MRS Communications (2019), 1 of 11

© Materials Research Society, 2019

doi:10.1557/mrc.2019.32

Artificial Intelligence Prospective

Machine learning for composite materials

Chun-Teh Chen, Department of Materials Science and Engineering, University of California, Berkeley, CA 94720, USA
Grace X. Gu, Department of Mechanical Engineering, University of California, Berkeley, CA 94720, USA
Address all correspondence to Grace X. Gu at [email protected]

(Received 3 February 2019; accepted 28 February 2019)

Abstract
Machine learning (ML) has been perceived as a promising tool for the design and discovery of novel materials for a broad range of applica-
tions. In this prospective paper, we summarize recent progress in the applications of ML to composite materials modeling and design. An
overview of how different types of ML algorithms can be applied to accelerate composite research is presented. This framework is envisioned
to revolutionize approaches to design and optimize composites for the next generation of materials with unprecedented properties.

Introduction is an inverse design problem. For a forward modeling problem,

Designing novel materials with superior tailored properties is the structure (e.g., atomic constituents, crystal structure, and
the ultimate goal of modern engineering applications.[1–5] In topology) of the material to be investigated is usually given
the past few decades, with rapid advances in high-performance and the properties are governed by physical laws such as
parallel computing, materials science, and numerical modeling, quantum mechanics, thermodynamics, and solid mechanics.
many essential properties of materials can now be calculated Consequently, various types of properties at different length-
using simulations with reasonable accuracy. For example, the and time-scales can be calculated numerically by solving the
chemical reactivity and stability of molecules can be estimated corresponding governing equations using proper physics-based
using density functional theory (DFT).[6,7] Molecular dynamics modeling tools such as DFT, MD, and FEM. However, there is
(MD) and the finite element method (FEM) can be applied to no physics-based modeling tool that can solve inverse design
simulate a wide range of mechanical behaviors of materials problems—that is, to generate the structure of a material with
at the nano-scale and continuum-scale, respectively.[8,9] a given set of required properties. In practice, one common
Nowadays, simulations of material properties can be performed approach to solve inverse design problems is using domain
on a laptop, workstation, or computer cluster, depending on the knowledge and experience (intuition) to narrow down the
computational cost. In general, performing simulations to pre- design space and propose new materials by trial and error, as
dict the properties of a material is much faster and less expen- shown in the flow chart of Fig. 1. For example, many biomate-
sive than synthesizing, manufacturing, and testing the material rials such as nacre and bone have remarkable mechanical prop-
in a laboratory. Moreover, simulations offer very precise con- erties despite being composed of relatively weak components.
trol over environments and offer more detailed information of Consequently, it is plausible to study the architectures of these
material behavior and associated mechanisms under different biomaterials and learn the structure–property relationships nec-
conditions, many of which cannot or would be very difficult essary for achieving improved properties of materials. Thus, the
to be observed using experiments. For instance, the stress knowledge gained can be applied to design new materials in a
field of a composite material under fracture and the motions case-by-case manner. This approach is referred to as biomimi-
of each molecule in an organic material under loading can be cry and pursuant designs are termed bioinspired.[8,19–25] If the
predicated in simulations but are difficult to measure in exper- design space can reasonably be parameterized into composi-
iments. It is this reason that many studies in the literature have tional and configurational degrees of freedom, it is possible
focused on advancing computational tools and methods to to solve inverse design problems by using a physics-based
model various types of materials.[10–18] modeling tool with a brute-force exhaustive search approach
Compared with solely predicting properties of known mate- to explore the entire design space. This approach is referred
rials, designing new materials to achieve tunable properties is a to as high-throughput computational screening and is explored
more important problem for scientific and engineering pur- for various materials. For example, Emery et al. used high-
poses. In fact, predicting materials’ properties and designing throughput DFT to screen ABO3 perovskites based on thermo-
materials are quite different problems, in which the former is dynamic considerations.[26] Chen et al. proposed the most sta-
often referred to as a forward modeling problem and the latter ble molecular structures for eumelanin and polydopamine by

MRS COMMUNICATIONS
Downloaded from https://www.cambridge.org/core. University of New England, on 30 Mar 2019 at 16:31:18, subject to the Cambridge Core terms • www.mrs.org/mrc ▪1
of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1557/mrc.2019.32
 Figure 1. Comparison of materials’ design approaches based on domain knowledge and ML. The flowchart shown in blue represents the
domain-knowledge-based materials’ design approach and the flowchart shown in green represents the machine-learning-based materials’ design approach.

using high-throughput DFT to screen thousands of probable points. Recently, ML has turned our daily life to be more con-
candidates.[6,7] However, when the design space is vast, using venient in numerous ways by influencing image recognition,
a brute-force approach is computationally infeasible even autonomous driving, e-mail spam detection, among oth-
with modern algorithms and supercomputers. Therefore, opti- ers.[31–38]
mization methods such as greedy algorithms and gradient- In the field of materials science, most ML applications are
based algorithms are typically implemented in inverse design concentrated on discovering new chemical compounds or mol-
problems to search for optimal designs without having to ecules with desired properties. Those studies can be categorized
explore the entire design space. Despite the high computational into a subfield of ML for materials chemistry. One of the most
efficacy of optimization methods, the optimal solutions often challenging and active research topics in this subfield is finding
depend on the initial configuration (initial values of design var- a suitable representation (e.g., descriptors and fingerprints) of
iables) adopted in the optimization process. Therefore, the solu- a molecule or crystal structure to be used as input variables
tions obtained from those optimization methods not only vary in ML models.[39,40] This process is referred to as feature engi-
from one initial configuration to another but also, in some neering (Fig. 1). Such representation can be coarse-level
cases, can be stuck in local minima or critical points. chemo-structural descriptors or something containing informa-
Consequently, it is crucial to investigate alternative methods tion of molecular electronic charge density.[34] In this area,
to make the inverse design of materials possible. Dieb et al. applied ML models to search for the most stable
In addition to the old-fashioned materials’ design structures of doped boron atoms in graphene[41]. Hansen
approaches mentioned above, data-driven approaches based et al. applied a number of ML techniques to predict ground-
on machine learning (ML) techniques may transform the state atomization energies of small molecules.[32] Botu et al.
approaches of materials’ design in the future as shown in showed that energies and atomic forces may be predicted
Fig. 1. ML, a branch of artificial intelligence, uses a variety with chemical accuracy using an ML algorithm.[28] In the
of statistical and probabilistic methods that allow computers work of Mannodi-Kanakkithodi et al., the authors used ML
to learn from experience and detect hidden patterns (the corre- methods to design polymer dielectrics.[33] Meredig et al.
lations between input and output variables) from large and showed that the thermodynamic stability of compounds can
oftentimes noisy datasets.[27–30] Through evaluating only a por- be predicted using an ML model.[42] Pilania et al. used ML
tion of the possible data, ML algorithms can detect hidden pat- models trained on quantum mechanical calculations and com-
terns in the data and learn a target function that best maps input pared properties such as atomization energy, lattice parameter,
variables to an output variable (or output variables)—a proce- and electron affinity with their model.[34] As with other appli-
dure referred to as the training process. With the extracted pat- cations, applying ML to discover new materials requires a
terns, predictions for unseen data points can be made and allow large amount of training data, which can either be attained com-
for generalizations with a limited amount of data rather than putationally or experimentally. In addition to generating
using an exhaustive approach to explore all possible data required training data using physics-based modeling tools,

2▪ MRS COMMUNICATIONS • www.mrs.org/mrc

Downloaded from https://www.cambridge.org/core. University of New England, on 30 Mar 2019 at 16:31:18, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1557/mrc.2019.32
 Artificial Intelligence Prospective

numerous materials’ databases provide universal access to all supervised learning algorithms, linear regression is the
abundant materials’ data. Some of the most comprehensive most basic one and has been studied and used extensively.
materials’ databases include the Materials Project, Automatic Unlike other more complex (deep) ML models (e.g., NN and
Flow for Materials Discovery, Open Quantum Materials CNN), which are often being considered as “black box” models
Database, and Novel Materials Discovery. De Jong et al. pre- because of their complexity, a linear (regression) model has
dicted bulk and shear moduli for polycrystalline inorganic com- high interpretability of the relationship between input and out-
pounds using training data from the Materials Project.[43] Ward put variables. The hypothesis of linear regression is:
et al. used training data from the Open Quantum Materials
Database to discover new potential crystalline compounds for y = wT x (1)
photovoltaic applications.[44]
There are many instructive and inspiring review papers in where w represents the learnable parameters or weights (includ-
the literature on applying ML to accelerate the discovery of ing a bias w0), x represents the input variables (including a
new compounds.[31,34,45–48] In this prospective paper, we constant x0 for the bias term), and y represents the output var-
focus instead on another type of material—composite materi- iable (dependent) of the model. One common technique to esti-
als. Composites, composed of two or more base materials, mate the loss (error) in a regression model is using the mean
are commonly used as structural materials.[2,4,49,50] The base squared error (MSE), which is used to quantify how close or
materials oftentimes have vastly distinctive properties and the far the predictions are from the actual quantities. The MSE is
combined architecture built from the base materials allows computed as follows:
composites to possess unprecedented properties. Despite the
vast design space of composites, traditional manufacturing 1
N
E= (yn − ŷn )2 (2)
methods have limited composite architectures to mostly lami- N n=1
nate structures. In the past, more complex composites with hier-
archical and porous features were difficult to be realized. where N is the number of data points (samples) used to calcu-
Recent advances in additive manufacturing, however, have late the MSE, yn is the prediction of sample n (from ML mod-
opened up the design space of composites and allowed for els), and ŷn is the actual quantity of the sample (ground truth).
the creation of complex materials with internal voids and mul- Once the error function of a model is defined, the weights of the
tiple materials.[51–60] With this newfound freedom when it model can be calculated by an optimization algorithm such as
comes to the realization of complex shapes, optimization of the classical stochastic gradient descent or the Adam optimiza-
structures become essential to achieve superior materials.[61–65] tion algorithm. Linear regression has been widely applied to
Consequently, finding a rational way to select optimal architec- various research problems because of its simplicity and high
tures would be crucial when it comes to realizing the full interpretability. For the applications in composite materials,
potential of composite materials in engineering applications. Tiryaki et al. showed that linear regression can be applied to
Recently, ML has been perceived as a promising tool for the predict the compressive strength of heat-treated wood based
design and discovery of new composite materials.[35,66–70] on experimental data (Table I).[71] In the authors’ work, the
In this prospective paper, we discuss recent progress in the input variables include wood species, heat treatment tempera-
applications of ML to composite design. Here, the focus will be ture, and exposure time. The output variable is the compressive
on the applications of using ML models to predict mechanical strength of heat-treated woods. Khademi et al. and Young et al.
properties (e.g., toughness, strength, and stiffness) of compos- applied linear regression to predict concrete compressive
ites as well as applying ML models to design composites with strength, in which the input variables include experimentally
desired properties. We review some basic ML algorithms measured concrete characteristics and mixture proportions
including linear regression, logistic regression, neural networks (Table I).[72,73] In those studies, in addition to the linear models,
(NN), convolutional neural networks (CNN), and Gaussian the authors also performed more complex ML models such as
process (GP) in the context of materials design. Recent studies NN models for their regression tasks. The authors showed that
(both experimental and computational) on applying those ML more complex ML models, in general, offered more accurate
algorithms to composite research (including nanocomposites) predictions. However, in those studies, the linear models pro-
is also discussed and highlighted in Table I. Lastly, we con- vided valuable information like which input variables (e.g.,
clude with a summary and future prospects in this rapidly grow- treatment conditions and material features) were more impor-
ing research field. tant (with higher influences) to the prediction (e.g., compres-
sive strength). Note that linear regression assumes a linear
Linear regression relationship between input variables and the prediction. Most
Numerous ML algorithms have been developed for different problems have different degrees of nonlinear characteristics
types of learning purposes such as supervised learning, semi- and linear regression is not appropriate for highly nonlinear
supervised learning, and unsupervised learning. Supervised problems in which the relationship between the input and out-
learning (i.e., predictive modeling) is the most widely used put variables cannot be approximated by a linear function.
learning approach in scientific and engineering fields. Among However, it is always a good idea to try linear regression (or

MRS COMMUNICATIONS
Downloaded from https://www.cambridge.org/core. University of New England, on 30 Mar 2019 at 16:31:18, subject to the Cambridge Core terms • www.mrs.org/mrc ▪3
of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1557/mrc.2019.32
 Table I. Recent studies on applying ML algorithms to composite research.

Material Type ML models Input Output Samples

[71]
Heat-treated woods Exp. Linear, NN Treatment conditions Compressive strength 48

Concrete[72] Exp. Linear, NN, ANFIS Mixture proportions Compressive strength 173

Concrete[73] Exp. linear, NN, SVM, Mixture proportions Compressive strength 10,000
other

Composites[74] FEM Linear, CNN Topology of base materials Toughness, strength 1,000,000

Cemented paste Exp. NN Mixture proportions, Compressive strength 396

backfill[75] treatment conditions

Composites[70] FEM CNN Topology of base materials Stiffness 8,550

Hierarchical FEM CNN Topology of unit cells Toughness, strength 100,000

composites[66]

Graphene kirigami[77] MD NN, CNN Topology of unit cells Yield stress, yield 29,791
strain

Stainless steel[80] Exp. GP Processing parameters Porosity 82

[81]
Assembled hairy NPs CG-MD GP Mixture proportions Modulus, toughness 100

other simple ML algorithms) first to see how difficult the prob- Softmax regression, a generalization of logistic regression, is
lem is before applying other more complex ML algorithms. often applied, which extends logistic regression to multi-class
classification.
Using logistic regression algorithms, Gu et al. presented
Logistic regression
pioneer work on applying ML to the design of composites
Unlike linear regression which is used for the prediction of
for Mode I fracture (Table I).[74] Unlike the applications of
continuous quantities, logistic regression is mostly used for
using ML to predict the compressive strength of woods and
the prediction of discrete class labels. Logistic regression
concrete, in which experimental measurements were used as
predicts the probabilities for classification problems with two
the training data,[71–73] Gu et al. performed finite element anal-
possible categories. A logistic function is used in logistic
ysis to generate the training data for their composite design
regression to force the output variable to be between 0 and
problems. Figure 2 shows the ML approach using logistic
1. Thus, the output variable can be used to represent the prob-
regression (i.e., linear model) for an 8 by 8 composite system
ability for a sample being in a category:
in the authors’ work. The input variables represent the topology
ew x
T of composites, where two base materials (stiff and soft) were
h= (3) considered and denoted as 0 and 1. As can be seen in the figure,
1 + ewT x
instead of using the actual performance (i.e., toughness and
where w represents the weights and x represents the input var- strength) of composites to train a regression model, the training
iables of the model. Logistic regression is a generalized linear samples were categorized into two classes, namely “good”
model since the predicted probabilities only depend on the design and “bad” design, and logistic regression was performed
weighted sum of the input variables. As only two categories for the classification task. After the training process, the linear
are considered in logistic regression, 0.5 is used as the classifi- model not only could distinguish whether an unseen composite
cation threshold. For logistic regression, the cross-entropy error design was a good or bad design but also could estimate how
is often used to estimate the error in the model: good (or how bad) the composite design was by comparing
the probabilities of being in each class (Fig. 2). This study dem-
1
N
onstrated that ML could be applied to learn structure–property
ln(1 + e−ŷn w xn )
T
E= (4)
N n=1 relationships of materials even when the performance of mate-
rials could not be accurately measured. This ML approach
where N is the number of samples used to calculate the cross- could be applied to other material modeling and design prob-
entropy error, xn represents the input variables of sample n, lems in which the performance of materials is difficult (or
and ŷn is the actual category of the sample (either 1 or −1). expensive) to measure in experiments or simulations. The
As mentioned above, logistic regression can only perform authors showed that using ML to predict mechanical properties
binary classification tasks. For multi-class classification tasks, of composites is orders of magnitude faster than conventional

4▪ MRS COMMUNICATIONS • www.mrs.org/mrc

Downloaded from https://www.cambridge.org/core. University of New England, on 30 Mar 2019 at 16:31:18, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1557/mrc.2019.32
 Artificial Intelligence Prospective

Figure 2. Overall flow chart. The flow chart shows the ML approach using the linear model for an 8 by 8 system. The ML approach using the CNN model is
similar to this flow chart but without the step of converting to 1-D arrays. Note that for a 16 by 16 system, the amount of input data, training data, and testing data
are 1 million, 0.9 million, and 0.1 million, respectively. Reprinted with permission from Ref. 74. Copyright 2017 Elsevier.

finite element analysis. Moreover, although it was shown in the (the layers except for the input and output layers) and each layer
authors’ work that the prediction accuracy could be improved comprises a number of neurons. Nonlinear functions such as
when using a more complex ML model such as CNN, the linear rectified linear unit and sigmoid are used as activation functions
model provided high interpretability of the relationship to introduce nonlinearity. The hidden layers in a NN are called
between input and output variables. Thus, after the linear the fully-connected layers as all the neurons in adjacent layers
model is trained, the optimal designs (the designs with the high- are fully connected. The output of the first layer goes into the
est toughness or strength) of composites could be generated second layer as input, and so on. On the basis of the universal
directly by using the information of the weights in the linear approximation theorem, a NN with one hidden layer containing
model, without requiring any sampling, optimizing, or search- a finite number of neurons can approximate any continuous
ing process. In the authors’ work, the linear model generated functions. In addition to NN, CNN is another widely used
optimal designs of composites with the toughness and strength NN architecture, which is commonly applied to analyzing
orders of magnitude higher than the training samples, and images and videos. The main building block in the CNN archi-
required a much less computational cost compared with tecture is the convolution layer. The convolution layer com-
exhaustive methods. prises a number of convolution filters. Each filter is
convolved with the input from the previous layer and generates
Neural networks feature maps, which are fed into the next layer as input.
Linear models assume that the relationship between input and Compared with the fully-connected layers, this convolution
output variables is linear. Thus, the learning capacity of linear operator can significantly reduce the number of weights needed
models is limited. Although it is possible to increase the learn- in an ML model and makes the computation more efficient. The
ing capacity by adding a few nonlinear terms (e.g., polynomi- purpose of doing convolution is to extract relevant features
als) to the hypothesis, it is not practical as there is an infinite from data. This feature learning ability is essential for image-
number of nonlinear functions that can be chosen from and related problems. In practice, CNN models usually outperform
in general little is known which nonlinear functions are more NN models on image classification tasks and many others.
suitable for our problems. An alternative approach to achieve NN and CNN have been applied to a few composite studies
high learning capacity is to use linear regression with an activa- mostly for regression tasks. Qi et al. applied a NN model to pre-
tion function to make an artificial neuron (perceptron) and con- dict the unconfined compressive strength of cemented paste
nect a number of neurons to make an artificial NN. The most backfill (Table I).[75] In the authors’ work, the training data
basic architecture of NNs is multilayer perceptron, which is were collected from experiments. The input variables include
also denoted to as NN here. It consists of multiple hidden layers the tailing type, cement-tailings ratio, solids content, and curing

MRS COMMUNICATIONS
Downloaded from https://www.cambridge.org/core. University of New England, on 30 Mar 2019 at 16:31:18, subject to the Cambridge Core terms • www.mrs.org/mrc ▪5
of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1557/mrc.2019.32
 time while the output variable is the unconfined compressive can be learned by ML directly based on training data, without
strength of specimens. The authors applied particle swarm opti- the need for conventional homogenization and localization pro-
mization to tune the architecture (the numbers of hidden layers cesses. For this hierarchical composite problem, CNN was
and neurons) of their NN model and showed that their NN implemented to perform a regression task since linear regres-
model can accurately predict and calculate the unconfined com- sion was incapable of capturing the highly nonlinear correla-
pressive strength of cemented paste backfill. For composite tions between the input and output variables. However,
modeling and design problems, the topology of a composite unlike linear models, CNN models (and most ML models) can-
can naturally be represented as an image (2-D or 3-D). The not generate optimal designs directly based on the values of
unit cells or base materials of a composite are converted to pix- weights. To overcome this limitation, Gu et al. further aug-
els (with different values) of the corresponding image. mented CNN models to discover high-performing hierarchical
Accordingly, all the advantages of CNN found in image-related composite designs with a self-learning algorithm. The concept
tasks are valid when applying to composite problems. Yang of the self-learning algorithm is similar to that of the genetic
et al. applied CNN models to predict the effective stiffness of algorithm. The sampling process consisted of many sampling
3-D composites (Table I).[70] In the authors’ work, the training loops. In each sampling loop, a large number of new composite
data were generated by 3-D Gaussian filters with different designs (candidates) were fed into their CNN model and the
covariance matrixes. The performance of the training data model predicted the performance of those designs with a negli-
was calculated by finite element analysis. The input variables gible computational cost. From there, high-performing candi-
include the topology of composites made up of two base mate- dates were identified and used to generate new candidates for
rials (stiff and soft) and the output variable is the effective stiff- the next sampling loop. The mechanical properties of the ML
ness. Different architectures of CNN models were explored to designs were then calculated using finite element analysis. As
tune hyperparameters (the numbers of layers and filters). The can be seen in Fig. 3, simulation results show that the ML
authors showed that their CNN models can produce highly designs are much tougher and stronger than the training samples.
accurate predictions of the effective stiffness of composites Most importantly, the simulation results were validated through
based on a given topology. additive manufacturing and experimental testing in the authors’
Recently, CNN has been employed to design hierarchical work. Note that knowing the optimal designs gives us an oppor-
composites. Gu et al. applied CNN models to accelerate the tunity to probe design strategies and physical insights (Fig. 3) of
design of bioinspired hierarchical composites (Table I).[66] In how to create tougher and stronger hierarchical composites.
the authors’ work, three unit cells made up of stiff and soft Recently, Hanakata et al. applied ML to accelerate the
materials were used to create hierarchical composites. Those design of stretchable graphene kirigami, which is a patterning
unit cells exhibited distinctive anisotropy responses because graphene sheet with cuts (Table I).[77] In the authors’ work,
of different geometrical configurations of stiff and soft materi- NN and CNN models were applied to predict the yield stress
als in the unit cells. The input variables represent the topology and yield strain of graphene sheets based on the cutting pattern
of hierarchical composites, where the three unit cells are described by two unit cells. The input variables represent the
denoted as 1, 2, and 3. Compared with using complete micro- topology of graphene sheets, where the two unit cells (with a
structural data with the geometrical configuration of all local cut and with no cut) are denoted as 0 and 1. The training
elements (stiff and soft), using macrostructural data with the data generated by MD and ML models was applied to perform
geometrical configuration of unit cells drastically reduces the a regression task to achieve accuracy close to the MD simula-
data space. Thus, a less complex ML architecture and fewer tions. After the training process was completed, an iterative
training samples are needed for this learning task. The output screening process was used to search for high-performing (high-
variable is the toughness of hierarchical composites calculated yield strain) candidates. In each screening iteration, their CNN
using finite element analysis. The idea of using unit cells to rep- model was applied to search for high-performing candidates
resent the topology of composites is similar to the homogeniza- and the performance of those candidates was calculated using
tion process commonly applied in estimating effective material MD. The new simulation results were then added to the training
properties of composites. However, the conventional homoge- data for the next iteration of training and screening processes.
nization process requires to extract effective (homogenized) The authors showed that when more and more samples were
parameters of unit cells based on a rigorous mathematical the- used to train the CNN model, the yield strains of the high-
ory,[76] which is not required in the authors’ ML approach. performing candidates in each iteration were also improved.
Moreover, the fracture of composites occurs on the microstruc-
tural scale (local elements) and cannot be captured in a homog- Gaussian process
enized FEM model. Consequently, in homogenized finite Although deep neural networks (e.g., NN and CNN) with many
element analysis, the localization process is required to convert hidden layers and neurons, theoretically, can capture any com-
the averaged strains of unit cells to the strains of local elements. plex patterns in data, they require a large amount of training
In contrast, Gu et al. showed that the correlations between the data in order to learn the hidden patterns without overfitting.
geometrical configuration of unit cells (macrostructural data) For some problems, simple ML algorithms cannot capture the
and the corresponding toughness (microstructural property) complex patterns in data and generating a large amount of

6▪ MRS COMMUNICATIONS • www.mrs.org/mrc

Downloaded from https://www.cambridge.org/core. University of New England, on 30 Mar 2019 at 16:31:18, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1557/mrc.2019.32
 Artificial Intelligence Prospective

Figure 3. ML-generated designs. (a) Strength and toughness ratios of designs computed from training data and ML output designs. Strength ratio is the
strength normalized by the highest training data strength value. Toughness ratio is the toughness normalized by the highest training data toughness value. The
ML output designs are shown from training loops of 1000 and 1,000,000. Envelopes show that ML material properties exceed those of training data. (b) Effects
of learning time on ML models for minimum, mean, and maximum toughness ratio start to converge as training loops increase. (c) Microstructures from
partitions A (lowest toughness designs in training data) and B (highest toughness designs from ML) in part (a) of the figure with the corresponding colors for
unit cell blocks (blue = U1, orange = U2, yellow = U3). Also shown in the right-most columns for the designs A and B are the strain distributions, which show
lower strain concentration at the crack tip for the ML-generated designs[66].

training data for deep learning algorithms is unfeasible due to points. Therefore, a GP model not only can make a prediction
the time and cost of conducting experiments or simulations. for an unseen data point based on the training data but also
A GP, which is a non-parametric approach, serves as an alter- can naturally quantify the uncertainty of the prediction. In gene-
native method for highly nonlinear problems. A GP is a collec- ral, the uncertainty of the prediction for an unseen data point
tion of random variables and assumes that all input and output increases when the point is away from the training data points.
variables have joint Gaussian distributions. Instead of setting a Although a GP is a powerful and elegant method with numerous
hypothesis for an ML model and finding optimal values for applications in ML and statistics, the implementation requires a
the weights in the model, a GP produces a distribution of all pos- computational cost which grows as O(n 3), where n is the number
sible functions that are consistent with the observed (training) of training data points. Thus, conventional GP models have been
data. Therefore, the number of parameters in a GP is unbounded mostly applied to problems with a small amount of training data.
and grows with the amount of training data. A GP is defined as: There are a number of sparse GP techniques proposed to reduce
the computational cost for large data sets.[78,79]
y = f (x)  GP(m(x), k(x, x′ )) (5) Tapia et al. applied a GP model to predict the porosity in
metallic parts produced by selective laser melting, which is a
where m(x) is the mean function and k(x, x′ ) is the covariance laser-based additive manufacturing process (Table I).[80] In
function. When a GP model is applied to predict the output the authors’ work, the input variables are the laser power and
value for an unseen data point, it generates a Gaussian distribu- laser scanning speed, which are two of the most influential pro-
tion with a covariance matrix produced by a kernel function (e.g., cessing parameters. The output variable is the porosity of spec-
squared exponential) which measures the similarity between imens (17-4 PH stainless steel). A GP can be computationally

MRS COMMUNICATIONS
Downloaded from https://www.cambridge.org/core. University of New England, on 30 Mar 2019 at 16:31:18, subject to the Cambridge Core terms • www.mrs.org/mrc ▪7
of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1557/mrc.2019.32
 Figure 4. (a) Young’s modulus versus toughness obtained from the metamodel. Pareto frontier obtained by sampling 1 million input parameters over the entire
design space. (b) One hundred initial CG-MD designs (blue dots) are used to build the metamodel. Using the metamodel, a Pareto frontier (red curve) is
obtained. Seven random points from the Pareto curve are chosen (purple squares) and tested by running CG-MD simulations (green diamonds). Comparison of
(c) Young’s modulus and (d) toughness obtained from the metamodel and CG-MD simulations. The error bars represent a 95% confidence interval. Reprinted
with permission from Ref. 81. Copyright 2018 American Chemical Society.

very efficient when the amount of training data is small. The of some basic ML algorithms and review recent studies using
authors used the GP model to predict the porosity of specimens ML models to predict mechanical properties of composites.
over the entire design space. Thus, the resulting porosity at any In those studies, ML models were applied to approximate
combination of the laser power and laser scanning speed was physics-based modeling tools such as MD and FEM with a
revealed. Recently, Hansoge et al. applied GP models to predict computational cost orders of magnitude less. Given the com-
mechanical properties of assembled hairy nanoparticles (NPs) puting speed advantage of ML approaches compared with
(Table I).[81] Assembled hairy NPs is a type of polymer nano- physics-based modeling tools, ML models were applied to
composite with relatively regular spacing between particles. In explore the design space of composite design problems and
the authors’ work, the input variables are the polymer chain to generate optimal designs of composites. By knowing the
length, grafting density, polymer–NP interaction strength, and optimal designs obtained from ML for different composite
NP-edge length. The output variable is the modulus or toughness design problems, it is possible to probe design strategies for
of assembled hairy NPs. The training data was generated using achieving high-performing composites. Although ML models
coarse-grained molecular dynamics (CG-MD) and GP models are often being considered as “black box” models, the discov-
were applied to perform regression tasks based on the CG-MD ered design strategies can provide new physical insights into
simulation results. The authors used the GP models (metamodel) how to make tougher and stronger composites.
to predict the modulus and toughness of assembled hairy NPs Future implementations for further breakthroughs in ML
over the entire design space. Figure 4 shows the modulus and applications on composite materials include developing more
toughness obtained from the metamodel by sampling 1 million efficient topology representations for composites and more effi-
assembled hairy NPs with different combinations of input vari- cient inverse design techniques. Although the topology of a
ables. As can be seen in the figure, the predicted modulus and composite can naturally be represented as an image by convert-
toughness obtained from the metamodel and those obtained ing the unit cells or base materials of the composite to pixels,
from the CG-MD simulations are very close. this representation is not as efficient for large composite sys-
tems (e.g., aircraft wings). For example, considering each
Summary and future perspectives base element in an FEM model as an input variable for ML nat-
This prospective paper presents recent progress in the applica- urally causes computational problems when the model is large.
tions of ML to composite materials. We give a brief overview Thus, developing new topology representations (with fewer

8▪ MRS COMMUNICATIONS • www.mrs.org/mrc

Downloaded from https://www.cambridge.org/core. University of New England, on 30 Mar 2019 at 16:31:18, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1557/mrc.2019.32
 Artificial Intelligence Prospective

variables) for large composite systems is imperative to apply 11. H.-S. Nam and D.J. Srolovitz: Molecular dynamics simulation of Ga pen-
ML to design large composite systems. In this prospective etration along grain boundaries in Al: A dislocation climb mechanism.
Phys. Rev. Lett. 99, 025501 (2007).
paper, we review recent studies using ML to solve inverse 12. Y.-J. Wang, G.-J.J. Gao, and S. Ogata: Atomistic understanding of diffu-
design problems. Gu et al. applied a linear model to generate sion kinetics in nanocrystals from molecular dynamics simulations. Phys.
optimal designs of composites by using the information of Rev. B 88, 115413 (2013).
13. P.D. Zavattieri, L.G. Hector Jr, and A.F. Bower: Cohesive zone simulations
the weights in the model.[74] However, the performance of
of crack growth along a rough interface between two elastic–plastic sol-
the ML designs was limited to the learning ability of the linear ids. Eng. Fract. Mech. 75, 4309 (2008).
model. In their later work, Gu et al. augmented CNN models 14. H. Jeong, S. Signetti, T.-S. Han, and S. Ryu: Phase field modeling of crack
with a self-learning algorithm to discover high-performing propagation under combined shear and tensile loading with hybrid for-
mulation. Comput. Mater. Sci. 155, 483 (2018).
hierarchical composite designs.[66] Hanakata et al. applied
15. A. Masud and A.A. Al-Naseem: Variationally derived discontinuity
ML with an iterative screening process to search for high- capturing methods: Fine scale models with embedded weak and
performing graphene kirigami.[77] Recently, Peurifoy et al. strong discontinuities. Comput. Methods Appl. Mech. Eng. 340, 1102
showed that NN models could be used to solve an inverse (2018).
16. H. Gajendran, R.B. Hall, and A. Masud: Edge stabilization and consistent
design problem for NPs by using backpropagation to train the tying of constituents at Neumann boundaries in multi‐constituent mixture
inputs of the NN models.[82] Although the gradient in this models. Int. J. Numer. Methods Eng. 110, 1142 (2017).
inverse design approach is analytical, not numerical, it is still 17. Y. Peng, A. Zhang, S. Li, and F. Ming: A beam formulation based on
a gradient-based optimization approach. Thus, this approach RKPM for the dynamic analysis of stiffened shell structures. Comput.
Mech. 63, 35 (2018).
can still get stuck in local minima or critical points. Note 18. X. Ma, S. Hajarolasvadi, G. Albertini, D.S. Kammer, and A.E. Elbanna: A
there is no optimization approach that can guarantee finding hybrid finite element‐spectral boundary integral approach: Applications
global minima of a non-convex optimization problem without to dynamic rupture modeling in unbounded domains. Int. J. Numer.
exploring the entire design space. Future work on understand- Anal. Methods Geomech. 43, 317 (2019).
19. G.X. Gu, F. Libonati, S.D. Wettermark, and M.J. Buehler: Printing nature:
ing the performance (theoretically or numerically) of different Unraveling the role of nacre’s mineral bridges. J. Mech. Behav. Biomed.
inverse design approaches for the applications of ML on com- Mater. 76, 135 (2017).
posites is required for the accelerated discovery of new materi- 20. G.X. Gu, M. Takaffoli, and M.J. Buehler: Hierarchically enhanced impact
resistance of bioinspired composites. Adv. Mater. 29, 1700060 (2017).
als. Though true inverse design is currently a challenge in the
21. F. Libonati, G.X. Gu, Z. Qin, L. Vergani, and M.J. Buehler: Bone‐inspired
field, it is believed that ML will play a promising role in this materials by design: toughness amplification observed using 3D printing
engineering problem in the future. and testing. Adv. Eng. Mater. 18, 1354 (2016).
22. P.-Y. Chen, J. McKittrick, and M.A. Meyers: Biological materials:
Functional adaptations and bioinspired designs. Prog. Mater. Sci. 57,
Acknowledgments 1492 (2012).
The authors acknowledge support from the Regents of the 23. J.W. Dunlop, R. Weinkamer, and P. Fratzl: Artful interfaces within biolog-
ical materials. Mater. Today 14, 70 (2011).
University of California, Berkeley. 24. G.X. Gu, I. Su, S. Sharma, J.L. Voros, Z. Qin, and M.J. Buehler:
Three-dimensional-printing of bio-inspired composites. J. Biomech.
Eng. 138, 021006 (2016).
References 25. J. Mueller, J.R. Raney, D.M. Kochmann, and K. Shea: Stiffness‐indepen-
1. B.G. Compton and J.A. Lewis: 3D‐printing of lightweight cellular compos- dent toughening of beams through coaxial interfaces. Adv. Sci. 5,
ites. Adv. Mater. 26, 5930 (2014). 1800728 (2018).
2. A.R. Studart: Biological and bioinspired composites with spatially tunable 26. A.A. Emery, J.E. Saal, S. Kirklin, V.I. Hegde, and C. Wolverton:
heterogeneous architectures. Adv. Funct. Mater. 23, 4423 (2013). High-throughput computational screening of perovskites for thermo-
3. Y. Sun and N. Pugno: Hierarchical fibers with a negative Poisson’s ratio chemical water splitting applications. Chem. Mater. 28, 5621 (2016).
for tougher composites. Materials (Basel) 6, 699 (2013). 27. A.L. Blum and P. Langley: Selection of relevant features and examples in
4. J. Wang, Q. Cheng, and Z. Tang: Layered nanocomposites inspired by the machine learning. Artif. Intell. 97, 245 (1997).
structure and mechanical properties of nacre. Chem. Soc. Rev. 41, 1111 28. V. Botu and R. Ramprasad: Adaptive machine learning framework to
(2012). accelerate ab initio molecular dynamics. Int. J. Quantum Chem. 115,
5. J. Yeo, G.S. Jung, F.J. Martín-Martínez, S. Ling, G.X. Gu, Z. Qin, and 1074 (2015).
M.J. Buehler: Materials-by-design: computation, synthesis, and charac- 29. I. Kononenko, and M. Kukar: Machine Learning and Data Mining:
terization from atoms to structures. Phys. Scr. 93, 053003 (2018). Introduction to Principles and Algorithms (Horwood Publishing,
6. C.-T. Chen, F.J. Martin-Martinez, G.S. Jung, and M.J. Buehler: Chichester, UK, 2007).
Polydopamine and eumelanin molecular structures investigated with ab 30. R.S. Michalski, J.G. Carbonell, and T.M. Mitchell: Machine Learning: An
initio calculations. Chem. Sci. 8, 1631 (2017). Artificial Intelligence Approach (Springer Science & Business Media,
7. C.-T. Chen and M.J. Buehler: Polydopamine and eumelanin models in Malaysia, 2013).
various oxidation states. Phys. Chem. Chem. Phys. 20, 28135 (2018). 31. K.T. Butler, D.W. Davies, H. Cartwright, O. Isayev, and A. Walsh:
8. C.-T. Chen, F.J. Martin-Martinez, S. Ling, Z. Qin, and M.J. Buehler: Machine learning for molecular and materials science. Nature 559,
Nacre-inspired design of graphene oxide–polydopamine nanocomposites 547 (2018).
for enhanced mechanical properties and multi-functionalities. Nano 32. K. Hansen, G. Montavon, F. Biegler, S. Fazli, M. Rupp, M. Scheffler, O.A.
Futures 1, 011003 (2017). Von Lilienfeld, A. Tkatchenko, and K.-R. Müller: Assessment and valida-
9. W. Xu, Z. Qin, C.-T. Chen, H.R. Kwag, Q. Ma, A. Sarkar, M.J. Buehler, and tion of machine learning methods for predicting molecular atomization
D.H. Gracias: Ultrathin thermoresponsive self-folding 3D graphene. Sci. energies. J. Chem. Theory Comput. 9, 3404 (2013).
Adv. 3, e1701084 (2017). 33. A. Mannodi-Kanakkithodi, G. Pilania, T.D. Huan, T. Lookman, and R.
10. T. Hatano and H. Matsui: Molecular dynamics investigation of dislocation Ramprasad: Machine learning strategy for accelerated design of polymer
pinning by a nanovoid in copper. Phys. Rev. B 72, 094105 (2005). dielectrics. Sci. Rep. 6, 20952 (2016).

MRS COMMUNICATIONS
Downloaded from https://www.cambridge.org/core. University of New England, on 30 Mar 2019 at 16:31:18, subject to the Cambridge Core terms • www.mrs.org/mrc ▪9
of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1557/mrc.2019.32
 34. G. Pilania, C. Wang, X. Jiang, S. Rajasekaran, and R. Ramprasad: 57. F.P. Melchels, M.A. Domingos, T.J. Klein, J. Malda, P.J. Bartolo, and D.W.
Accelerating materials property predictions using machine learning. Sci. Hutmacher: Additive manufacturing of tissues and organs. Prog. Polym.
Rep. 3, 2810 (2013). Sci. 37, 1079 (2012).
35. I. Santos, J. Nieves, Y.K. Penya, and P.G. Bringas: Machine-learning- 58. Q. Chen, P.F. Cao, and R.C. Advincula: Mechanically robust, ultraelastic
-based mechanical properties prediction in foundry production, in hierarchical foam with tunable properties via 3D printing. Adv. Funct.
ICCAS-SICE, 2009 (IEEE2009), pp. 4536. Mater. 28, 1800631 (2018).
36. S. Chen, K.-R.G. Reyes, M.K. Gupta, M.C. McAlpine, and W.B. Powell: 59. J.Z. Manapat, Q. Chen, P. Ye, and R.C. Advincula: 3D printing of polymer
Optimal learning in experimental design using the knowledge gradient nanocomposites via stereolithography. Macromol. Mater. Eng. 302,
policy with application to characterizing nanoemulsion stability. SIAM/ 1600553 (2017).
ASA J. Uncertainty Quantif. 3, 320 (2015). 60. N. Alsharif, A. Burkatovsky, C. Lissandrello, K.M. Jones, A.E. White, and
37. S.M. Neumayer, L. Collins, R. Vasudevan, C. Smith, S. Somnath, V.Y. Shur, K.A. Brown: Design and realization of 3D printed AFM probes. Small 14,
S. Jesse, A.L. Kholkin, S.V. Kalinin, and B.J. Rodriguez: Decoupling meso- 1800162 (2018).
scale functional response in PLZT across the ferroelectric–relaxor phase 61. M.P. Bendsøe, O. Sigmund, M.P. Bendsøe, and O. Sigmund: Topology
transition with contact Kelvin probe force microscopy and machine learn- Optimization by Distribution of Isotropic Material (Springer, Berlin,
ing. ACS Appl. Mater. Interfaces 10, 42674 (2018). Heidelberg, 2004).
38. L. Li, Y. Yang, D. Zhang, Z.-G. Ye, S. Jesse, S.V. Kalinin, and R.K. 62. A.T. Gaynor, N.A. Meisel, C.B. Williams, and J.K. Guest: Multiple-material
Vasudevan: Machine learning–enabled identification of material phase topology optimization of compliant mechanisms created via PolyJet
transitions based on experimental data: Exploring collective dynamics three-dimensional printing. J. Manuf. Sci. Eng. 136, 061015 (2014).
in ferroelectric relaxors. Sci. Adv. 4, eaap8672 (2018). 63. G.X. Gu, and M.J. Buehler: Tunable mechanical properties through texture
39. F. Faber, A. Lindmaa, O.A. von Lilienfeld, and R. Armiento: Crystal struc- control of polycrystalline additively manufactured materials using adjoint-
ture representations for machine learning models of formation energies. based gradient optimization. Acta Mech., 229, 4033 (2018).
Int. J. Quantum Chem. 115, 1094 (2015). 64. C.-C. Lin, and Y.-J. Lee: Stacking sequence optimization of laminated
40. K. Schütt, H. Glawe, F. Brockherde, A. Sanna, K. Müller, and E. Gross: composite structures using genetic algorithm with local improvement.
How to represent crystal structures for machine learning: Towards fast Compos. Struct. 63, 339 (2004).
prediction of electronic properties. Phys. Rev. B 89, 205118 (2014). 65. T. Zegard and G.H. Paulino: Bridging topology optimization and additive
41. T. M. Dieb, Z. Hou, and K. Tsuda: Structure prediction of boron-doped manufacturing. Struct. Multidisc. Optim. 53, 175 (2016).
graphene by machine learning. J. Chem. Phys. 148, 241716 (2018). 66. G.X. Gu, C.-T. Chen, D.J. Richmond, and M.J. Buehler: Bioinspired hier-
42. B. Meredig, A. Agrawal, S. Kirklin, J.E. Saal, J. Doak, A. Thompson, K. archical composite design using machine learning: simulation, additive
Zhang, A. Choudhary, and C. Wolverton: Combinatorial screening for manufacturing, and experiment. Mater. Horiz. 5, 939 (2018).
new materials in unconstrained composition space with machine learn- 67. C. Bisagni and L. Lanzi: Post-buckling optimisation of composite stiffened
ing. Phys. Rev. B 89, 094104 (2014). panels using neural networks. Compos. Struct. 58, 237 (2002).
43. M. De Jong, W. Chen, R. Notestine, K. Persson, G. Ceder, A. Jain, M. Asta, 68. A. Paul, P. Acar, W.-K. Liao, A. Choudhary, V. Sundararaghavan, and A.
and A. Gamst: A statistical learning framework for materials science: Agrawal: Microstructure optimization with constrained design objectives
application to elastic moduli of k-nary inorganic polycrystalline com- using machine learning-based feedback-aware data-generation. Comput.
pounds. Sci. Rep. 6, 34256 (2016). Mater. Sci. 160, 334 (2019).
44. L. Ward, A. Agrawal, A. Choudhary, and C. Wolverton: A general-purpose 69. Z. Yang, Y.C. Yabansu, D. Jha, W.-k Liao, A.N. Choudhary, S.R. Kalidindi,
machine learning framework for predicting properties of inorganic mate- and A. Agrawal: Establishing structure–property localization linkages for
rials. NPJ Comput. Mater. 2, 16028 (2016). elastic deformation of three-dimensional high contrast composites
45. R. Ramprasad, R. Batra, G. Pilania, A. Mannodi-Kanakkithodi, and C. Kim: using deep learning approaches. Acta Mater. 166, 335 (2018).
Machine learning in materials informatics: recent applications and pros- 70. Z. Yang, Y.C. Yabansu, R. Al-Bahrani, W.-k Liao, A.N. Choudhary, S.R.
pects. NPJ Comput. Mater. 3, 54 (2017). Kalidindi, and A. Agrawal: Deep learning approaches for mining struc-
46. Y. Liu, T. Zhao, W. Ju, and S. Shi: Materials discovery and design using ture–property linkages in high contrast composites from simulation
machine learning. J. Materiomics 3, 159 (2017). datasets. Comput. Mater. Sci. 151, 278 (2018).
47. A. Lavecchia: Machine-learning approaches in drug discovery: Methods 71. S. Tiryaki and A. Aydın: An artificial neural network model for predict-
and applications. Drug Discov. Today 20, 318 (2015). ing compression strength of heat treated woods and comparison with
48. G.B. Goh, N.O. Hodas, and A. Vishnu: Deep learning for computational a multiple linear regression model. Constr. Build. Mater. 62, 102
chemistry. J. Comput. Chem. 38, 1291 (2017). (2014).
49. J. Bauer, S. Hengsbach, I. Tesari, R. Schwaiger, and O. Kraft: 72. F. Khademi, M. Akbari, S.M. Jamal, and M. Nikoo: Multiple linear regres-
High-strength cellular ceramic composites with 3D microarchitecture. sion, artificial neural network, and fuzzy logic prediction of 28 days
Proc. Natl. Acad. Sci. USA 111, 2453 (2014). compressive strength of concrete. Front. Struct. Civil Eng. 11, 90
50. J.W. Dunlop and P. Fratzl: Biological composites. Annu. Rev. Mater. Res. (2017).
40, 1 (2010). 73. B.A. Young, A. Hall, L. Pilon, P. Gupta, and G. Sant: Can the compressive
51. Y. Ding, J. Muñiz-Lerma, M. Trask, S. Chou, A. Walker, and M. Brochu: strength of concrete be estimated from knowledge of the mixture propor-
Microstructure and mechanical property considerations in additive man- tions?: New insights from statistical analysis and machine learning meth-
ufacturing of aluminum alloys. MRS Bull. 41, 745 (2016). ods. Cem. Concr. Res. 115, 379 (2019).
52. I. Gibson, D. Rosen, and B. Stucker: Additive Manufacturing Technologies 74. G.X. Gu, C.-T. Chen, and M.J. Buehler: De novo composite design based
(Springer, New York, 2014). on machine learning algorithm. Extreme Mech. Lett. 18, 19 (2018).
53. G.X. Gu, S. Wettermark, and M.J. Buehler: Algorithm-driven design of 75. C. Qi, A. Fourie, and Q. Chen: Neural network and particle swarm optimi-
fracture resistant composite materials realized through additive manufac- zation for predicting the unconfined compressive strength of cemented
turing. Addit. Manuf. 17, 47 (2017). paste backfill. Constr. Build. Mater. 159, 473 (2018).
54. N. Guo and M.C. Leu: Additive manufacturing: Technology, applications 76. J.-C. Michel, H. Moulinec, and P. Suquet: Effective properties of compos-
and research needs. Front. Mech. Eng. 8, 215 (2013). ite materials with periodic microstructure: a computational approach.
55. Y. Huang, M.C. Leu, J. Mazumder, and A. Donmez: Additive manufactur- Comput. Methods Appl. Mech. Eng. 172, 109 (1999).
ing: current state, future potential, gaps and needs, and recommenda- 77. P.Z. Hanakata, E.D. Cubuk, D.K. Campbell, and H.S. Park: Accelerated
tions. J. Manuf. Sci. Eng. 137, 014001 (2015). search and design of stretchable graphene kirigami using machine learn-
56. K. Makiewicz, S. Babu, M. Keller, and A. Chaudhary: Microstructure ing. Phys. Rev. Lett. 121, 255304 (2018).
Evolution During Laser Additive Manufacturing of Ti6A14V Alloys. In 78. J. Quiñonero-Candela, and C.E. Rasmussen: A unifying view of sparse
Proceedings of the International Conference Trends Weld. Res. Chicago, approximate Gaussian process regression. J. Mach. Learn. Res. 6,
IL (2012). 1939 (2005).

10 ▪ MRS COMMUNICATIONS • www.mrs.org/mrc

Downloaded from https://www.cambridge.org/core. University of New England, on 30 Mar 2019 at 16:31:18, subject to the Cambridge Core terms of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1557/mrc.2019.32
 Artificial Intelligence Prospective

79. E. Snelson and Z. Ghahramani: Sparse Gaussian processes using

pseudo-inputs, in Advances in neural information processing systems
(2006), pp. 1257.
80. G. Tapia, A. Elwany, and H. Sang: Prediction of porosity in metal-based
additive manufacturing using spatial Gaussian process models. Addit.
Manuf. 12, 282 (2016).
81. N.K. Hansoge, T. Huang, R. Sinko, W. Xia, W. Chen, and S. Keten:
Materials by design for stiff and tough hairy nanoparticle assemblies.
ACS Nano 12, 7946 (2018).
82. J. Peurifoy, Y. Shen, L. Jing, Y. Yang, F. Cano-Renteria, B.G. DeLacy, J.D.
Joannopoulos, M. Tegmark, and M. Soljačic:́ Nanophotonic particle sim-
ulation and inverse design using artificial neural networks. Sci. Adv. 4,
eaar4206 (2018).

MRS COMMUNICATIONS
Downloaded from https://www.cambridge.org/core. University of New England, on 30 Mar 2019 at 16:31:18, subject to the Cambridge Core terms • www.mrs.org/mrc ▪ 11
of use, available at https://www.cambridge.org/core/terms.
https://doi.org/10.1557/mrc.2019.32