is based on trial and error/ or iterative
method to make predictions while interacting with the dynamic environment.

Out of all the types of ML models mentioned above, supervised learning is the most widely used
learning approach used for material properties prediction while reinforced learning is the least used
model. ML provides material property prediction at both macroscopic and microscopic levels, usually
using regression analysis method and design and discovery of new material using probabilistic ML
models. Its fundamental concept for material property prediction is to determine the relationships
between properties and existing factors. The predictions carried out by ML models assists to carry out
quick analysis allowing patterns and properties’-based predictions for the composite materials [2] [3].

3.2 Overview of relevant machine learning models

Various ML algorithms are used for material analysis and properties’ prediction for composites and
choosing the specific algorithm directly affects it. Size of dataset is one of the essential characteristics
for appropriate model selection, that is, large datasets are complex and are not suitable for an ordinary
computation system to handle, they are suitable to handle small datasets. Out of all the algorithms, the
commonly used are the following:

3.2.1 Support Vector Machine (SVM) [2]

SVM or Support Vector Machine is the most popular and commonly used ML algorithm that comes
under supervised learning. It is used for both classification and regression analysis. It uses data points
to form the best line (in 2D) or plane (in 3D) or decision boundary called hyperplane to form the
categories and the gap between the hyperplanes or decision boundaries is called margin. SVM aims to
find an optimal hyperplane that has maximal margin. The samples that are nearest to the or are on the

decision boundary are called support vectors and govern the decision boundary equation. Support
Vector Regression or SVR is an extension of SVM for regression analysis for both linear and non-linear
fitting problems

Figure 2. General schematic of SVM [2]

SVM provides high prediction speed and possesses the ability to handle high dimensional data using
Kernel function. Support Vector Machine Learning (SVML), Support Vector Machine Polynomial
(SVMPOLY), Support Vector Machine Radial Basis Function (SVMRBF), and Support Vector
Machine Sigmoid (SVMSIGMOID) are the most usually used Kernel functions. However, they are
inefficient for large datasets and not suitable for handling noisy data and outliers.

3.2.2 k-Nearest Neighbour (k-NN) [2]

The k-NN or k- Nearest Neighbour algorithm falls under supervised learning which can be utilized for
both classification and regression. The model is trained and tested with the labelled data which is then
stored by it. When new data is entered in the algorithm for class prediction or regression analysis, it
measures the distance between the new data and existing training datasets is measured using Euclidean
or Manhattan distance measurement. To make predictions, the nearest number of neighbours are
considered for accuracy, that is, the k value. For regression, averaging of the nearest k neighbours is
done.

Fig 3. General schematic of k-NN [19]

The algorithm is simple and easy to understand and implement, however, with large datasets the
performance speed is slow. Since it works on an instance-based approach utilizing the training datasets,
it requires storage space. The value of k affects the decision greatly.

3.2.3 Decision Tree [2]

Decision tree algorithm is a tree-based algorithm, used for both classification and regression, which
comes under supervised learning approach and is utilized to predict the mechanical properties of
composites. Its schematic is of hierarchical form where after providing input, at every node a decision
is made (at decision node) on the basis of attributes until the output is deduced (at leaf node), that is,
the predicted class label, that is, from root node to leaf node (terminal point of the tree). It establishes
relationships to predict material properties.

Figure 4. General schematic of Decision Tree [2]

3.2.4 Ensemble Learning Techniques [2]

Random Forest regression is an ensemble learning technique where its ensembles trees or SVM models
randomly for higher accuracy in prediction of mechanical properties of composites. AdaBoost or
Adaptive Boosting is another ensemble learning technique used to predict the mechanical properties of
composites. The AdaBoost regression follows an iterative approach to improve prediction accuracy of
the tree by forcing them to learn from the errors of the previous trees. Another effective ensemble
learning algorithm XGBoost or Extreme Gradient Boosting utilizes gradient descent optimization for
enhancing prediction accuracy of the model. Such regression models used alongside decision trees
increase efficiency of decision trees for composite material analysis and properties’ prediction.

While decision trees are easy to understand and interpret, they are prone to overfitting, where due to
excess training, their predictions are right for the training datasets but not for the new datasets.

3.2.5 Artificial Neural Network (ANN) [2]

Artificial Neural Networks or ANN is one of the most popular supervised machine learning algorithms
for prediction of mechanical properties of composite materials connecting various nodes (artificial
neurons). Inspired by the biological neural network, it consists of an input layer to intake features, a
hidden layer where each neuron receives data from the input layer, integrates it and then using a
straightforward computation gives predictions to the output layer. It has the ability to handle complex
non-linear networks along with high speed and accuracy with parallel information processing
capability.
 Fig 5. General schematic of ANN [2]

However, it is computationally expensive, requiring large amounts of training data and special
hardware requirements compared to SVM, k-NN and XGBoosting.

3.2.6 Regression Models [1]

3.2.6.1 Linear Regression

Linear regression is the most basic and popular supervised learning algorithm that interprets the
relation between the input and output and thus is widely used due to its simplicity and ability of
prediction of continuous quantities. Inputs such as treatment conditions and material properties are
entered to predict mechanical properties such as strength.

3.2.6.2 Logistic Regression

Logistic regression is an alternative used for classification of problems. It computes the probability of
the input of being in each class thus estimating the structure properties.

In conclusion, SVM provides high speed and accuracy while it is inefficient for small datasets whereas
k-NN being simple and possessing the ability to compute large datasets, its speed decreases with the
increase in size. Decision Tree and Artificial Neural Network are high-powered and can be used to
solve complex problems with large datasets but require additional domain knowledge along with their
individual weaknesses [2] [3].

3.3 Comparison of ML algorithms

3.3.1 Comparison of machine learning methods and finite element analysis on the fracture
behavior of polymer composites

Composite materials have high application in various sectors such as aerospace, building, automotive
as well as biomedical and thus are forced to strive in tough surrounding, experiencing uneven stress
distribution and structural damages, hence it is necessary to know the fracture toughness of composites.

Testing Methods:

a. Experimental Testing: Arcan Test

b. Finite Element Analysis

c. Machine Learning Algorithms: Decision Tree, k-Nearest Neighbour, Random Forest, Support Vector
Mechanism with Kernel Functions

In Arcan Test, specimen is required in butterfly form, that is, it is notched from both sides which was
created using a milling cutter.

Materials:

a. Pure Carbon/ Epoxy Laminate

b. Pure Glass/ Epoxy Laminate

c. Intra-ply hybrid knitting Glass-Carbon/ Epoxy fiber reinforced composite plate (4 laminae, 50%
carbon, 50% glass fiber, widths 12.5 mm, 25 mm, 50 mm)

Loading Methods:

Mode I: Load applied tries to open the crack tip causing damage to specimen, applied parallel to wale
direction, angle 0°

Mode II: Load applied shear the crack tip causing damage to specimen, applied parallel to the course
direction, angle 30°, 45°, 60°

Mode I/II: both Mode I and Mode II effects appear simultaneously at the crack tip, angle 90°

A. Experimental Testing: Arcan Test

Fig 5. Pure and hybrid cracked Arcan Test specimens [20]

B. Finite Element Analysis

ANSYS FEM program was used to model and solve damage behavior of composite structures.
Table 1: Mechanical Properties of pure knitting fabric reinforced composite specimens

The mechanical properties from Table 1 were used to perform the analysis.

Fig 6: 2D FEA model of Arcan Test specimen (a) before and (b) after numerical analysis

C. Machine Learning Algorithms:

The data for ML algorithms is divided into three parts: training, testing and validation. Validation is
required since the test evaluation methods may not be reliable. K Fold cross validation is used where
training and testing data are integrated and divided into K equally sized subjects. For performance
evaluation and model selection, the commonly used regression metrics are root-mean-square error,
mean average error and coefficient of determination. Best hyperparameters are selected using Grid
Search (GS) or Parameter Optimization (PO).

Results and Discussion

Table 2: Fracture Toughness of carbon/ epoxy pure composites having a crack at carbon side
Table 3: Fracture Toughness of glass-carbon/ epoxy hybrid composites with 50 mm pattern width
having a crack at carbon side

Table 4: Fracture toughness of glass/ epoxy pure composites having a crack at glass side

Table 5: Fracture Toughness of glass-carbon/ epoxy hybrid composites with 50 mm pattern width
having a crack at glass side

Table 6: error performance metrics of results according to Experimental and FEA

Conclusion

MLA methods are particularly preferred in applications that are difficult to extract equations or
modelling of experimental structure. In the following analysis, it was shown that MLA can be used in
fracture mechanics problems in terms of computer computational methods when compared to FEA. DT,
KNN and RF algorithms produce good results for experimental data.
4. Machine Learning Applications in Composite Material Analysis

4.1 Prediction using ANN on mass loss of hybrid composites

In many applications, metal matrix composites (MMCs) are gradually taking the place of conventional
materials, particularly in the automotive and aerospace industries where performance standards are
becoming more and more stringent. Artificial neural networks (ANNs) can be used to model and
forecast the mechanical properties of the composites, including their density, tensile strength, hardness,
and yield strength. The application of artificial neural networks (ANNs) is a novel approach in various
fields, such as materials science and materials processing operations (Bhadeshia, 1999; Genel et al.,
2003). The current study aims to construct an artificial neural network model to forecast the mass loss
of hybrid composites, as well as to produce Al6061 metal matrix composites reinforced with silicon
carbide (SiC) and graphite particles and examine their wear behavior. The pin-on-disc wear testing
machine was used for dry wear tests

Experimental Design [4]

● Factor Levels: For every factor that was determined, there were five levels: applied load (X3),
sliding distance (X4), weight percentage of graphite particles (X2), and weight percentage of
SiC particles (X1).
● Experimental Design: For testing, the central composite rotatable design was selected. Table
III displays this design matrix, with each column denoting a factor. Thirty-one trials were
carried out with various factor level combinations.

Data Collection [4]

● Data Points: Information on wear behaviour, including specimen weight loss, was gathered
throughout the trials.
● Experimental Data Base: The investigation's experimental database was created using the data
that was gathered.

Table 7: Factors and its levels

4.1.1 Artificial Neural Network

An artificial neural network (ANN) is a mathematical model whose structure and operation have been
compared to that of the human brain. It is made up of several densely interconnected processing
components arranged into layers. To minimize the sum square error, the current work applied a
multilayer feed forward AN with error BP training technique (Figure 2) based on weight updates.
In this work, supervised learning was employed, and the neural network toolbox included in
"MATLAB 6.5" software was utilized to train the ANN for 24 input-output patterns (MathWorks
Incorporation, 2005). Four neurons make up the input layer of the multilayer feed forward artificial
neural network design (X1, X2, X3, and X4), and one neuron makes up the output layer (equivalent to
one output, i.e. mass loss). A neural network's architecture includes details about how many layers the
network has, how many neurons are in each layer, and how the layers are connected to one another. In
this work, the neuron architecture was found by a process of trial and error. In the current investigation,
one hidden layer with eight neurons was used. [4]

Figure 7: Multilayer feed forward ANN architecture adopted in this study

Table 8: Design matrix
The network is successfully trained using the subsequent learning factors:
● The quantity of layers: 3.
● The quantity of input layer neurons: 4.
● The quantity of hidden layer neurons: 8.
● The quantity of output layer neurons: 1.
● Rate of learning (h) ¼ 0.05.
● Momentum constant (a) ¼ 0.85.
● Minimum number of epochs ¼ 1,000.
● Tolerance for MSE ¼ 0.00001.
● Training parameter learning rule: BP

4.1.2 ANN results

Feed forward BP neural network that was fully created was used to train and execute the ANN. To
assess the effectiveness of the suggested ANN, absolute percentage errors were utilized. It was
discovered that there was very little difference between the experimental and predicted values. [4]

(a) (b)
Figure 8: (a) Comparison for the training stage of mass loss (b) Comparison for the testing stage of
mass loss
4.2 Predicting the Ablation Performance of Ceramic Matrix Composites

Predicting the ablation performance of ceramic matrix composites involves assessing how well these
materials withstand high-temperature and high-velocity environments, such as those encountered
during atmospheric re-entry or in rocket propulsion systems. Ablation refers to the process where
material is removed from the surface of an object due to various processes like heat, vaporization, or
erosion. Ceramic matrix composites (CMCs) are materials composed of a ceramic matrix reinforced
with fibres or whiskers. They are known for their high strength, stiffness, and thermal stability, making
them attractive for applications where high-temperature resistance is crucial. Researchers use various
experimental techniques and computational modelling approaches to predict how CMCs will behave
under specific ablation conditions, allowing for the optimization of material design and performance
for particular applications. [5]

4.2.1 Data Preparation for Predictive Modelling

Material and Fabrication Process: A silicon oxy-carbide composite reinforced with continuous fibres
was utilized as the CMC sample for this investigation. In the polymer infiltration and pyrolysis
technique (PIP), woven carbon textiles and poly-siloxane (PSX) resin are used in the fabrication
process.
Experimental Setup: An oxyacetylene torch is used to generate high temperatures, up to 2200°C, on
the CMC sample during a torch test. Heat flux sensors are used to monitor the heat flow at various
distances from the flame while the sample is fastened to a slidable panel. The sample's rear surface
temperature is tracked using thermocouples.
Pyrolysis and Testing: After being post-cured and autoclave-cured, the composite samples are
pyrolyzed in a tube furnace. The samples are burned through during the torch test, which yields
information on ablative properties such as insulating index and erosion rate.
Data Collection: Three torch tests are carried out in a consistent environment, maintaining a distance
of 25 mm from the torch nozzle and a heat flux of 524 W/cm2. Each test's burning through temperature
data is gathered and used to create machine learning prediction models. [5]

4.2.2 Machine Learning Predictive Modelling

In this work, machine learning (ML) models were run using Python 3.9 from Google Collab. Every
machine learning algorithm used in this investigation adhered to the typical procedure outlined in scikit
learn [38] using Python. This study's process for creating each machine learning model is graphically
represented in Figure 1.

The following ML models were employed in the study:

● Decision Tree Regression Modelling

● Random Forest Regression Modelling
● Support Vector Machine (SVM) Regression Modelling
● Gradient Boost Regression (GBR) Modelling
● Extreme Gradient Boost Regression (XGBoost) Modelling
● Adaptive Boosting Regression Modelling

Figure 9: Methodology used to create regression models for machine learning

4.2.3 Parameters for Measuring Performance Accuracy

The performance of the machine learning regression models is evaluated in this study using four
evaluation metrics: coefficient of determination (R2), mean absolute error (MAE), mean absolute
percentage error (MAPE), and root mean square error (RMSE) analysis for the test dataset. [5]

4.2.4 Comparing the Performance of ML Models

Figure 10: Performance (a) RMSE and (b)MAE results of individual replicates with test data
percentage.

This paper aims to investigate how well different machine learning (ML) regression models perform in
predicting the behaviour of ceramic matrix composites (CMCs) during ablation. The correctness of the
models is evaluated using measures like root mean squared error (RMSE), mean absolute percentage
error (MAPE), coefficient of determination (R-squared score), and mean absolute error (MAE). The
most accurate boosting ensemble algorithms are Gradient Boosting Regression (GBR) and XGBoost,
with R-squared scores almost hitting 1. This indicates good agreement between the observed and
predicted temperature data from torch tests. These models demonstrate the capacity to effectively
forecast the burning through temperature of CMCs during ablation, outperforming previous models in
terms of MAPE, RMSE, and MAE. Computational efficiency is also taken into account; real-time
applications can benefit from XGBoost and AdaBoost's quick execution rates. The sequential boosting
technique of GBR enables it to identify intricate, non-linear patterns in the data, while the parallel and
distributed computing capabilities of the XGBoost model make it stand out. Both models perform
exceptionally well in terms of accuracy overall, but the quantity and complexity of the dataset might
affect how computationally efficient they are.

This study highlights the value of machine learning in predicting the operational longevity of ceramic
matrix composites in difficult applications like gas turbines and aero engines. It also represents the
beginning of an efficient method for predicting thermal properties using oxy-acetylene torch test data.
[5]

4.3 Identification of material defects and anomalies

The design of composite structures is often complex due to the challenges associated with their
manufacture and the effect on the resulting mechanical properties. The manufacturing process may
introduce defects such as voids, resin rich regions, and misalignment of fibers. The design of composite
structures becomes more critical where the concerned structures are subjected to anomalous or cyclic
loading conditions when in service. In addition, the wide range of damage and defects that can occur
independently and simultaneously has made it fundamental to monitor structures in real-time to detect
anomalies at an early stage. Furthermore, as aircraft travel is increasing considerably, the research
community is driven towards the simplified application of structural inspection methods, leading to
quicker maintenance procedures [7].

Machine Learning as a Tool for Structural Health Monitoring

The data acquired from SHM systems (both active and passive) may be considered as belonging to the
“big data” type, due to their wide complexity, variability, and diversity. Many challenges may be
encountered when dealing with big data, such as difficulty in searching, capturing, and storing
significant data, as well as in building adequate computational architectures and data-processing
models. So, the appeal of the big-data field has led to the development of sophisticated algorithms and
computing platforms that make it possible to reduce the size of the data itself.
Damage is defined as something that causes a modification to a material and/or its geometry, which
leads to a change in its mechanical performance. Machine learning based SHM in the context of
composite materials requires the definition of indicators that relate to different damage, based on the
evaluation of measured data and features. Machine learning approaches can be described as either
supervised or unsupervised. Supervised and partially supervised approaches require prior label
information, i.e. there is a set of training data available, where the relationship between the measured
data and physical damage phenomena is known. This prior knowledge is then applied when dealing
with the analysis of new datasets. Unsupervised approaches do not make use of prior information and
can be considered as ‘blind’. They make use of patterns and similarities in data features in order to find
labels [7].

ANN-based Damage Diagnosis Procedure

The current diagnosis procedure utilizes transfer functions to correlate strains experienced in
neighboring locations on the host structure, regardless of external loads. Strains are measured at
discrete locations, termed 'master' and 'slave' nodes, for relative strain comparisons. Neural networks
(ANNs) establish correlations between strains at these master and slave nodes, serving as a unique
'fingerprint' of the structure. Each master-slave node pair has its own neural network, describing the
strain relationship. Once initialized, the algorithm is operational. It operates on the assumption that
only nearby locations to a damage event are affected, while those further away remain unaffected.
Figure 1 outlines the steps of the algorithm applied to recorded strain data. A step-by-step description
of the procedure is given below:
Step 1. First, the reference structure is subjected to operational loads – in this case, 0.55 kN. During
loading, the strains in both master and slave locations R m;s n o are recorded.
Step 2. Then introducing some damage on the structure and re-subjecting it to the operational loads, the
strains D m;s n o is measured again.
Step 3. After the acquisition of all the necessary data, the core of the procedure is implemented. The
strains measured on the reference structure are used to train the neural networks. The strains measured
in slave locations (R s) are used as input data, while those measured in master locations (R m) are used
as output data.
Step 4. At this point, the strains measured in the slave locations on the damaged structure (D s) are
provided as new input data to the trained networks. The networks produce the strain predictions at
master locations ( P m ) as output data.
Step 5. The strain predictions output by Step 4 are compared to the actual strains measured on the
damaged structure (D m). The evaluation of the mismatch ð Þe , shown in Eq. (1), leads to the detection
of damage:

Figure 11: Damage Diagnosis procedure flowchart.

The current study introduces a method for structural health monitoring (SHM) of composite structures
using artificial neural networks (ANNs). The algorithm relies on establishing correlations among
strains measured in a specific area of the structure under examination. Changes in these correlations,
potentially caused by damage presence or defect propagation, serve as indicators for detection. The
algorithm was applied to data gathered by a distributed optical fiber sensor embedded in a composite
plate subjected to multiple four-point bending tests [7].
4.4 Characterization of Microstructures and Phase Analysis

In this study, Sergei Evsevleev investigated the 3D microstructure of an AlSi12CuMgNi alloy

reinforced with 7%vol of Al2O3–Saffil short fibers and 15%vol of SiC particles. It show the
application of an advanced DL method for the semantic segmentation of all phases present in this
multiphase material.The investigated composite was produced by squeeze casting.[12] Saffil short
fibers, which consist of 96–97% of Al2O3 and 3–4% of SiO2 binders, have planar-random orientation,
whereas SiC particles are expected to be randomly distributed within the volume.[8]
 Figure 12: a) SEM micrograph of the five-phase composite and b) enlarged view with indicated
constituent phases. Note that the micrographs are taken for the sample’s cross-section parallel to the
fiber plane (XY); c) SEM of the deep-etched sample and d) enlarged view with indicated constituent
phases.
The SEM micrographs of a polished composite sample reveal five distinct phases: Al2O3 fibers and
SiC particles appear dark gray, eutectic Si is light gray, and IMs are white. It's noted that SiC particles
tend to cluster between Al2O3 fibers. Further analysis using SXCT, after chemically removing the
Al-alloy matrix, highlights the intricate network of reinforcements. Various IM particles are
interconnected with eutectic Si and ceramic reinforcements. Eutectic Si forms bridges between SiC
particles and Al2O3 fibers, creating an interconnected network. However, these observations are
qualitative, lacking volumetric quantitative information. Thus, synchrotron X-ray computed
tomography (SXCT) is utilized for true 3D imaging, providing a comprehensive understanding of
phase distribution, orientation, and interconnectivity [8].

DL based segmentation

The assessment of the DL segmentation results compares the DL segmentation output with a manually
annotated slice. Despite a few discrepancies, primarily at object borders, the DL segmentation
generally detects most objects accurately. Quantitative evaluation employs the DICE coefficient, a
standard metric for assessing convolutional network performance in semantic segmentation tasks.[8]
Figure 13: Qualitative assessment of the segmentation result. a) Original 2D SXCT slice; b) Manually
segmented 2D SXCT slice: Al2O3 fibers in lilac, SiC particle in white, Si in red, and IMs in green; c)
U-Net segmentation result; d) Qualitative example of U-Net segmentation performance for Al2O3
fibers, SiC particles, IMs, and Si. Green: true positive, blue: false negative, and red: false positive (d).
Note that regions shown in (d) are also indicated in (a) by color-coded frames.

In summary, the study demonstrates the successful application of DL techniques to solve a

segmentation problem that conventional methods, aside from labor-intensive manual approaches,
couldn't address. The achieved accuracy enables estimation of volumetric characteristics for each phase
in a complex five-phase composite. 3D rendering reveals the formation of a eutectic Si phase between
ceramic reinforcements and IMs, consistent with observations from similar studies using SXCT in Al
alloy matrix composites with single ceramic reinforcement phases. These results present new
opportunities for quantitative 3D microstructural characterization of complex materials through CT.
Looking ahead, the study suggests using microstructural features extracted from segmented CT data to
refine analytical models and develop new finite element models, ultimately improving the prediction of
micromechanical behavior in multiphase metal matrix composites, enhancing their practical
applications [8].

5. Optimization of Composite Material Properties

5.1 Importance of material optimization in composite design

Material optimization of composite design focuses on achieving maximizing stiffness, minimizing

weight, cost reduction, and ensuring stability and strength by selecting the right materials based on
their mechanical properties and compatibility. The key parameters influencing the structural behavior
of composites include ply orientations, fiber volume fraction, number of layers, stacking sequence,
material of fibers and matrix, and layer thickness. These parameters serve as design variables in
optimization problems. [10][11]

5.2 Overview of optimization techniques in material science

The optimization techniques in material science involve adjusting the composition, structure, and
processing of materials to achieve desired properties and performance characteristics. These techniques
aim to find the most efficient configurations of materials to achieve optimal results based on specific
criteria such as strength, weight, cost, and sustainability. The following are some optimization
techniques used in material science:

5.2.1 Topology Optimization

Topology optimization is a structural design method that optimizes material distribution within a design
space to satisfy specified load conditions, performance, and constraints. In the topology optimization
process the initial design space is not limited to any predefined shape but explores a broader range of
design options. The methods for topological optimization are:

1. Evolutionary Structural Optimization (ESO): In this method the inefficient material is gradually
removed until the desired material volume is achieved.
2. Level Set Method (LSM): Level set method uses high dimensional level set functions to describe
structural boundaries. The optimum configuration with structural boundary is obtained by
 iteratively solving the Hamilton-Jacobi equation to update the level set functions.
3. Feature-Driven Optimization (FDO): Feature driven optimization is a method derived from level
set method. In this method the complex engineering structures are decomposed into simple
geometric features.

The topologically optimized structures often have complex geometries that can be challenging to
manufacture using conventional methods, such as machining and casting and require additional
treatments to improve manufacturability. Topology optimization is computationally intensive, requiring
advanced algorithms and computing power. Despite these challenges the technique is used to achieve
significant improvements in performance and efficiency [12].

5.2.2 Genetic Algorithm

Genetic algorithm involves finding the best stacking sequence or arrangement of layers to meet specific
design requirements. The algorithm uses mechanisms such as selection, crossover, and mutation to
evolve solutions over successive generations. The performance of genetic algorithms is improved using
various techniques, such as adjusting crossover and mutation probabilities.

The algorithm is used to optimize the fiber orientation and angle to achieve highest natural frequency
while considering the specified boundary conditions and minimize the displacement of the fibers when
loaded. The fiber orientation and angle can take one of four possible angles: 0, +45, -45, or 90 degrees
and give optimum result [13].

5.2.3 Multi-objective optimization (MOO)

Since composite materials are made of two or more materials it is difficult to optimize one material as
it may lead to the reduction of another material. The multi objective optimization finds the best
trade-offs between competing properties, optimizing multiple objectives simultaneously. The optimal
solutions in multi-objective optimization (MOO) problems are usually represented as a set of trade-off
solutions, the so-called Pareto set. The Pareto set provides a clear understanding of the trade-offs
between different objectives in the optimization problem. The most appropriate solution is chosen
based on the requirements [14].

5.3 Application of machine learning for material property optimization

Machine learning helps optimize composite structures and predict their mechanical properties. Machine
learning can identify patterns and relationships in large datasets, leading to the discovery of new
materials with tailored properties. The machine learning models used to predict the mechanical
properties of the composite materials when provided with input data. They provide valuable insights
into which input variables have a greater impact on predictions and can classify the design into "good"
or "bad" based on the mechanical properties (toughness and strength). Thus, preventing potential
failure and assessing the reliability of materials under different conditions. They are especially useful
for problems involving images or data that can be represented as images, such as the topology of
composites [1].

5.4 Steps in Developing Machine Learning Models for Predicting Polymer Composite Properties
from

Microstructural Data
 Fig13: Steps involved in the machine learning algorithms for predicting the properties of polymer
composite

1. Data Collection: Gathering a comprehensive and representative dataset of composite testing

outcomes, including microstructural details and property data. This involves acquiring information
about the composition, structure, geometry, physical, and mechanical characteristics of polymer
composites. Various methods such as microscopy, computed tomography (CT), image processing, and
testing methods like tensile, compression, and impact testing are used to collect this data.

2. Data Quality Assurance: Ensuring that the gathered data accurately reflects the materials and
microstructures of interest and is of superior quality. This may involve gathering data from multiple
sources, different contexts, and cleaning/filtering to eliminate outliers or mistakes.
3. Feature Engineering: Extracting pertinent features from microstructural data that can be used as
inputs for machine learning algorithms. These features include numerical features (e.g., average fiber
diameter, volume fraction of reinforcement), textural features (e.g., arrangement of reinforcement
fibers or particles, pore form, surface roughness), and statistical features (e.g., mean, standard
deviation, skewness, kurtosis of microstructural characteristics).
4. Model Training: This step involves calculating the parameters of the machine learning model
based on input and output characteristics of polymer composites. The process typically includes
dividing the available data into training and validation datasets. The training dataset is used to
determine the model parameters, while the validation dataset is used to evaluate the model's accuracy
and generalization to new, unseen data. Iterative modification of model parameters is done to
minimize the difference between predicted and actual composite characteristics, often assessed using
metrics like mean squared error or mean absolute error. Model improvement may involve altering
parameters or input features.
5. Model Validation: Model validation is essential to assess the performance of machine learning
models on unseen data and ensure their suitability for different data distributions. This step also
involves splitting data into training and validation datasets, with the training set used to determine
model parameters and the validation set used to evaluate performance on new data. Evaluation metrics
like mean squared error, mean absolute error, R-squared, and correlation coefficient are used to assess
model accuracy. Visual evaluation, such as plotting predicted values against actual values or the
discrepancy between predicted and actual values, helps understand model performance across different
feature spaces and identify potential issues or outliers.

5.5 Various machine learning models used for optimization:

 Fig. 15: Types of machine learning models

1. Linear Regression Models: These models forecast continuous output based on a linear
combination of input characteristics. They are simple and easy to understand but may not capture
intricate connections effectively.
2. Nonlinear Regression Models: Used for forecasting continuous output based on a nonlinear
combination of input variables. They are more flexible and can capture more complex relationships
but might be harder to interpret.
3. Decision Tree Models: These models forecast categorical output based on a sequence of decisions
derived from input characteristics. They are straightforward but may not capture intricate connections
effectively.
4. Random Forest Models: Employed to forecast categorical outcomes by combining multiple
decision trees. They are more flexible and can handle complex relationships but may be more
challenging to interpret [15].

5.6 Prediction and optimization of mechanical properties of composites using convolutional

neural networks

The convolutional neural networks (CNNs) are employed to predict mechanical properties based on the
spatial distribution of materials within the composite. CNNs are widely utilized for image recognition
tasks due to their ability to preserve spatial relationships between pixels, which is beneficial for
analyzing image-like data.

The typical architecture of a CNN includes convolutional layers, pooling layers, fully connected layers,
and activation functions. Convolutional layers apply filters to input images, preserving spatial
information, while pooling layers combine information from neighboring pixels, such as using
maximum pooling to extract key features. Fully connected layers and activation functions are similar to
those used in standard neural network architectures.

Two CNN architectures are developed, tailored for 8 × 8 and 16 × 16 grid sizes. Each architecture
comprises six composite layers, a dropout layer for regularization, and a fully connected layer. Each
composite layer consists of a 2D convolutional layer, 2D batch normalization, and rectified linear unit
(ReLU) activation function. The kernel sizes of the convolutional layers differ between the two
architectures, and the same hyperparameters are used for both models. Fig. 16 depicts the architecture
of the CNN developed for this study, illustrating the arrangement of convolutional layers, dropout
layers, and fully connected layers. This CNN architecture enables the prediction of mechanical
properties based on the microstructure of the composite material, facilitating efficient and accurate
analysis of complex material systems.
 Fig. 16: Illustration of the CNN model used in the present study.

The PyTorch framework and Nvidia Pascal Titan XP GPU platform was used for model development
and testing. During training, we minimize the mean squared error (MSE) to adjust model parameters,
while the mean absolute percentage error (MAPE) evaluates model accuracy on unseen data.
Optimization employs a genetic algorithm (GA), leveraging selection, crossover, and mutation
operators, with elitism preserving top-performing solutions. Chromosomes represent microstructure
configurations, and genes encode material types. Hyperparameters are carefully chosen for efficient
optimization. Overall, this approach integrates deep learning and evolutionary optimization for
effective composite microstructure modeling.

The successful training of a CNN model using epochs, where each epoch represents an iteration
through the training dataset. Convergence history plots for both 8x8 and 16x16 grid sizes indicate
effective optimization, with minimal overfitting managed through early stopping at 200 epochs. The
CNN model demonstrates high accuracy in predicting material properties, as evidenced by low mean
absolute percentage errors (MAPE) and maximum errors, even for the more complex 16x16 grid.
Notably, the CNN model surpasses the performance of linear models. Integration of CNN models with
genetic algorithms (GA) facilitates efficient optimization of composite configurations. GA strikes a
balance between exploration and exploitation, guiding the search towards optimal solutions.

Table 9: Performance of the CNN model: 8 × 8 grid.

MAPE Max error % of data points with

error > 5%

Modulus 0.6% 0.7% 0.0%

Strength 0.2% 0.4% 0.0%

Toughness 0.8% 38.8% 1.7%

 Table 10: Performance of the CNN model: 16 × 16 grid.

MAPE Max error % of data points with

error > 5%

Modulus 0.2% 2.4% 0.0%

Strength 0.4% 2.1% 0.0%

Toughness 2.4% 188.7% 0.082%

Furthermore, the strategic incorporation of soft materials within composite configurations to enhance
specific mechanical properties, such as toughness, while preserving other desirable characteristics like
strength and modulus. The optimization process with GA illuminates that introducing soft materials in
regions neighboring crack tips can notably augment toughness without compromising other mechanical
properties. This tactical placement of soft materials aligns intuitively with prior research findings.
Despite variations in the volume fraction of soft materials based on the optimized property, the results
consistently showcase a harmonious balance among different objectives. This balance minimizes the
need for significant compromise in achieving optimal composite configurations. Overall, strategically
integrating soft materials enhances the performance and resilience of composite structures, as
corroborated by the study's findings.

Fig. 17. Multi-objective optimization using genetic algorithm: The volume fractions of the soft
materials are: (a) 23.4% (modulus and strength), (b) 27.3% (modulus and toughness), (c) 28.1%
(strength and toughness), and (d) 32.8% (modulus, strength, and toughness).
 Table 11: Results of the multi-objective optimization using GA.

Output Modulus and Modulus and Strength and Modulus,

strength toughness toughness strength, and
toughness

Volume fraction 23.4% 27.3% 28.1% 32.8%

of soft material

Fm 0.84 0.81 0.79 0.76

Fs 0.99 0.99 0.99 0.97

Ft 0.86 0.94 0.94 0.97

5.7 Case study on application of machine learning for process optimization in Plastic Injection
Molding

Injection molding is a widely used manufacturing process for producing plastic components with high
precision and efficiency. However, achieving optimal process parameters to minimize defects such as
blush remains a challenge. Traditional methods rely on trial and error or simulation-based optimization,
which may be time-consuming and resource-intensive. This case study explores an innovative approach
that combines experimental testing with machine learning algorithms to enhance the injection molding
process.

Experimental Setup and Procedure:

The experimental setup involved feeding PVC compound into an injection molding machine and
setting process parameters using a digital control panel. After molding, high-resolution imaging was
employed to visualize blush defects, and measurements were taken to calculate defect area using an
ellipse area measurement formula. Experimental tests were conducted to validate finite element
analysis (FEA) simulations, ensuring accurate representation of process conditions.

Data Analysis and Modeling:

A fractional factorial design and central composite design (CCD) experiments were performed to
identify influential process parameters. Analysis of variance (ANOVA) was applied to analyze
experimental results and develop regression models for predicting blush defect area. Additionally, basic
and hybrid machine learning algorithms, including artificial neural networks (ANN), were trained and
validated using experimental data to further optimize process parameters.

Optimization Using Machine Learning:

The trained ANN models were used to predict blush defect area, which served as the fitness function
for genetic algorithm (GA) optimization. GA iteratively adjusted process parameters to minimize defect
area, leading to optimized parameter levels. The effectiveness of the optimization approach was
evaluated through comparison with experimental results and traditional ANOVA-based optimization.

Results and Discussion:

Experimental validation confirmed the accuracy of FEA simulations, providing a reliable basis for
subsequent optimization efforts. ANOVA analysis identified key process parameters affecting blush
defects, which were further refined using machine learning algorithms. The integrated approach yielded
significant reductions in defect area, demonstrating the effectiveness of machine learning-based
optimization.

Fig. 18: The impact of: (a) flow rate, (b) melt temperature, (c) holding pressure, and (d) runner
diameter on the extent of defect occurrence.

Conclusion:

This case study highlights the potential of integrating experimental testing and machine learning
techniques for process optimization in injection molding. By leveraging empirical data and advanced
analytics, manufacturers can achieve more efficient and reliable production processes while minimizing
defects and improving product quality. Future research may explore additional machine learning
algorithms and optimization strategies to further enhance process performance.[21]

6. Challenges & Considerations

In the realm of composite materials and structures, machine learning (ML) offers promising avenues
for design and analysis, yet several challenges impede its widespread application. The research papers
shed light on the intricate relationship between machine learning (ML) techniques and the design and
analysis of composite materials. One of the central challenges highlighted across these papers is the
issue of representative training data. Composite materials are inherently complex, and their constitutive
relations pose difficulties in obtaining sufficient and relevant data for training artificial neural network
(ANN) models. This challenge is further exacerbated in reinforced composite technology, where data
availability and interpretability become critical factors. ML algorithms demand abundant and reliable
data, which might be scarce in this domain, and the black-box nature of some algorithms complicates
the understanding of underlying physical principles.

Moreover, the integration of ANN models with mechanical systems for composite analysis presents
technical hurdles, particularly in multiscale inverse modeling scenarios. Composite materials exhibit
behaviors across different length scales, making it challenging to develop ANN models that accurately
capture these complexities. Additionally, the complexity of material behavior, such as
matrix-reinforcement interaction and manufacturing process, adds another layer of difficulty. Polymer
composites, in particular, pose challenges due to their complex behaviors influenced by various factors.

ML models are susceptible to overfitting and underfitting issues, impacting their accuracy and
generalizability. This is especially problematic given the high dimensionality and uncertainty inherent
in composite properties, leading to further challenges in traditional analysis methods. Challenges
specific to characterizing compressive damage in composite laminates using ML include the
complexity of strain-softening curves and the need to avoid overfitting noisy data. Furthermore, the
scarcity of high-quality data for training ML models and the necessity for algorithm verification against
experimental data contribute to the overarching challenges in the design and optimization of polymer
composites.

Despite these obstacles, the potential of ML to revolutionize composite design and optimization is
evident. Addressing these challenges through further research and development efforts will be crucial
for unlocking the full potential of ML in composite materials research. This necessitates
interdisciplinary collaboration between materials scientists, mechanical engineers, and data scientists to
overcome the technical hurdles and harness the transformative power of ML in advancing composite
materials technology.

7. Future Directions & Emerging Trends

7.1 Potential advancements in machine learning techniques for composite material analysis and
optimization

The future of machine-learning approaches for composite material analysis and optimization may be
genuinely transformative, tackling pressing problems while also opening up new opportunities. Recent
work has concentrated on developing deep learning architectures such as convolutional neural networks
(CNNs) to handle the intricacy of composites data. Such designs allow for accurate representations of
composite microstructures that enhance performance prediction and optimization. Additionally,
combining physics-constrained machine learning models allows a more comprehensive approach
towards the understanding of underlying physic laws governing composite materials behaviour leading
to improved accuracy of predicted responses and interpretability. Machine learning models integrated
with mechanical frameworks such as coupling artificial neural networks (ANNs) with finite element
(FE) analysis facilitate better optimization processes, improve efficiency and ensure model training
implicitly enforces physical constraints.
Though there have been significant advancements, computational cost in sampling and uncertainties in
design optimization remain challenges that continue to persist. Nevertheless, ongoing research efforts
are striving to solve these problems through new techniques. Surrogate models like ANNs enable
substantial reduction of computational time, simplification of the optimization process, as well as
bridging the gap between several optimization levels. Simultaneously, concerns about uncertainties
within the design optimization course are being addressed with machine learning models becoming
faster ways for quantifying and mitigating uncertainties than physics-based models ever were. While
scientists come up with innovative approaches in this area, it is predicted that machine learning will
advance rapidly in composite materials analysis and optimization enabling an efficient material design
and engineering.[9]

7.2 Integration of multi-scale modelling and machine learning approaches

The union of multi-scale modelling and machine learning in composite materials is a powerful strategy
for enhancing analysis and optimization. This can be achieved by combining advanced techniques such
as representative volume element (RVE) methods and mathematical homogenization theories with
artificial neural networks (ANNs) based on machine learning models, which enables researchers to
capture complex material behaviour across multiple scales. The use of machine learning to approximate
nonlinear material behaviour from multiscale data arises as an answer to computational challenges,
while physics-informed neural networks enable the incorporation of physical constraints into it by
ensuring that known laws are not violated. As a result, this collaboration leads to decreases in
computational costs, increase in interpretability with application of insights from physics into ANN’s as
well as some composite material physics that would drive material design or engineering
innovations.[9]

7.3 Prediction for the future of machine learning in composite material science

The recent improvements in this area, however, have shown that a major shift is possible. As of late,
many researchers have managed to predict the mechanical properties of composite materials by means
of simple machine learning algorithms unlike traditional molecular dynamics (MD) or finite element
method (FEM) techniques. Composite designs can be explored more efficiently using ML models
thereby resulting in the generation of optimized designs. However, even though machine learning has
always been known for its “black box” models, various design challenges are now starting to reveal
possible ways for improving toughness and strength in composites. There will be more future
breakthroughs concerning topology representations as well as inverse design methods that are effective
in large composite systems. For example, one interesting development which falls within this category
is augmentation of CNN models with self-learning algorithms for hierarchical composites designs and
application of NN models with backpropagation for solving inverse design problems. This, however
does not mean achieving full inverse designs but ongoing studies on the performance improvement
strategies will continue providing to us how important ML can be in hastening material discovery for
composite engineering.[1]

8. Conclusion

The study highlights the application of machine learning algorithms for the analysis and prediction and
optimization of composite materials. Supervised learning models are the widely ML algorithms
including SVM, k-NN, DT, Random Forest and other ensembling techniques and regression models for
the purpose where regression models; linear and logistic are simple and basic, the complex models
provide more accurate results and are able to solve complex problems. Several case studies emphasize
the actual application of ML for composite material properties, detecting material anomalies and
microstructure and phase analysis of composite materials. Utilizing ML for optimizing the composites
focuses on selecting the right materials based on their mechanical properties and compatibility while
ensuring stability and strength. ML driven optimization techniques are developed for optimization in
material science focusing on topology optimization, genetic algorithm and MOO. Each machine
learning model has its own strengths and weaknesses whereas as a whole, the biggest challenge noticed
is the availability and interpretability of data. However, future development in ML and its integration
with other techniques surely maximizes the utility factor of machine learning in the domain of
composites analysis.

9. References

[1] Chen, Chun-The, and Grace X. Gu. “Machine Learning for Composite Materials.” MRS
Communications, vol. 9, 2019, pp. 556-566.

[2] Faiskaw Kibrete, Tomasz Trzepiecinski, Hailu Shimels Gebremedhen and Dereje Engida
Woldemichael. “Artificial Intelligence in Predicting Mechanical Properties of Composite Materials.”
Journal of Composites Science (2023)

[3] Cheuk Hei Chan, Mingzi Sun and Bolong Huang. Application of machine learning for advanced
material prediction and design. EcoMat Functional Materials for Green Energy and Environment
(2022)

[4] C. Velmurugan, R. SubramanianS, Thirugnanam, B. Anandavel. Experimental study and prediction

using ANN on mass loss of hybrid composites, Industrial Lubrication and Tribology (2012)

[5] Jayanta Bhusan Deb, Jihua Gou, Haonan Song and Chiranjit Maiti. Machine Learning Approaches
for Predicting the Ablation Performance of Ceramic Matrix Composite, Journal of Composite Science
(2024)

[6] K. Genel a, S.C. Kurnaz b, M. Durman. Modeling of tribological properties of alumina fiber
reinforced zinc–aluminum composites using artificial neural network, Materials Science and
Engineering (2003)

[7] America Califano,Neha Chandarana. Damage Detection in Composites By Artificial Neural

Networks Trained By Using in Situ Distributed Strains, Applied Composite Materials (2020)
27:657–671

[8] Sergei Evsevleev, Sidnei Paciornik, Giovanni Bruno. Advanced Deep Learning-Based 3D
Microstructural Characterization of Multiphase Metal Matrix Composites, Advance Engineering
materials

[9] Su Tian, Fei Tao,Haodong Du, Wenbin Yu , How machine learning can help the design and analysis
of composite materials and structures (Oct 2020)

[10] S. Nikbakt, S. Kamarian, M. Shakeri. A Review on Optimization of Composite Structures Part I:

Laminated Composites (2018)

[11] Karam Maalawi. Introductory Chapter: An Introduction to the Optimization of Composite

Structures (2018)
[12] Jihong Zhu, Han Zhou, Chuang Wang, Lu Zhou, Shangqin Yuan and Weihong Zhang. A review of
topology optimization for additive manufacturing: status and challenges (2020)

[13] R.H. Lopez, M.A. Luersen and E.S. Cursi. Optimization of laminated composites considering
different failure criteria (2009)

[14] Bofeng Shi, Turab Lookman and Dezhen Xue. Multi-objective optimization and its application in
materials science (2023)

[15] J. Maniraj, Felix Sahayaraj Arockiasamy, C. Ram Kumar, D. Ashok Kumar, I. Jenish, Indran
Suyambulingam, Sanjay Mavinkere Rangappa and Suchart Siengchin. Machine Learning Techniques
for the Design and Optimization of Polymer Composites: A Review (2023)

[16] Diab W. Abueiddaa , Mohammad Almasrib , Rami Ammourahc , Umberto Ravaiolib , Iwona M.
Jasiuka, Nahil A. Sobhc. Prediction and optimization of mechanical properties of composites using
convolutional neural networks

[17] Christian Emeka Okafora, Sunday Iweriolorb, Okwuchukwu Innocent Anic, Shahnawaz Ahmadd,
Shabana Mehfuzd, Godspower Onyekachukwu Ekwuemee, Okechukwu Emmanuel Chukwumuanyae,
Sylvester Emeka Abonyif, Ignatius Echezona Ekengwua and Okechukwu Peter Chikelua. Advances in
Machine Learning-aided Design of Reinforced Polymer Composite and Hybrid Material Systems
(2023)

[18] A. Sharmaa, T. Mukhopadhyayb, S. M. Rangappac, S. Siengchinc, V. Kushvahaa. Advances in

Computational Intelligence of Polymer Composite Materials: Machine Learning Assisted Modeling,
Analysis and Design

[19] Christopher, Antony. A complete explanation of K-NN

[20] Balcıoğlu, H.E., Seçkin, A.Ç. Comparison of machine learning methods and finite element
analysis on the fracture behavior of polymer composites.

[21] Alireza Mollaei Ardestani, Ghasem Azamirad, Yasin Shokrollahi, Matteo Calaon, Jesper Henri
Hattel, Murat Kulahci, Roya Soltani, Guido Tosello. Application of Machine Learning for Prediction
and Process Optimization—Case Study of Blush Defect in Plastic Injection Molding