See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/383231911

Research Advances on the Role of Deep Learning in Materials Informatics

Article in Journal of Materials and Engineering · August 2025

DOI: 10.61552/JME.2025.02.004

CITATIONS READS

0 317

12 authors, including:

Opeyemi Akanbi Odumuwagun Adebowale

Ladoke Akintola University of Technology Pennsylvania State University
10 PUBLICATIONS 2 CITATIONS 7 PUBLICATIONS 0 CITATIONS

SEE PROFILE SEE PROFILE

Victor Ayanwunmi Gbemi Abass

University of Ilorin Hohai University
5 PUBLICATIONS 1 CITATION 8 PUBLICATIONS 16 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Oluwakunle Moyofoluwa Ogunsakin on 19 August 2024.

The user has requested enhancement of the downloaded file.

 Vol. 03, Iss. 2 (2025) 172-185, DOI: 10.61552/JME.2025.02.004

Journal of Materials and

Engineering
www.jme.aspur.rs

Research Advances on the Role of Deep Learning in

Materials Informatics

Opeyemi S. Akanbia, Oluwakunle M. Ogunsakinb, Idowu J. Ojoc, Adebowale A. Odumuwagund,

Victor Ayanwunmie, Abass F. Gbemif, Samuel Ayanwunmig, Julius A. Oladejod, Paul Adegbiteh,
Omogbeme Angelai, Olusola Akinrinolaa
aDepartment of Pure and Applied Physics, Ladoke Akintola University of Technology, Ogbomosho, Nigeria,
bDepartment of Geology, Geological Engineering and Petroleum Engineering, Missouri University of Science and Technology, USA,
cDepartment of Chemical Engineering, Ladoke Akintola University of Technology, Ogbomoso, Nigeria,
dDepartment of Mechanical Engineering, Federal University of Technology, Akure, Nigeria,
eDepartment of Mechanical Engineering, University of Ilorin, Kwara State, Nigeria,
fDepartment of Materials Science and Engineering, Hohai University, Nanjing, China,
gDepartment of Electrical Engineering, Yaba College of Technology, Lagos State, Nigeria,
hDepartment of Computer Science and Engineering, Ladoke Akintola University of Technology, Nigeria,
iDepartment of Business Analytics, University of West Georgia, USA.

Keywords: ABSTRACT
Deep learning
Deep materials informatics is a rapidly evolving field that employs deep
Deep materials informatics
learning techniques to develop predictive models for materials science. It
Neural networks
involves the use of large datasets, advanced algorithms, and high-
Data-driven science
performance computing to extract key features from complex materials data.
The aim of deep materials informatics is to speed up the process of locating
* Corresponding author: materials with desired attributes, by taking advantage of machine learning.
There are numerous applications of this technology, and it can forecast the
Oluwakunle Ogunsakin properties of materials at all levels - from the atomic to the macroscopic. For
E-mail: [email protected] instance, deep materials informatics are used to predict the electrical, thermal
and mechanical characteristics of materials, which are critical for designing
Received: 19 June 2024 new materials for various applications. This paper offers a thorough
Revised: 28 July 2024 examination of the fundamentals of deep learning, its benefits and drawbacks,
Accepted: 18 August 2024 and its current usage in the analysis of numerous materials datasets.
Additionally, it can be utilized to optimize the processing parameters for
creating materials with desired characteristics.
© 2025 Journal of Materials and Engineering

1. INTRODUCTION simulations, and other sources has created the

emergence of data-driven science. Advanced
In the current age of massive data, the amount of techniques for data-driven analytics are
information gathered from experiments, necessary to find useful information from the

172
 Opeyemi S. Akanbi et al., Journal of Materials and Engineering Vol. 03, Iss. 2 (2025) 172-185

data, which in turn can aid in accelerating Proper knowledge of complex systems must be
materials discovery and realizing the Materials obtained to effectively explore and develop novel
Genome Initiative (MGI) [1]. This fourth materials with the preferred characteristics.
paradigm of science uses ML and data mining Combining experiments and computer simulations
approaches to uncover actionable have enabled us to reduce the time and cost of
understandings from the data, as illustrated in materials design, [6,9], and the progress in
Figure 1 [2]. computing power and development of efficient
codes have allowed for computational high-
throughput studies [10] of large material groups.
These large-scale simulations and calculations,
together with experimental high-throughput
studies, [11,12] are providing a vast amount of data
that make it possible to use machine learning in
materials science. The lengthy and intricate process
has given us essential lessons that can be utilized in
materials science to generate successful outcomes.
Fig. 1. The four approaches of scientific exploration as
they relate to materials [2].
Fourth paradigm science data-driven methods are
being employed in various domains, for instance,
Researchers in materials science and engineering marketing, healthcare, climate science,
use experiments and simulations to try to bioinformatics, social media, materials science, and
comprehend the processing-structure-property- cosmology. Recently, Deep Learning has shown
performance (PSPP) relationships, which are not itself to be a successful and popular tool in data-
thoroughly understood [3]. Historically, driven analytics, particularly when applied to
experiments were the primary method for Materials Informatics. This paper will provide an
locating and characterizing new materials. overview of Deep Learning, its challenges and
Despite its potential, this method requires a great potential, before demonstrating examples of Deep
deal of equipment and resources and can take a Materials Informatics. Additionally, it will discuss
protracted period to yield results; thus, the exploration of PSPP linkages in materials. These
discoveries often occur by accident or intuition. A include an overview of workflow [13] and learning
computational revolution in materials science the chemistry of materials from elemental
was driven by the emergence of computational composition [14], predicting properties based on
techniques, such as molecular dynamics, density structure, [15,16] crystal structure prediction [17],
functional theory (DFT) [4,5], and Monte Carlo multiscale homogenization [18,19] and localization
simulations, which enabled researchers to [20] of high-contrast composites, structure
investigate the phase and composition space characterization [21,22] and quantification, [23,24]
more effectively. The PSPP relationships are and microstructure reconstruction [25] and design
integral to the study of materials and their [26]. We also provide a glimpse into the future of
properties. As depicted in Figure 2, the cause- deep materials informatics and wrap up the paper
effect relationships of science flow from left to with a summary and conclusion.
right, and the goals–means relationships of
engineering flow from right to left [2].
2. OVERVIEW OF DEEP LEARNING

Deep learning, a family of techniques in ML and AI,

is a re-examination of neural networks (NNs) which
were theorized in the 1980s [27,28]. The advent of
big data and big computing has enabled these
networks to become deeper, and they are capable
of learning and representing a wide selection of
nonlinear functions [29]. Deep learning has been a
powerful tool for automating the extraction of
Fig. 2. The interplay between engineering and materials
meaningful data from large datasets and has
science. This involves science moving from the left to the resulted in remarkable progress in a number of
right, and engineering from the right to the left [2]. areas, such as computer vision [30,31] and speech

173
 Opeyemi S. Akanbi et al., Journal of Materials and Engineering Vol. 03, Iss. 2 (2025) 172-185

recognition [32,33]. We will overview the deep have been some attempts to better
learning’s benefits and drawbacks, then discuss the understand the workings of the neural
constituent part of a deep NN, and at last, review network, they are not as easily interpretable
some of the networks which are used for deep as some traditional statistical models such
materials informatics. as linear regression [37].
Neurons, which are the basic components of
When compared to traditional ML methods, Artificial Neural Networks (ANNs), take their
deep learning has several advantages. By using design from biological neurons found in the
a hierarchical approach, it can identify human brain. Each neuron processes multiple
pertinent features in data without manual inputs and produces an output based on a
feature engineering. Deep learning, with its weighted sum of the inputs, and the output is then
ability to eliminate the need for feature transformed by a non-linear activation function
engineering, is particularly advantageous when (AF). ANNs commonly employ rectified linear
dealing with large datasets. Moreover, deep units (ReLU), linear, sigmoid, and leaky ReLU
learning is more accurate for massive data than activation functions. A fully-connected artificial
traditional machine learning models, which neural network (ANN) uses the ReLU activation
reach their limits much faster. Once the neural function, as demonstrated in Figure 3.
network is trained, it also has the advantage of
being able to make predictions quickly and with
great precision.

Deep learning has many advantages over

traditional ML techniques, yet, in certain cases,
there are still properties that complicate its
usage. Below are four of the major challenges of Fig. 3. (A). Depiction of a deep ANN (fully connected)
applying deep learning: with one output, four inputs, and five layers found
between the input and output layers with an
1. Lack of sufficient training data: Data used in inconsistent amount of neurons. (B). The graph
deep learning is typically large, curated and utilizes the ReLU activation function [24].
labelled, though this is a rarity in many
scientific and engineering fields such as Deep learning networks are a kind of ANN that
materials science [2]. consist of two or more hidden layers; a hidden layer
is situated between the input and output layers. The
2. Computing power: Creating deep learning structure of the network is established by how the
models necessitates a lot of computing neurons are interconnected. In addition, weights
capabilities and can take an extended period are assigned to the edges between neurons and are
to finish when working with large datasets, adjusted during the training process of the ANN so
even when utilizing the most up-to-date that the outputs become more accurate. To use an
hardware. Research is ongoing in the realm artificial neural network (ANN) to make
of parallelizing neural network training predictions, the input data is passed through the
algorithms [34,35]. network to yield the outputs. The loss or errors are
then calculated, and the weights of the connections
3. Network architecture search: Since neural between neurons are adjusted in order to minimize
networks are composed of interconnected the error, a process known as back-propagation.
neurons, there are a number of possibilities One pass of the entire training data is referred to as
for the network architecture, and though an epoch (which consists of one or more batches
there are some general guidelines for and a part of the dataset is used to train the neural
choosing an architecture for a given network) and it is repeated iteratively until the
problem, there is no formal method to weights have converged [38]. Normalization of the
determine the optimal architecture for a inputs and at times the inputs of the internal layers
task; this is an open research topic [36]. are carried out (known as batch normalization) to
stabilize the artificial neural networks. Dropouts
4. Model interpretability: Deep learning-based are a noteworthy and valuable idea in Artificial
models are usually seen as black-box models Neural Networks (ANNs) that involve randomly
due to their complexity, and though there

174
 Opeyemi S. Akanbi et al., Journal of Materials and Engineering Vol. 03, Iss. 2 (2025) 172-185

turning off certain neurons when undergoing a as the discriminator, which takes a picture as
backward or forward pass. This is a regularization input and reduces its resolution, generating a
method for reducing overfitting, and it is also found probability distribution to determine if the
to be an effective method for approximating multi- image is genuine or false. It can be likened to a
model averaging [38]. criminal creating counterfeit money, with the
discriminator as the police tasked with
2.1 Convolutional Neural Network (CNNs) recognizing the false currency. As the two
neural networks are trained, they become more
A Convolutional Neural Network (CNN) is an and more proficient, till they reach the Nash
Artificial Intelligence algorithm that is capable equilibrium [42]. GANs have been utilized in a
of analyzing an input image, determining the variety of image analysis applications, such as
relative importance (via learnable weights and high-resolution image synthesis [43], text-to-
biases) of various objects and features, and image synthesis [44],image editing [45],
differentiating among them. It is composed of blending [46], inpainting [47], and beyond
three main hidden layers: a convolutional layer, image domains, like music generation [48].
a fully connected layer, and a pooling layer. The
convolutional layer uses numerous kernels on 3. THE APPLICATION OF DEEP LEARNING TO
the image to generate abstract features, which UNDERSTAND PSPP RELATIONSHIPS IN
then serve as inputs for the next layer. The MATERIALS SCIENCE
outputs of this layer are then condensed into a
one-dimensional vector, which is used as the By examining how deep materials informatics
input for the fully connected layers, producing has been used to explore PSPP links through
the final prediction. The most common type of inverse and forward models, we can gain insight
CNNs are 2-D CNNs which work with 2-D input into the specific characteristics of deep learning
matrices, however, there are other variants in the materials sector.
such as 1-D and 3-D CNNs which take 1-D
vectors and 3-D matrices respectively as input, 3.1 Understanding Material Chemistry
and graph CNNs [39,40] that can work with Through Elemental Analysis
graphs as input.
Jha et al. recently developed a neural network
2.2 Generative Adversarial Networks (GANs) called ElemNet which requires only the
elemental composition of a crystalline
A generative adversarial network (GAN)[41] is a compound as input and predicts its formation
sophisticated deep learning architecture that has enthalpy [14]. This deep learning method does
been developed recently, based on game theory. not require any feature engineering, which is a
Figure 4 illustrates two neural networks paired major benefit when compared to traditional
together in a competitive way. machine learning algorithms. To test the
efficacy (considering its mean absolute error
(MAE)) of ElemNet, a dataset from the Open
Quantum Materials Database (OQMD) of
approximately 276, 000 compounds were used.
The Random Forest model - well-known for
being the most efficient of the traditional
machine learning approaches - achieved a
0.157 (electron volts per atom) MAE when only
Fig. 4. A GAN is composed of two neural networks, a using elemental components as features. By
generator and a discriminator, and when given the including composition-derived physical
right instruction, it has the ability to create lifelike attributes, however, the MAE was reduced
images/data from random noise [24]. significantly to 0.071 electron volts per atom.
On the other hand, ElemNet attained an even
The first one, labelled as the generator, takes a lower MAE of 0.055 electron volts per atom
random noise vector and increases its when only the elemental compositions were
resolution to produce a synthetic image. The considered. Thus, this result indicated that the
second is a convolutional network, referred to ElemNet was more successful than the

175
 Opeyemi S. Akanbi et al., Journal of Materials and Engineering Vol. 03, Iss. 2 (2025) 172-185

traditional machine learning algorithms, even 3.2 Crystal Structure Examinations for
when physical attributes were present. In Predicting Characteristics of a Material
addition, ElemNet was successful in predicting
convex hull phase diagrams for unseen material Even though composition-based models are
systems when two specifically designed highly accurate, the structure is still a major
training-test splits were evaluated. Further factor in materials because different
research into the effects of varying training set polymorphs and allotropes can have different
sizes demonstrated that ElemNet was more characteristics with the same chemical
successful than RF when over 4000 compounds makeup. Therefore, it is necessary to create
were present. With a training and test set of models that are sensitive to structures for
250,000 compounds each, ElemNet took predicting material properties. Several
approximately 7 hours to train on a GPU but research projects have been conducted, using
was much quicker when it came to prediction diverse sets of features to represent structure
time, requiring 0.08 and 9.28 seconds on a GPU information [49-52] for the ML algorithms to
and CPU respectively. build predictive models. Recently, deep
learning has been used directly on the crystal
Jha et al. [14] investigated the workings of structure, as discussed by Xie and Grossman
ElemNet to explain its successful results. It was [15]. A CNN framework was devised to extract
examined that the representations of the material properties from the connections
network had learned and studied the between atoms in the crystal. To begin, the
activations of the network when given certain crystal composition is expressed as a crystal
inputs. It was revealed that ElemNet had taught graph, in which the atoms in the unit cell are
itself features that are related to chemistry represented as nodes. A CNN is then
(such as element similarity and interactions), implemented on the graph, consisting of
despite not receiving any periodic table pooling layers, fully connected layers, and
information during training. When using convolutional layers, in order to draw out the
Principal Component Analysis (PCA) to map the best representations for modelling the
activations of the second and first layers specified characteristics. The database
generated by Alkali Metals (Caesium, Lithium, contained 46,744 materials from the Materials
Sodium, Potassium, and Rubidium) into a two- Project [53] which spanned 87 elements, 7
dimensional space, the results were arranged lattice systems, and 216 space groups. A model
in a straight line. Experiments with binary was initially used to predict formation energy
mixtures of Alkali Metals/Alkaline Earth Metals with a mean MAE of 0.108 electron volts per
and Chalcogens/Halogens elements conducted atom. This model had two layers, and the same
through the model showed that the charge- weights were used for all neighbouring atoms
unbalanced and balanced compositions in each layer. This was improved by the
clustered independently in the eighth layer of introduction of a convolution function that
ElemNet. ElemNet employs a deep learning considered the varying interaction strength
approach which is akin to that used for image between neighbouring atoms, resulting in an
recognition. In this process, initial layers are MAE of 0.039 electron volts per atom.
designed to recognize fundamental features Additionally, a similar framework was then
such as corners and edges, after which, for used in determining other density functional
instance, ‘shape’, a more complex element can theory-computed properties like shear moduli,
be identified in subsequent layers. This absolute energy, bulk moduli, Fermi energy,
technique proved to be especially accurate and Poisson ratio, and band gap from the Materials
rapid, allowing the authors to explore the Project. Recently, Ye et al. [16] demonstrated
quaternary space, analyzing nearly 500 million that through the use of artificial neural
compounds. Their findings included a variety of networks and only two descriptors, the density
hitherto unseen systems with at least one functional theory formation energies of
potentially stable compound. To reinforce the C3A2D3O12 garnets and ABO3 perovskites can
reliability of the results, some of these be accurately predicted, with mean absolute
compounds were found to be present in the errors of 0.007-0.010 electron volt per atom
Inorganic Crystal Structure Database (ICSD) and 0.020-0.034 electron volt per atom,
but were missing from the OQMD. respectively. The two descriptors are (1).

176
 Opeyemi S. Akanbi et al., Journal of Materials and Engineering Vol. 03, Iss. 2 (2025) 172-185

Pauling electronegativity. (2). Ionic radii of the 118-class classification difficulty. When the test
elements involved. Ye et al found a way to set was compared to its baseline value (which
model total cation disorder for mixed garnets was 20 per cent of the full atomic fingerprint
its type of species in D, A, and C sites are more data), the average level of inaccuracy was 31
than one, using an averaging technique and a per cent, an impressive result for such a
binary encoding scheme that only slightly complex system. Blunders that were chemically
reduced accuracy. logical made by the model were discovered to
be chemically (for instance 4f and 3d elements).
3.3 Prediction of Crystal Structure Also, the t-distributed stochastic neighbour
embedding (t-SNE), which is mostly used for
Foretelling crystalline formations is a difficult exploring data and representing high-
task in the realm of materials science, akin to dimensional data, was applied to the sigmoid
the way protein structure prediction is carried classifier weights of the elements shown. This
out in bioinformatics. Breaking the problem resulted in groupings similar to those of the
down into two parts, the first step is to generate periodic table. This indicates that the deep
possible structures and the second part is to learning model was capable of understanding
evaluate them to determine which is the most periodic trends and chemical similarities
probable. Usually, evolutionary algorithms without any explicit instruction, just like the
using random starting points are used for the ElemNet model [14] did with composition data.
structure generation step and quantum Turning to the crystal structure prediction
mechanical methods assess the structures. problem, Ryan et al [17] used a list of known
Recently, Ryan et al. applied deep learning to structure types as the basis for creating new
crystal structure prediction, particularly for crystal structures, and the deep learning model
structure evaluation [17]. The problem in to evaluate the structures. The generator used
predicting crystal structure was modified by 51,723 known structures as a base and
Ryan et al. by predicting the likelihood of each combined substitution across all elements to
atomic site. This was done by multiplying the generate a total of 2,703,834 ternary and
chance of the atom appearing in a certain spot. 623,380 binary possible crystal structures. This
To do this, data from the Crystallographic Open is a unique process that has never been done
Database and ICSD, which consists of around before. Ryan et al [14] combined a structure
704,000 distinct sites in around 52,000 crystal generation technique with deep learning based
structures, was used. In order to capture the structure evaluation to make crystal structure
local environment surrounding each site, the predictions. A holdout test set of 5845 crystal
model was trained using an input structures was employed to evaluate the
representation of atomic sites that was performance of the model. It was found that the
composed of normalized atomic fingerprints. correct structure was the highest-ranked choice
The input featured properties like three- 27% of the time, and it was part of the top 10
dimensional data retention, translational predicted structures 59% of the time. The
invariance, and a fixed number of dimensions authors of this study demonstrated the capacity
which enabled the model to master structural of deep learning models to autonomously gain
layouts as opposed to specific crystal the knowledge associated with chemistry from
formations. The deep neural network was made geometric and topological data, as well as how
up of three separate networks. To start with, a it can be applied to the intricate undertaking of
forty-two-layer convolutional variational crystal structure prediction. They
autoencoder compressed the 3072-dimensional demonstrated this in a case study of
atomic fingerprints into sixty-four-dimensional Manganese–Germanium and Lithium–
latent representations. Afterwards, a 5-layer Manganese–Germanium systems, which
sigmoid classifier was used to predict the revealed unique, novel chemical compositions
element combinations likely to form certain with matching predicted structure templates.
structural topologies based on latent Despite the fact that the accuracy of the
representations. This was then fed into a five- predictions may be underestimated due to the
layer auxiliary softmax classifier with batch isostructural crystal structures, this study
normalization and 118 output neurons, thus demonstrates the potential of deep learning
changing the element prediction query into a models to facilitate crystal structure prediction.

177
 Opeyemi S. Akanbi et al., Journal of Materials and Engineering Vol. 03, Iss. 2 (2025) 172-185

3.4 Relationship Between Localization and 6.79%). Moreover, another recent work [19] on
Homogenization at Multiple Scales in using deep learning for homogenization has
High-Contrast Composites shown that the filters/kernels learned by the
CNN during training can be interpreted as
Examining the connection between structure and microstructure features that the model learns to
properties at various lengths scales, an be influential for improving the macroscale
exploration of deep learning to three- property of interest. The inverse problem of
dimensional microstructure data of two-phase design exploration could be significantly aided by
composites is conducted in this section. this approach, ultimately providing feedback to
Homogenization is the process by which inform materials design.
microstructure information is transferred from
smaller scales to larger ones, such as predicting Yang et al. [20] proposed a deep-learning
macroscale properties based on microstructure technique that does not require any feature
information. Localization is when salient engineering for localization. They experimented
microstructure information is transferred from with two datasets composed of 2500 MVEs
larger scales to smaller scales, like how a load is (contrast-10) and 3000 MVEs (contrast-50) of
distributed in a material's microstructure under size 21 x 21 x 21, with periodic boundary
macroscopic loading conditions. These conditions and varying volume fractions. To
homogenization and localization connections are represent each voxel, a 3-D neighbourhood of
modelled using numerical techniques [24]. In size 11 x 11 x 11 was used, which would have
recent years, Materials Knowledge Systems been computationally expensive if 3-D CNNs
(MKS) [54] have been created to understand were used. The authors employed a 2-D CNNs
localization connections, which are non-iterative approach instead of 3-D CNNs. This entailed
series solutions using calibrated Green's treating the 3-D image of 11 x 11 x 11 as 11
function-based kernels, as well as ML-based channels of a 2-D image of 11 x 11 which was
methods which require feature engineering to perpendicular to the maximum principal strain
capture the local microstructure neighbourhood. direction. The CNN architecture featured six
layers, with two fully-connected layers and two
Yang et al. [18] proposed a feature-engineering- convolution layers. The accuracy of the deep
free deep learning-based homogenization learning model was compared to the MKS method
solution to predict macroscale effective stiffness [47], as well as single-agent [56] and multi-agent [57]

in two-phase composites with a contrast of 50. By ML-based methods. The single-agent and multi-
employing 3-D convolutional neural networks agent methods recorded a mean absolute strain
(CNNs) to map microstructure information to error (MASE) of 13.02% and 8.04%, respectively,
effective stiffness, a homogenization link was on the contrast-10 dataset. The MKS method,
established utilizing a dataset of 8550 simulated however, yielded a MASE of 10.86%. The deep
51 × 51 × 51 3-D microstructures (known as learning CNN model, though, had a clearly better
MVEs). The most efficient deep learning network MASE of 3.07%. The MKS method had a MASE of
was identified as a 14-layer model containing five 26.46% on contrast-50, which was much higher
convolution blocks (each consisting of a than the 5.71% MASE obtained by the deep
convolution layer followed by a pooling layer) learning model. An analysis of what CNN learned
and two fully connected layers, with 3.8 million showed that the impact of different level
trainable parameters. This network was used to neighbours decreases with increasing levels of
compute the effective stiffness through neighbours, which is in accordance with existing
micromechanical finite element simulations with domain knowledge.
periodic boundary conditions. The accuracy of
the deep learning model (MAE of 1.04 GPa or 3.5 Quantification and Characterization of
3.10%) was significantly better than simple Microstructure
physics-based approaches or rule of mixtures
method (MAE of 15.68 GPa or 46.66%) and Materials characterization is a process that can
sophisticated physics-inspired data science help us understand the structure of a specific
approaches that utilize principal component material [58], By using advanced techniques at
analysis (PCA) on two-point statistics as input varying lengths and times, such as different
features for regression [55] (MAE of 2.28 GPa or microscopies and spectroscopies, materials

178
 Opeyemi S. Akanbi et al., Journal of Materials and Engineering Vol. 03, Iss. 2 (2025) 172-185

imaging has become more prolific. Electron hypercolumn pixel features to the corresponding
backscatter diffraction (EBSD) is one of these target. The convolutional layers were based on
techniques that can be used to measure the three the pre-trained VGG16 network [61], and the
Euler angles of a crystalline material which are MLP layers were trained from scratch, with batch
represented by 〈φ1, Φ, φ2〉 [59]. Known as EBSD normalization, dropout, weight decay
indexing, the problem of learning orientation regularization, and data augmentation. The
angles from an EBSD pattern is an inverse segmentation model performance was tested
structure characterization. The Hough with focal loss (FL) and cross-entropy (CE)
transform-based method, which is often used classification loss on a dataset of 24 micrographs.
commercially, is effective but can be impaired by While the model was successful in accurately
noise in the pattern. An alternative method of categorizing spheroidite and particle
indexing is dictionary-based, which is a strong segmentation, it encountered difficulties when
and effective method of finding the nearest trying to recognize particles of less than 5 pixels
neighbour. The results of this search are the in radius. This indicates that higher-quality input
angles of the EBSD pattern closest to the query. is essential for training the model. Patton et al.
Unfortunately, this method is computationally [36] recently developed MENNDL, a 167-petaflop
expensive. deep learning system for automated raw electron
microscopy image-based atomic defect
Liu et al. recently tested a deep learning method identification and analysis on a supercomputer.
to index EBSD patterns, training a CNN model on An asynchronous, parallel, scalable, genetic
300,000 of the 333,227 that are simulated with algorithm combined with a support vector
EBSD patterns (60 x 60 grayscale images). machine enables the system to swiftly identify
Despite this, the performance of the CNN model the optimal network and can generate and
could have been improved had it been trained evaluate several deep neural networks (DNNs)
using the entire lexicon and optimized for the comprising various hyperparameters and
discrepancy between the predicted and actual architectures in a few hours. This deep learning
orientations. Jha et al. [22] addressed these network further facilitates the development of a
limitations in their research by training their CNN mapping of atomic transformation pathways, a
model on two dictionaries comprised of 374,852 library of imperfections, examination of local
EBSD patterns (including 1000 independent atomic environment distortions near defects, and
simulated EBSD patterns) and optimizing for the automated selection of the most favourable area
mean disorientation error by creating a in the sample for measurements or
differentiable approximation to the manipulations without human intervention,
disorientation function. This network, comprised forming the base for an independent microscope.
of 17 layers and roughly 200 million parameters,
was able to beat the dictionary approach by 16%. 3.6 Design and Reconstruction Microstructure
The mean disorientation was 0.548° for the
network, while the dictionary technique yielded Reconstructing a disordered heterogeneous
0.652°. material's structure with limited structural
information about the original system is still a
DeCost et al. [23] presented a deep learning significant issue in modelling heterogeneous
solution for the quantitative analysis of ultrahigh materials [61]. A deep transfer learning method
carbon steel microstructure, consisting of two was developed by Li et al [25] to generate
tasks: (i) segmentation of steel micrographs into statistically similar microstructures of any
four regions (grain boundary carbide, material system from only one provided
spheroidized particle matrix, particle-free grain microstructure. In their technique, the input
boundary denuded zone, and Widmanstätten microstructure with k-labelled materials is
cementite); and (ii) segmentation of cementite initially encoded into a three-channel (RGB) form
particles within the spheroidized particle matrix. to be used as an input to a pruned version of a
This was accomplished by using the PixelNet [60] pre-trained CNN known as VGG19 [62]. Also sent
architecture, which represents each pixel by the to the pruned VGG19 network is a randomly
concatenation of its representations in each initialized RGB image, which will be iteratively
convolutional layer, and then applying a changed as it is reconstructed into the encoded
Multilayer perceptron (MLP) to map the microstructure. The loss function to be

179
 Opeyemi S. Akanbi et al., Journal of Materials and Engineering Vol. 03, Iss. 2 (2025) 172-185

minimized is the difference of Gram-matrix (a microstructures and 17.2 per cent superior when
measure of the texture of a given image) [63] optimized using Bayesian optimization. The
between the activations of the original and generator, which was trained, was capable of
reconstructed microstructure, summed up over creating microstructures of any size and the
chosen convolutional layers of the pruned discriminator could be employed to initially train
VGG19. The weights of the network are fixed, thus models for predicting structure-property
requiring the optimization of the pixel values of relationships.
the microstructure reconstruction. Gradients of
the loss function are calculated through 4. PROSPECT AND FUTURE IMPACT
backpropagation and then used in a nonlinear
optimization algorithm to iteratively update the Deep learning is a rapidly expanding field of
pixel values until convergence is attained. To research which has been garnering considerable
achieve the desired result, the authors carried out attention, leading to cutting-edge algorithms being
a post-processing step with simulated annealing created at an astonishing rate. Below is a summary
after decoding the reconstruction through k- of the essential principles of deep learning as they
means clustering. This clustering divides the relate to material informatics, which is likely to
pixels into k groups, the number of which have a large influence on how data-driven
corresponds to the number of material phases in approaches are used in materials science.
the original microstructure. However, it does not
guarantee that the volume fractions will match 4.1 Other Forms of DNNs
the original microstructure. To remedy this,
simulated annealing was employed to switch the This paper has discussed three different types of
phase label of some boundary pixels in order to deep-learning neural networks: CNNs, GANs, and
make the volume fractions of the original and MLPs. However, there are also many other kinds
reconstructed microstructures equivalent. of networks that can process different forms of
data. For instance, recurrent neural networks are
Yang and Li et al. [25,26] recently developed a designed for sequence data (or temporal data) of
deep adversarial learning methodology of GANs different lengths, and are commonly used for
for a low-dimensional and non-linear embedding speech recognition [68], natural language
of microstructures for microstructural materials processing [69], and recently in materials
design. This approach was tested on various informatics [70,71]. Geometric deep learning has
structural morphologies, such as carbonate, recently been developed as a form of deep
polymer composites, sandstone, ceramics, a learning capable of dealing with non-Euclidean
block copolymer, a metallic alloy, and three- data, for example, nodes and edges in graphs,
phase rubber composites, and was found to be which cannot be handled by standard deep
more efficient than other methods (decision tree- learning kernels, such as convolution. This
based synthesis [64], Gaussian random field [65], method is applicable in quantum chemistry
two-point correlation [66], and physical [72,73], particularly in analyzing data from
descriptor [67] in four out of the five material molecular dynamics simulations.
systems. After training a GAN on 5000 synthetic
microstructure images generated with the GRF 4.2 Transfer Learning
method, it was then used to randomly sample a
latent variable vector and generate new images. Deep learning traditionally needs a great amount
The realism of the images was verified by their of data, yet transfer learning enables its use even
lineal-path correlation and two-point correlation in situations with a limited amount of data. This is
functions. To prove the power of the design done by taking the knowledge obtained from a
optimization framework, a GAN was utilized deep learning model that was built with a much
together with Bayesian optimization and larger dataset related to a different problem and
rigorous coupled wave analysis so as to simulate creating a new model with a much smaller dataset
the optical absorption performance of the for the current problem. This approach is
microstructure in question. The authors anticipated to be especially helpful for materials
observed that the GAN-generated informatics, as the datasets used are usually much
microstructures had an optical performance that smaller than those used in other areas such as
was 4.8 per cent superior to randomly sampled cosmology, bioinformatics, and social media.

180
 Opeyemi S. Akanbi et al., Journal of Materials and Engineering Vol. 03, Iss. 2 (2025) 172-185

Transfer learning is also widely used for image 4.4 Interpretability of Model
classification, where pre-trained deep learning
models, such as VGG [62] trained on the large The importance of being able to interpret models
ImageNet [74] database (containing more than 14 has been a major concern when it comes to
million labelled images), can be used to extract applications like finance, self-driving, and
important features from small datasets and build materials engineering, where a single incorrect
ML models [75]. positive result can cost dearly. Lipton [78]
proposed a comprehensive taxonomy of
4.3 Active Learning and Uncertainty interpretability criteria and methods,
Quantification distinguishing between interpretations that are
transparent to humans and those that are post-
In the future, Uncertainty Quantification (UQ) will hoc explanations. Transparency refers to the
become increasingly significant in the realm of comprehension of the model's training, inputs,
materials informatics predictive analytics. UQ parameters, and overall functioning, although it is
enables the precise determination of uncertainty typically at odds with complexity. Although some
for ML model predictions, within a specified models are not easily understandable, deep trees
confidence range. An ensemble data mining and linear models with elaborate inputs and
approach [76] is one common methodology for UQ ensembles can be less transparent than a basic
in materials informatics, where multiple neural network that has been through minimal
predictive models are constructed for the same processing. In contrast, post-hoc analyses require
task by using different techniques on the same comprehension of what the model has
data or the same technique on different subsets of determined, making deep-learning models an
data, with the final prediction being calculated ideal tool due to their extensive representations.
with the individual predictions, such as by their Visualization techniques, such as t-SNE [79] or
mean. One can assess the degree of instability by saliency maps [80], can be used to map the latent
computing the standard deviation of individual representations of a model in 2D or to identify the
forecasts, or by combining it with the error of the most influential sections of the input.
model's prediction. An alternate method is Additionally, a model's low-dimensional latent
through deep learning models, which use dropout representation can be used to identify the k-
during the training of neural networks to make the nearest neighbors in the training set and explain
model more generalizable and to simulate model the model's decisions by showing similar
ensembling [38]. At the assessment phase of the examples from the training set. [78] There are
model, certain neurons are randomly deactivated several other methods and guidelines to
via dropout, which can cause the same input to understand deep learning networks [81]. A
produce various results each time it is processed. workflow practice of using a Python
This set of predictions can then be used to programming language in modeling using neural
generate uncertainty estimates, similar to network / deep learning-based models is
ensemble learning models or more sophisticated provided by Wang et al. [13] which contains
methods [77]. At UQ, there is a strong emphasis on training, validation, and test sets, as it is
active and reinforcement learning. In these customary for a machine learning project, the
methods, an algorithm is employed to determine dataset contains 309 columns each of which is in
which unlabeled inputs should be labelled to double data type representing an attribute of the
increase the precision of the model or to decide the material sample. Some data transformations were
ideal action in a given context to acquire the done, these included scaling and transformation.
highest reward. Data-driven material discovery Then they proceeded to build the deep learning
has practical implications. By using a set of project, which adopted a dense fully connected
experimental or simulated data, it is possible to neural network, The input layer of DenseNet
predict which experiment or simulation should be accepts input data in the dimension of each row of
carried out next to upgrade the predictive model. the input data, which is equal to the number of
By using this updated system which incorporates features in the composition-based feature vectors
active learning, the search for the ideal product (CBFV) featurization scheme. When the Oliynyk
with the necessary qualities can be hastened. This featurizer was used, the input dimension was 177.
diminishes the number of experiments needed to The model’s input layer is the dimension of the
reach a conclusion. dataset which is 309, with a hidden layer of 64 by

181
 Opeyemi S. Akanbi et al., Journal of Materials and Engineering Vol. 03, Iss. 2 (2025) 172-185

32 and an output dimension of 1. A dataloader 4.6 Conclusion and Summary

class specific was defined for loading CBFV-type
datasets and a batch size of 128. For the loss Deep Materials Informatics, similar to
function and optimizers, they adopted Loss and bioinformatics of two decades ago, is still in its
Adam optimizer and the training lasted 500 infancy, specifically in the area of deep learning
epochs. After training the model was evaluated for materials science. This paper examines the
and got an R2 score of 0.9863, MAE of 3.7027, most recent developments in deep materials
RMSE of 7.0814 and for the validation set, R2 informatics, such as categorizing elemental
score of 0.9102, MAE of 13.6182, RMSE of composition, crystal structure, and two-
23.9155 were realized, as shown in figure 5. dimensional and three-dimensional
microstructure images. Additionally, the
fundamentals of deep learning, its benefits and
obstacles, the deep learning networks’ various
types, and the future of deep materials
informatics are evaluated. With the ongoing
expansion of big data and materials databases,
coupled with headways in data science and deep
learning techniques, there is immense potential
to revolutionize the prediction of materials
properties, discovery, design, and application of
advanced materials. Model evaluation as done by
Wang et al, was done on the test set and test R2 of
0.8919, test MAE of 14.4120, and test of 25.6550.
It was realized that machine learning is better for
material science because Deep learning models
can rapidly analyze vast amounts of data,
accelerating the discovery process. In contrast,
traditional lab work often involves time-
Fig. 5. Evaluation of model that predicts the heat consuming experiments and manual data
capacity of inorganic materials [13]. analysis as well as uncovering hidden patterns
and relationships in data that may not be
4.5 The Possible Effects Over Time apparent through conventional analysis
methods. This can lead to new insights and
Material advancement is essential for the growth discoveries in material science. Deep learning
of the industry, and its advancement is vital for models can also predict material properties and
the prosperity of our society. Deep materials behaviors based on data, allowing researchers to
informatics techniques have the potential to explore a wide range of possibilities without
revolutionize the field of materials science, as extensive experimentation.
well as various industries. By providing the
means to explore near-endless possibilities in
materials, researchers can more easily identify REFERENCES
the best options and then test them through
experiments and simulations, ultimately [1] Executive Office of the President, "Materials
accelerating the process of acquiring and using Genome Initiative for Global Competitiveness,"
advanced materials while also reducing the cost 2011.
of materials production. The Materials Genome [2] A. Agrawal and A. Choudhary, "Perspective:
Initiative (MGI) strives to identify and create Materials informatics and big data: Realization of
innovative materials in an economical manner. the ‘fourth paradigm’ of science in materials
Utilizing deep materials informatics could science," APL Mater., vol. 4, no. 5, p. 053208,
accelerate this process, thereby providing 2016; [Chem. Modell.], vol. 17, pp. 91-152, 2022.
substantial advantages to businesses in [3] G. B. Olson, "Computational design of
numerous industries, for instance, automobiles, hierarchically structured materials," Science, vol.
healthcare, transportation, construction, clean 277, pp. 1237–1242, 1997.
energy, and aerospace.

182
 Opeyemi S. Akanbi et al., Journal of Materials and Engineering Vol. 03, Iss. 2 (2025) 172-185

[4] P. Hohenberg and W. Kohn, "Inhomogeneous [19] A. Cecen, H. Dai, Y. C. Yabansu, S. R. Kalidindi, and
electron gas," Phys. Rev., vol. 136, no. 3B, pp. L. Song, "Material structure-property linkages
B864-B871, 1964. using three-dimensional convolutional neural
networks," Acta Mater., vol. 146, pp. 76-84, 2018.
[5] W. Kohn and L. J. Sham, "Self-consistent
equations including exchange and correlation [20] Z. Yang et al., "Establishing structure-property
effects," Phys. Rev., vol. 140, no. 4A, pp. A1133- localization linkages for elastic deformation of
A1138, 1965. three-dimensional high contrast composites
using deep learning approaches," Acta Mater.,
[6] G. B. Olson, "Designing a new material world,"
vol. 166, pp. 335-345, 2019.
Science, vol. 288, no. 5468, pp. 993-998, 2000.
[21] R. Liu, A. Agrawal, W. K. Liao, A. Choudhary, and
[7] A. R. Oganov, Ed., Modern Methods of Crystal
M. De Graef, "Materials discovery: Understanding
Structure Prediction. John Wiley & Sons, 2011.
polycrystals from large-scale electron patterns,"
[8] A. R. Oganov and C. W. Glass, "Crystal structure in 2016 IEEE International Conference on Big
prediction using ab initio evolutionary Data (Big Data), 2016, pp. 2261-2269.
techniques: Principles and applications," J. Chem.
[22] D. Jha, S. Singh, R. Al-Bahrani, W. K. Liao, A.
Phys., vol. 124, no. 24, p. 244704, 2006.
Choudhary, M. De Graef, and A. Agrawal,
[9] R. E. Newnham, Properties of Materials: "Extracting grain orientations from EBSD
Anisotropy, Symmetry, Structure. Oxford patterns of polycrystalline materials using
University Press, 2005. convolutional neural networks," Microsc.
Microanal., vol. 24, no. 5, pp. 497-502, 2018.
[10] S. Curtarolo, G. L. Hart, M. B. Nardelli, N. Mingo, S.
Sanvito, and O. Levy, "The high-throughput [23] B. L. DeCost, B. Lei, T. Francis, and E. A. Holm,
highway to computational materials design," "High throughput quantitative metallography for
Nat. Mater., vol. 12, no. 3, pp. 191-201, 2013. complex microstructures using deep learning: A
case study in ultrahigh carbon steel," Microsc.
[11] M. L. Green et al., "Fulfilling the promise of the
Microanal., vol. 25, no. 1, pp. 21-29, 2019.
materials genome initiative with high-
throughput experimental methodologies," Appl. [24] A. Agrawal and A. Choudhary, "Deep materials
Phys. Rev., vol. 4, no. 1, p. 011105, 2017. informatics: Applications of deep learning in
materials science," MRS Commun., vol. 9, no. 3,
[12] H. Koinuma and I. Takeuchi, "Combinatorial
pp. 779-792, 2019.
solid-state chemistry of inorganic materials,"
Nat. Mater., vol. 3, no. 7, pp. 429-438, 2004. [25] X. Li, Y. Zhang, H. Zhao, C. Burkhart, L. C. Brinson,
and W. Chen, "A transfer learning approach for
[13] A. Y. T. Wang et al., "Machine learning for
microstructure reconstruction and structure-
materials scientists: An introductory guide
property predictions," Sci. Rep., vol. 8, no. 1, p.
toward best practices," Chem. Mater., vol. 32, no.
13461, 2018.
12, pp. 4954-4965, 2020.
[26] Z. Yang, X. Li, L. C. Brinson, A. N. Choudhary, W.
[14] D. Jha et al., "Elemnet: Deep learning the
Chen, and A. Agrawal, "Microstructural materials
chemistry of materials from only elemental
design via deep adversarial learning methodology,"
composition," Sci. Rep., vol. 8, no. 1, p. 1-13, 2018.
J. Mech. Des., vol. 140, no. 11, 2018.
[15] T. Xie and J. C. Grossman, "Crystal graph
[27] D. E. Rumelhart, G. E. Hinton, and R. J. Williams,
convolutional neural networks for an accurate
"Learning representations by back-propagating
and interpretable prediction of material
errors," Nature, vol. 323, no. 6088, pp. 533-536,
properties," Phys. Rev. Lett., vol. 120, no. 14, p.
1986.
145301, 2018.
[28] D. E. Rumelhart, G. E. Hinton, and R. J. Williams,
[16] W. Ye, C. Chen, Z. Wang, I. H. Chu, and S. P. Ong,
Learning internal representations by error
"Deep neural networks for accurate predictions
propagation, Parallel Distributed Processing:
of crystal stability," Nat. Commun., vol. 9, no. 1, p.
Explorations in the Microstructure of Cognition,
3800, 2018.
vol. 1, Cambridge, MA, USA: MIT Press, 1986.
[17] K. Ryan, J. Lengyel, and M. Shatruk, "Crystal
[29] K. Hornik, "Approximation capabilities of
structure prediction via deep learning," J. Am. Chem.
multilayer feedforward networks," Neural Netw.,
Soc., vol. 140, no. 32, pp. 10158-10168, 2018.
vol. 4, no. 2, pp. 251-257, 1991.
[18] Z. Yang et al., "Deep learning approaches for
[30] A. Krizhevsky, I. Sutskever, and G. E. Hinton,
mining structure-property linkages in high
"Imagenet classification with deep convolutional
contrast composites from simulation datasets,"
neural networks," Commun. ACM, vol. 60, no. 6,
Comput. Mater. Sci., vol. 151, pp. 278-287, 2018.
pp. 84-90, 2017.

183
 Opeyemi S. Akanbi et al., Journal of Materials and Engineering Vol. 03, Iss. 2 (2025) 172-185

[31] K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual [44] H. Zhang et al., "StackGAN: Text to photo-realistic
learning for image recognition," in Proc. IEEE image synthesis with stacked generative
Conf. Comput. Vis. Pattern Recognit., 2016, pp. adversarial networks," in Proc. IEEE Int. Conf.
770-778. Comput. Vis., 2017, pp. 5907-5915.
[32] G. Hinton et al., "Deep neural networks for [45] G. Perarnau, J. Van De Weijer, B. Raducanu, and J.
acoustic modelling in speech recognition: The M. Álvarez, "Invertible conditional GANs for
shared views of four research groups," IEEE image editing," arXiv preprint arXiv:1611.06355,
Signal Process. Mag., vol. 29, no. 6, pp. 82-97, 2016.
2012.
[46] H. Wu, S. Zheng, J. Zhang, and K. Huang, "GP-GAN:
[33] L. Deng, G. Hinton, and B. Kingsbury, "New types Towards realistic high-resolution image
of deep neural network learning for speech blending," in Proc. 27th ACM Int. Conf.
recognition and related applications: An Multimedia, 2019, pp. 2487-2495.
overview," in 2013 IEEE Int. Conf. Acoust. Speech
[47] D. Pathak, P. Krahenbuhl, J. Donahue, T. Darrell,
Signal Process., 2013, pp. 8599-8603.
and A. A. Efros, "Context encoders: Feature
[34] S. Lee, D. Jha, A. Agrawal, A. Choudhary, and W. K. learning by inpainting," in Proc. IEEE Conf.
Liao, "Parallel deep convolutional neural Comput. Vis. Pattern Recognit., 2016, pp. 2536-
network training by exploiting the overlapping 2544.
of computation and communication," in 2017
[48] L. C. Yang, S. Y. Chou, and Y. H. Yang, "MidiNet: A
IEEE 24th Int. Conf. High-Perform. Comput.
convolutional generative adversarial network
(HiPC), 2017, pp. 183-192.
for symbolic-domain music generation," arXiv
[35] V. Mnih et al., "Asynchronous methods for deep preprint arXiv:1703.10847, 2017.
reinforcement learning," in Int. Conf. Mach.
[49] K. T. Schütt et al., "How to represent crystal
Learn., 2016, pp. 1928-1937.
structures for machine learning: Towards fast
[36] R. M. Patton et al., "167-pflops deep learning for prediction of electronic properties," Phys. Rev. B,
electron microscopy: From learning physics to vol. 89, no. 20, p. 205118, 2014.
atomic manipulation," in SC18: Int. Conf. High-
[50] F. Faber, A. Lindmaa, O. A. Von Lilienfeld, and R.
Perform. Comput. Netw. Storage Anal., 2018, pp.
Armiento, "Crystal structure representations for
638-648.
machine learning models of formation energies,"
[37] Y. Lou, R. Caruana, and J. Gehrke, "Intelligible Int. J. Quantum Chem., vol. 115, no. 16, pp. 1094-
models for classification and regression," in Proc. 1101, 2015.
18th ACM SIGKDD Int. Conf. Knowl. Discov. Data
[51] A. Seko, H. Hayashi, K. Nakayama, A. Takahashi,
Min., 2012, pp. 150-158.
and I. Tanaka, "Representation of compounds for
[38] N. Srivastava, G. Hinton, A. Krizhevsky, I. machine-learning prediction of physical
Sutskever, and R. Salakhutdinov, "Dropout: A properties," Phys. Rev. B, vol. 95, no. 14, p.
simple way to prevent neural networks from 144110, 2017.
overfitting," J. Mach. Learn. Res., vol. 15, no. 1, pp.
[52] L. Ward et al., "Including crystal structure
1929-1958, 2014.
attributes in machine learning models of
[39] T. N. Kipf and M. Welling, "Semi-supervised formation energies via Voronoi tessellations,"
classification with graph convolutional Phys. Rev. B, vol. 96, no. 2, p. 024104, 2017.
networks," arXiv preprint arXiv:1609.02907,
[53] A. Jain et al., "Commentary: The Materials
2016.
Project: A materials genome approach to
[40] J. Zhou et al., "Graph neural networks: A review accelerating materials innovation," APL Mater.,
of methods and applications," AI Open, vol. 1, pp. vol. 1, no. 1, p. 011002, 2013.
57-81, 2020.
[54] D. B. Brough, D. Wheeler, and S. R. Kalidindi,
[41] I. Goodfellow et al., "Generative adversarial "Materials knowledge systems in Python—a data
networks," Commun. ACM, vol. 63, no. 11, pp. science framework for accelerated development
139-144, 2020. of hierarchical materials," Integr. Mater. Manuf.
Innov., vol. 6, pp. 36-53, 2017.
[42] M. J. Osborne and A. Rubinstein, A Course in Game
Theory. Cambridge, MA, USA: MIT Press, 1994. [55] A. Gupta, A. Cecen, S. Goyal, A. K. Singh, and S. R.
Kalidindi, "Structure–property linkages using a
[43] T. C. Wang et al., "High-resolution image
data science approach: Application to a non-
synthesis and semantic manipulation with
metallic inclusion/steel composite system," Acta
conditional GANs," in Proc. IEEE Conf. Comput.
Mater., vol. 91, pp. 239-254, 2015.
Vis. Pattern Recognit., 2018, pp. 8798-8807.

184
 Opeyemi S. Akanbi et al., Journal of Materials and Engineering Vol. 03, Iss. 2 (2025) 172-185

[56] R. Liu, Y. C. Yabansu, A. Agrawal, S. R. Kalidindi, [69] Y. Xie et al., "Reducing infrequent-token
and A. N. Choudhary, "Machine learning perplexity via variational corpora," in Proc. 53rd
approaches for elastic localization linkages in Annu. Meet. Assoc. Comput. Linguist. 7th Int. Joint
high-contrast composite materials," Integr. Conf. Natural Language Process., vol. 2, 2015, pp.
Mater. Manuf. Innov., vol. 4, pp. 192-208, 2015. 609-615.
[57] R. Liu et al., "Context aware machine learning [70] M. Mozaffar et al., "Data-driven prediction of the
approaches for modeling elastic localization in high-dimensional thermal history in directed
three-dimensional composite microstructures," energy deposition processes via recurrent neural
Integr. Mater. Manuf. Innov., vol. 6, pp. 160-171, networks," Manuf. Lett., vol. 18, pp. 35-39, 2018.
2017.
[71] A. Paul et al., "Chemixnet: Mixed DNN
[58] Y. Leng, Materials Characterization: Introduction architectures for predicting chemical properties
to Microscopic and Spectroscopic Methods. using multiple molecular representations," arXiv
Hoboken, NJ, USA: John Wiley & Sons, 2009. preprint arXiv:1811.08283, 2018.
[59] A. J. Schwartz, M. Kumar, B. L. Adams, and D. P. [72] J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals,
Field, Eds., Electron Backscatter Diffraction in and G. E. Dahl, "Neural message passing for
Materials Science, 2nd ed. New York, NY, USA: quantum chemistry," in Proc. Int. Conf. Mach.
Springer, 2009. Learn., 2017, pp. 1263-1272.
[60] A. Bansal, X. Chen, B. Russell, A. Gupta, and D. [73] K. T. Schütt, H. E. Sauceda, P. J. Kindermans, A.
Ramanan, "PixelNet: Representation of the Tkatchenko, and K. R. Müller, "SchNet–a deep
pixels, by the pixels, and for the pixels," arXiv learning architecture for molecules and
preprint arXiv:1702.06506, 2017. materials," J. Chem. Phys., vol. 148, no. 24, p.
241722, 2018.
[61] R. Piasecki, "Microstructure reconstruction using
entropic descriptors," Proc. R. Soc. A Math. Phys. [74] J. Deng et al., "Proceedings of the IEEE
Eng. Sci., vol. 467, no. 2127, pp. 806-820, 2011. Conference on Computer Vision and Pattern
Recognition," IEEE Conf. Comput. Vis. Pattern
[62] K. Simonyan and A. Zisserman, "Very deep
Recognit., 2009.
convolutional networks for large-scale image
recognition," arXiv preprint arXiv:1409.1556, 2014. [75] K. Gopalakrishnan, S. K. Khaitan, A. Choudhary,
and A. Agrawal, "Deep convolutional neural
[63] L. Gatys, A. S. Ecker, and M. Bethge, "Texture
networks with transfer learning for computer
synthesis using convolutional neural networks,"
vision-based data-driven pavement distress
Adv. Neural Inf. Process. Syst., vol. 28, 2015.
detection," Constr. Build. Mater., vol. 157, pp.
[64] R. Bostanabad, A. T. Bui, W. Xie, D. W. Apley, and 322-330, 2017.
W. Chen, "Stochastic microstructure
[76] A. Agrawal and A. Choudhary, "An online tool for
characterization and reconstruction via
predicting fatigue strength of steel alloys based
supervised learning," Acta Mater., vol. 103, pp.
on ensemble data mining," Int. J. Fatigue, vol. 113,
89-102, 2016.
pp. 389-400, 2018.
[65] M. Grigoriu, "Random field models for two-phase
[77] Y. Gal, J. Hron, and A. Kendall, "Concrete dropout,"
microstructures," J. Appl. Phys., vol. 94, no. 6, pp.
Adv. Neural Inf. Process. Syst., vol. 30, 2017.
3762-3770, 2003.
[78] Z. C. Lipton, "The mythos of model
[66] Y. Liu, M. S. Greene, W. Chen, D. A. Dikin, and W.
interpretability: In machine learning, the concept
K. Liu, "Computational microstructure
of interpretability is both important and
characterization and reconstruction for
slippery," Queue, vol. 16, no. 3, pp. 31-57, 2018.
stochastic multiscale material design," Comput.-
Aided Des., vol. 45, no. 1, pp. 65-76, 2013. [79] L. Van der Maaten and G. Hinton, "Visualizing data
using t-SNE," J. Mach. Learn. Res., vol. 9, no. 11, 2008.
[67] H. Xu, D. A. Dikin, C. Burkhart, and W. Chen,
"Descriptor-based methodology for statistical [80] K. Simonyan, A. Vedaldi, and A. Zisserman, "Deep
characterization and 3D reconstruction of inside convolutional networks: Visualising image
microstructural materials," Comput. Mater. Sci., classification models and saliency maps," arXiv
vol. 85, pp. 206-216, 2014. preprint arXiv:1312.6034, 2013.
[68] Y. Miao, M. Gowayyed, and F. Metze, "EESEN: [81] G. Montavon, W. Samek, and K. R. Müller,
End-to-end speech recognition using deep RNN "Methods for interpreting and understanding
models and WFST-based decoding," in Proc. IEEE deep neural networks," Digit. Signal Process., vol.
Workshop Autom. Speech Recognit. Understand., 73, pp. 1-15, 2018.
2015, pp. 167-174.

185

View publication stats