SCIENCE ADVANCES | RESEARCH ARTICLE

APPLIED SCIENCES AND ENGINEERING Copyright © 2021

The Authors, some
Deep learning model to predict complex stress rights reserved;
exclusive licensee
and strain fields in hierarchical composites American Association
for the Advancement
of Science. No claim to
Zhenze Yang1,2†, Chi-Hua Yu1,3†, Markus J. Buehler1,4,5* original U.S. Government
Works. Distributed
Materials-by-design is a paradigm to develop previously unknown high-performance materials. However, finding under a Creative
materials with superior properties is often computationally or experimentally intractable because of the astronomical Commons Attribution
number of combinations in design space. Here we report an AI-based approach, implemented in a game NonCommercial
theory–based conditional generative adversarial neural network (cGAN), to bridge the gap between a material’s License 4.0 (CC BY-NC).
microstructure—the design space—and physical performance. Our end-to-end deep learning model predicts
physical fields like stress or strain directly from the material microstructure geometry, and reaches an astonishing
accuracy not only for predicted field data but also for derivative material property predictions. Furthermore, the
proposed approach offers extensibility by predicting complex materials behavior regardless of component shapes,
boundary conditions, and geometrical hierarchy, providing perspectives of performing physical modeling and simu-
lations. The method vastly improves the efficiency of evaluating physical properties of hierarchical materials
directly from the geometry of its structural makeup.

INTRODUCTION to performance. For many design applications and engineering analyses,

Because of high demand for materials with superior mechanical we need access to physical fields like strain or stress tensor distribu-
properties and versatile functionalities, tuning composite designs has tions. These types of physical data will also improve the physical
become a crucial part in materials development (1–4). The essence relevance of AI approaches and provide more mechanistic insights.
of composites lies in introducing heterogeneity through combina- In recent years, ML methods, especially deep learning (DL), have
tions of multiple materials with distinct, often disparate, properties. revolutionized our perspective of designing materials, modeling physical
By optimizing the spatial distributions of composite constituents, phenomena, and predicting properties (21–26). DL algorithms de-
vast enhancements of mechanical performance can be realized such veloped for computer vision and natural language processing can

Downloaded from https://www.science.org on February 10, 2022

as demonstrated in modern ceramics matrix materials (5, 6), fiber-­ be used to segment biomedical images (27), design de novo proteins
reinforced polymers (7), or architected materials including biologically (28–30), and generate molecular fingerprints (31, 32). Within the
evolved and bioinspired (8–11). However, traditional manufacturing field of computational materials science, the abundance of data enables a
methods limited tunable composite designs due to the difficulties of boom of ML applications from quantum scale up to macroscale. By
conjoining different base materials and manipulating complex incorporating field-based (density function theory), particle-­based
microstructures (12). To address these issues, additive manufacturing (MD), and continuum-based (FEM) modeling, ML sheds light on
has recently enabled the production of composites with complex predictions of quantum interactions (33–35), molecular force fields
geometries (13–15). (36–38), and material mechanics (19, 20, 39–41). However, most of
To calculate physical properties of composites often relies on these models are limited by a minimal level of information included
multiscale modeling approaches such as finite element method (FEM) in the prediction and difficulty in generalization.
(16) or molecular dynamics (MD) (17, 18) simulations. However, Here, we overcome these challenges and present a general method
the design space of composites usually exists of an intractable num- to translate material composition, provided to the model as a simple
ber of combinations that hinders us from searching optimal designs image of the geometry and microstructure that fully encodes mate-
either via experimental measurements or computational simula- rial composition and boundary conditions, directly to strain or stress
tions. In earlier work, enabled by advances in AI methods, instead fields as physical fields containing integral information of material
of using a brute-force approach, machine learning (ML) algorithms behaviors to bridge physics and designs. We show that the physical
were implemented to optimize composite designs based on the cal- relationship between composite geometries and strain or stress fields
culation of FEM for various loading conditions, such as tensile or can be directly established at high levels of accuracy and predictability.
shear loading (19, 20). However, these methods do not yet predict While some earlier work has used convolutional neural networks to
physical fields and do not directly connect material microstructure predict localization linkages for elastic deformation in composites
(42), the results reported in this paper offers a method that is capa-
1
Laboratory for Atomistic and Molecular Mechanics (LAMM), Massachusetts Institute ble to deal with more generalized mechanical problems and can be
of Technology, 77 Massachusetts Ave., Cambridge, MA 02139, USA. 2Department more easily used to incorporate experimental data.
of Materials Science and Engineering, Massachusetts Institute of Technology, 77
Massachusetts Ave., Cambridge, MA 02139, USA. 3Department of Engineering Science,
National Cheng Kung University, No.1, University Road, Tainan City 701, Taiwan.
4
Center for Computational Science and Engineering, Schwarzman College of Com- RESULTS
puting, Massachusetts Institute of Technology, 77 Massachusetts Ave., Cambridge, Model development and training
MA 02139, USA. 5Center for Materials Science and Engineering 77 Massachusetts A generative adversarial network (GAN) is a type of deep neural
Ave., Cambridge, MA 02139, USA.
*Corresponding author. Email: [email protected] network that generates new data based on the data statistics of the
†These authors contributed equally to this work. training set (43). GANs consist of two key components known as

Yang et al., Sci. Adv. 2021; 7 : eabd7416 9 April 2021 1 of 10

SCIENCE ADVANCES | RESEARCH ARTICLE

the Generator and the Discriminator, which are trained against unloading is the target. This represents a complex physical scenario
each other based on the game theory. The Generator generates and will serve as a test case to assess the method reported here. Typical
candidates that are evaluated by the Discriminator. Although GANs results of strain or stress field predictions of the test dataset are shown
were originally proposed as a form of unsupervised learning, labels in Fig. 2A. Notably, both stress and strain fields as well as displace-
can be incorporated as constrains, which leads to a subtype of GANs ments (incorporated in the overall deformation of the geometry)
known as the condition GAN (cGAN). In our work, we develop a are captured in our prediction, as shown in Fig.  3A. To further
DL model implemented in a cGAN with a paired image as the con- quantitatively evaluate the similarity between the ground truth and
straint (44). The cGAN model has two key components, Generator the prediction, the L2 norm (see Materials and Methods) is calculat-
U-Net and Discriminator PatchGAN (44). The geometric images ed for all 400 data in the test dataset, considering von Mises stress
(labels or constraints) are fed into the Generator to generate field field prediction.
images of interest with random noise. The Discriminator evaluates The distributions of L2 norm of predicted contours are plotted
these field images generated by the Generator through comparing against two reference contours, clean contours and random contours
them with real field images obtained from FEM. The training objec- (Fig. 3A). According to the figure, the mean value of the L2 norm of
tive of the Generator is to increase the error rate of the Discriminator, predicted contours is much lower when comparing with random
while the Discriminator is trained to optimize the capacity of iden- contours (~1.8, normalized by predicted contour) and clean con-
tifying fake images produced by the Generator. Within the game tours (~2.9, normalized by predicted contour). Furthermore, a
theory framework used here, the GAN model converges when the narrower peak of predicted contours indicates that the variation is
Discriminator and the Generator reach a Nash equilibrium in which smaller as well (random contour ~2.8 and clean contour ~2.7, nor-
one component maintains its status regardless of the actions of malized by predicted contour). The results reveal that our ML model
the opponent. features both high accuracy and applicability in predicting strain or
Once the ML model is trained, new field images can be predicted stress fields directly from the composite geometry.
bypassing conventional numerical simulations. The field predictions With predicted fields, we can further derive secondary mechan-
can be further used to extract mechanical properties of the compos- ical properties of the composites. For instance, mechanical recover-
ite given its geometry. As we show in the following sections, not ability of the composite can be computed after subjected to
only mechanical properties including stiffness and recoverability compressive loading and unloading. Here, recoverability is defined
can be obtained but also local features such as stress concentrations as average residual stress of the composites. Our model is able to
around cracks or sharp inclusions. More details are provided in the predict the geometries associated with the top three recoverability
Supplementary Materials. levels in the test dataset exactly (Fig. 3C), with the information
obtained from the field images. And more broadly, our ML model

Downloaded from https://www.science.org on February 10, 2022

From geometry to strain or stress field performed an accurate prediction on the ranks and the values of
We focus on two-dimensional (2D) composites with two constitu- recoverability or residual stress over all 400 composites in the test
ent materials, defined as soft units (red) and brittle units (white). dataset (Fig. 3D).
Both units have linear plasticity and strain hardening, which is The R2 value of the linear fitting between ML model prediction
characterized by the “crushable foam” model in Abaqus (see and the ground truth is 0.96, and the average relative errors of ML
Materials and Methods). Brittle units have relatively larger Young’s model prediction of the global property recoverability is 7.5%. It is
modulus but smaller yield strain comparing with soft units. The ini- emphasized that the ML model was not trained directly for predict-
tial composite considered consists of 8 × 8 block units (see Materials ing recoverability, as this measure is a secondary extracted feature.
and Methods) (19, 20), keeping the size small so we can conduct However, the model is still able to predict the secondary informa-
a brute-force validation of the ML method against full-physics tion obtained from field images precisely with only 1600 training
simulations. While the generated patterns are at coarse resolu- data. The case of recoverability suggests that our model is not only
tions (8 × 8 in the first example), the overall image resolution is giving field images that look similar to the ground truth but also
256 pixels × 256 pixels. predicting specific pixel values at high accuracy. As a consequence,
The arrangement of block units is generated by a random geometry the model can be used to optimize mechanical properties of com-
Generator so that we can explore a broad range of arrangement posites by varying geometries. Furthermore, there is no longer a need
combinations (Fig. 1). For different geometries, strain or stress to develop different ML models for different mechanical properties
fields of composites under mechanical tests such as compression are (19, 20) as multiple properties can be derived from predicted fields
obtained using FEM (see Materials and Methods). The results from and deformation already included in our model.
FEM are regarded as the ground truth in this work. Both geometry In addition to mechanical property predictions, we also check
images and strain or stress field images are collected to compose the the reliability for local, smaller-scale patterns. To assess this capability,
dataset. The dataset consists of 2000 data in total, which is split into a ML model trained with 8 × 8 representation is used to study
a training dataset with 80% of the data and a test dataset with the rest high-resolution patterns at an increased 16 × 16 resolution (Fig. 3E).
20% of the data. Clear boundaries of chessboard pattern are predicted even if the
16 × 16 representation has never been seen by the model. The con-
Evaluation of model performance sistence across different scales reveals that the GAN has learned a
To evaluate model performance, we consider the mechanical be- physical understanding about mechanical phenomena, in this case,
haviors of composites under compressive loading and unloading stress concentrations and how they emerge from complex micro-
(loading applied in the x direction, or horizontal axis, Fig. 2A) as an structural patterns. Although the checkerboard-based stress field agrees
example (see Materials and Methods and movie S1). The geometry well, it is important to assess the divergence of the stress values or
image is the input, while the field image of residual stress after high-order information such as the gradient of strain or stress fields

Yang et al., Sci. Adv. 2021; 7 : eabd7416 9 April 2021 2 of 10

SCIENCE ADVANCES | RESEARCH ARTICLE

Fig. 1. Depiction of the workflow of the method reported here. Starting from a random Generator, geometry images (8 × 8) of composites are created. A FEM analysis
is performed to obtain real field information of composites under mechanical tests. To predict strain and stress fields from geometry images, we train a ML model known
as GAN (red window) which consists of two components, Generator U-Net (blue window) and Discriminator PatchGAN (green window). The Generator takes geometry
images as input to generate fake images with wrong fields. Discriminator compares fake images from Generator with real images from FEM. Once the model is trained,

Downloaded from https://www.science.org on February 10, 2022

accurate field predictions are made that are validated against high-fidelity FEM models, as well as composites with new geometries. In addition, predicted field images
can be postprocessed to extract global or local mechanical features.

across different scales. In terms of reducing the divergence regarding boundary conditions, which are embedded directly in the images.
different scales in the future work, an alternative model can be trained The loading conditions are specified by adding green lines to our
by incorporating multiple representations (e.g., 8 × 8, 16 × 16, and geometry images and field images (Fig. 4A). These green lines are
32 × 32 resolutions, each with 1000 images) into the dataset. With such representative of rigid bodies that exert loads during simulations of
multiscale information, the model will likely be able to learn how to upscale FEM. The straight green lines are two rigid lines used to compress
or downscale the strain or stress fields across different hierarchies. the composites along the x direction. The circular green line is a
In conclusion, our model’s capacity for recognizing scale-inde- spherical (2D) rigid indenter used for nanoindentation. Hence, now
pendent local patterns can be applied to predict strain or stress field the dataset consists of two sets of data: One set of images contains
of hierarchical structures and is useful for design applications that two straight green lines showing the results of compression, and the
explore a broad range of de novo microstructures. other set includes one circular line exhibiting the results of nano­
indentation. The geometries of composites in the two sets are the same.
Nonsquare features After being trained with the mixed dataset, the model is able to
In the earlier examples, the composite was composed of square units. recognize different loading conditions and make relevant predictions
Now, we explore whether our model can be extended to investigate (Fig. 4A). More specifically, the model transfers loading conditions
composites with nonsquare elementary design units. To demonstrate from input geometry images to the prediction and predicts von Mises
such possibility, both hexagonal and triangular tessellations are used stress field according to the specified loading conditions. More
for designing composite elements. To do this, the same model trained unexpectedly, when two loading conditions are combined in the
on square representation is now trained with a dataset in hexagonal geometry image that is not included in our training dataset, the ML
and triangular representations, respectively. von Mises stress fields model is still able to predict the stress field accurately. As is shown
calculated by FEM and the trained model show high similarity, as in Fig.  4A, both stress fields caused by compression (the general
visualized in two example cases (Fig. 4A). The results indicate that stress pattern similar to single compression) and nanoindentation
the model can be easily generalized to composites of different shapes. (stress concentration on the top) are predicted by our model.
The ability of our model to read two distinct loading conditions
Multiple loading conditions from images and predict fields accordingly shows the potential of
The input geometry images we used above do not yet include any applying one model to predict multiple mechanical tests and offers
information about loading conditions, which were encoded in the evidence for transferability. In addition, in principle, the model is
training set used. Now, we extend the model to learn variations in also able to predict strain or stress fields for multistage mechanical tests

Yang et al., Sci. Adv. 2021; 7 : eabd7416 9 April 2021 3 of 10

SCIENCE ADVANCES | RESEARCH ARTICLE

Downloaded from https://www.science.org on February 10, 2022

Fig. 2. Strain or stress field predictions along with displacement fields, depicting a comparison the FEM based ground truth with the prediction from the ML model. (A) Sche-
matic of loading conditions in FEM. The outputs of the end frame after compressive loading and unloading are collected to build the dataset (see Materials and Methods). (B) Three
examples of strain and stress fields predictions, shown for illustration. In geometry images, red represents soft units, and white represents brittle units. The stress field is von Mises
stress field, and the strain refers to plastic strain PE11 (“1” refers to the x direction). The solid black background is plotted to show the reference frame, to visualize the predicted
deformation of the specimen. Note that while the patterns are at coarse resolutions (8 × 8 in the example shown here), the overall image resolution is 256 pixels × 256 pixels.

such as multiple loading cycles. One straightforward way to achieve geometries, we trained an even deeper model (see Materials and Methods)
that within the framework reported here is by leveraging different with 32 × 32 representation (Fig. 4C). Similarly, as before, the loading
color codes for different loading stages. For instance, as the loading condition is a one-cycle compressive loading and unloading with
cycles increase, we can vary the colors of the green lines in Fig. 4B von Mises stress field images being the target outputs. To show an ex-
from light green to dark green to encode the loading stages for training. ample of a more complicated composition and the associated pre-
With our well-trained model, FEM simulations that are usually carried diction, a geometry image in the shape of the MIT dome is used as the
out by conventional codes can be performed at much higher speed input. Both the ground truth and the field image predicted by the
and lower computational costs. Last, our model can also be used for model exhibit a similar stress pattern (Fig. 4C). In other words, a deeper
complex loading conditions due to its capacity of predicting fields model trained with 32 × 32 representation is more powerful than that
with mixed loading conditions. in 8 × 8 representation because it can better present nonsquare pat-
terns such as the sector in Fig. 4C and cover more situations when a
High-resolution geometry desired resolution is requested. The example reveals the potential of
For real-world composites, their geometries can be quite complicated the model to predict high-resolution fields for general inputs.
(45, 46). As a consequence, a relatively simple 8 × 8 representation To quantitatively investigate the accuracy of high-resolution pre-
may not be sufficient for practical applications. To study complex diction, we calculated the L2 norm as we did for 8 × 8 representation.

Yang et al., Sci. Adv. 2021; 7 : eabd7416 9 April 2021 4 of 10

SCIENCE ADVANCES | RESEARCH ARTICLE

Fig. 3. Model performance. (A) Direct comparison of both deformation (green line) and stress field (yellow line) between the ground truth and the prediction of ML
model. (B) Quantitative comparison of image difference between the ground truth and the prediction of ML model. The L2 norm indicates the image difference. “Random
contour” and “clean contour” are two reference frames (see Materials and Methods). (C) Geometry images of top three mechanical recoverability in the test dataset. The
results from FEM and ML models are consistent, showing excellent predictive power. (D) Rank of recoverability given by FEM and ML considering all 400 data in the test
dataset. The inset displays the relative errors of ML predictions of the residual stress/recoverability for all 400 data in the test dataset. (E) The ML model trained with 8 × 8

Downloaded from https://www.science.org on February 10, 2022

geometry images accurately predicts a downscaled pattern of 16 × 16, providing evidence that the model can predict features at multiple length scales. (Note that the
physical dimensions of the composites are fixed.) The chessboard geometry is used for clarifying resolution.

However, when using a 32  ×  32 representation, there are much DISCUSSION

more local patterns than in the 8 × 8 representation. As a result, a We proposed a DL-based approach to provide a direct translation
comprehensive comparison of differences across the whole image is of the composite geometry to strain or stress fields, realized using a
more suitable than a single value. Hence, we use an L2 norm map game-theoretic approach implemented as a GAN. The neural net-
instead of the L2 norm. The L2 norm map is obtained by calculating works are trained with a relatively small amount of data but reach
local L2 norm region by region when comparing two images (see astonishing accuracy, transferability, and, hence, broad applica-
Materials and Methods). We randomly generated a 32 × 32 geome- bility. Multiple mechanical fields are predicted with the same trained
try image to check the similarity between the ground truth and pre- model covering multiple aspects of materials’ behaviors (Fig. 2 and
diction using L2 norm map. fig. S3). By extracting information from the field predictions, the
Two reference contour images, known as clean contour and ran- model can identify top designs of recoverability exactly and capture
dom contour, are selected in the same way as we did for the 8 × 8 scale-consistent local patterns. In terms of the acceleration of the
representation (Fig.  3A). The L2 norm map of the clean contour computation speed gained by our method, the predictions of pre-
exhibits the positions of stress concentration, while for the random trained model take less than a second on a Intel i7-9800X (3.80 GHz)
contour, the map includes mixed information of two fields. Accord- CPU core, whereas the standard FEM simulations usually last minutes,
ing to the L2 norm map, the predicted contour is globally accurate hours, or even days. Therefore, the approach is a promising tool to
with low values of L2 norm across the map. We find that for those accelerate discovery of optimal geometric designs for materials such
points with large stress concentration in the image, the model shows as metamaterials and hierarchical materials.
less accuracy. The reason is that the ML model is inclined to By using soft materials, the model can predict analogous stress
smoothing the sharp peaks—stress concentrations—to achieve an fields around cracks (see Materials and Methods). The results agree
overall low loss. However, the tendency of displaying large stress in with the knowledge of the patterns of stress fields around cracks
those points is clearly predicted by our model, as is evidenced by even if the ratio of soft materials for these cases is extremely skewed
comparing with the L2 norm map of the clean contour. With the in the data distribution (fig. S4). The predictions are consistent re-
capacity of predicting fields for high-resolution geometries, the gardless of shapes and positions of cracks or sharp inclusions and
model can be used to investigate composites with complex patterns could even be used to describe the evolution of fields as cracks prop-
as is necessary in design applications. agate (movie S1). As a result, the model could also be used for

Yang et al., Sci. Adv. 2021; 7 : eabd7416 9 April 2021 5 of 10

SCIENCE ADVANCES | RESEARCH ARTICLE

Downloaded from https://www.science.org on February 10, 2022

Fig. 4. Extension of model to include multiple loading conditions, as well as image-based coding of boundary conditions. (A) von Mises stress field prediction on
nonsquare representation. Composites with hexagonal and triangular units are presented. (B) ML model trained for composites under various loading conditions. In this
part, loading conditions are embedded in images by using green lines (straight/semicircular). Straight line indicates that the loading condition is compression, while
semicircular line is used for nanoindentation. The training dataset contains images each with one loading condition, either compression or nanoindentation. von Mises
stress field is the target output. The ML model trained on the dataset is able to recognize different loading conditions and even predict the stress field with both loading
conditions exerted (Combination). (C) ML model trained for composites with high-resolution geometry. The geometry image is 32 × 32 with a ratio of soft/brittle units
unfixed. von Mises stress field is the target output. A logo of MIT dome is used as an example of complicated geometry, which can only be presented with high resolution.
The overall image resolution for high-resolution geometry is 512 pixels × 512 pixels. (D) Evaluation of ML model performance on predicting high-resolution field [ML
model from (C)]. Image difference is calculated by using L2 norm map, which can display local differences of two images (see Materials and Methods). L2 norm maps of
comparing the ground truth with “clean contour,” “predicted contour,” and “random contour” (see Materials and Methods) are generated.

crack-related design problems such as crack-resistant materials complex systems of interest. Moreover, the approach reported here
(47), providing added evidence for the transferability of the method. can also be directly applied to experimental images for training of
Our model can capture mechanical behaviors of composites the model, which underscores the transferability of the approach,
with various shapes, different loading conditions, and complicated and provides a previously unidentified way to combine bottom-up
geometries. It can predict mechanical behaviors from random ge- modeling with experimental data sources for predictive methods.
ometries and open the gate to accelerate searching optimal designs Future work could also focus more on global mechanical properties
of composites with multiple mechanical functionalities from the derived from local stress and strain fields, as opposed to local fields
bottom up. In that context, one of the extraordinary advantages of as focused on in this study. Although we covered a detailed analysis
using predicted stress and strain fields is that fields contain compre- of one global property—recoverability—some global properties may
hensive information for various design purposes. Thus, the approach possibly be more sensitive to divergence within the predicted maps.
offers a high level of efficiency of predicting physical properties and For instance, the fracture toughness is strongly correlated with local
accelerate the design process based on a transferrable ML method. predictions of stress fields around a crack tip. In the case, the train-
Furthermore, we extended our model to incorporate nonsquare repre- ing data may need to be modified accordingly, such as incorporating
sentations, multiple loading conditions, cracks, and high-resolution sharp edges and dynamic effects.
geometries. These extensions enabled us to cover tremendous ap- Moreover, combined with optimization algorithms (20), the model
plications of FEM with lower computational costs and investigate can speed up the discovery of optimal designs without heavy

Yang et al., Sci. Adv. 2021; 7 : eabd7416 9 April 2021 6 of 10

SCIENCE ADVANCES | RESEARCH ARTICLE

computational loads. Moreover, the proposed approach is a general split them into training dataset with 80% data and test dataset with
method for geometry-to-field translation. A similar protocol as de- the rest 20% data. The split is also stochastic, thus the distribution is
veloped here could be applied to other areas in the sciences, such as being conserved.
density functional theory fields, fluid mechanical fields, or electro-
magnetism. Image processing
The images of mechanical fields are generated first in Abaqus Visu-
alization and then postprocessed by Python code. As for the gener-
MATERIALS AND METHODS ation, the images are plotted with both deformation (to capture the
Graphical representation of composites displacement field in the geometry change) and strain or stress field
We use 8 × 8 graphical representation for composite profile, which contours. The color spectrum for field contours is realized using “white”
is shown in fig. S1A. For any composite with the given geometry, it [RGB = (216, 216, 216)] for lower bound and “red” [RGB = (216, 7, 0)]
consists of two different block units, labeled as brittle unit and soft for the upper bound. The lower and upper bounds vary from cases
unit. Brittle unit and soft unit are mechanically distinct in elasticity to cases but are consistent within one dataset. The same spectrum is
and plasticity. A brutal force algorithm is executed to generate all also used for geometry input with white [RGB = (216, 216, 216)] for
possible  combinations of two different units without fixing the brittle units and red [RGB = (216, 7, 0)] for soft units. The contour
composition. To simplify the geometry and shorten the calculation style used is “DISCRETE,” and the number of intervals is set to 24.
time in Abaqus, all composites created by the algorithm are sym- As for options for edges, “FEATURE” is used for visible edges, and
metric about Y axis. “THICK” is used specifically for green lines, which represent ana-
lytical rigid bodies. Any triad, legend, title, state, annotation, com-
FEM pass, and reference point are turned off to exclude useless information
The dataset is generated by FEM, using a commercial Abaqus/ in the images. In addition, the background of the image is set to
Explicit code (Dassault Systemes Simulia Corp., 2010). In our work, solid black [RGB = (0, 0, 0)]. After all settings are applied, an image
strain or stress fields calculated by FEM are considered as the ground is generated with the “print” command in Abaqus.
truth when comparing with ML results. All simulations are carried On the basis of the images created by Abaqus, a Python code is
out in 2D. To include plastic deformation, crushable foam model executed to cut, resize, and stitch images. Input image (geometry
with volumetric hardening embedded in Abaqus is implemented images) and target image (field images) are cut to keep the propor-
for composites (48, 49). Detailed material properties of two differ- tion of black background similar vertically and horizontally. The
ent units in composites are exhibited (tables S1 and S2). Table S1 purpose of cutting operation is to make sure that the composites
displays overall mechanical behavior of materials, while the specific in input image and target image are generally matching in shape to

Downloaded from https://www.science.org on February 10, 2022

hardening curve is defined by table S2. To briefly summarize the be read by the GAN model. Once the images are cut, they are
difference of two materials shown by the data, the brittle unit has all resized to images with a size of 512 × 512. After resizing, two
relatively larger Young’s modulus but a smaller yield strain com- images are stitched in accordance with the format of model input.
pared with the soft unit. OpenCV (50) is used for reading, cutting, plotting, and stitch-
Once the material properties and geometry are defined, the ing images.
composite is put under compressive loading and unloading for one
cycle in Abaqus (movie S2). Element “CPE4R” with the global size Image comparison
of 6 is used for composites during the calculation. In terms of To compare filed images, L2 norm is calculated to determine the
boundary conditions, the middle line of the composite is set immo- difference between two images. Given two images P1 and P2, L2
bile along the x direction due to the symmetry. The loading and norm between these two images are defined as
unloading are realized by moving two symmetric analytical rigid
shells. During the loading process, the magnitude of compressive
strain is 10%. The contact type between composite and 2D shell is Norm _ L2 = ​
​ ​ ​​ ​ P1(i, j ) ‐P2(i, j ) ∣​​ 2​​
∑ i,j∣ (1)
surface-to-surface contact using kinetic contact method. For either
loading or unloading, 2D shell is moving in a time period of 500 In Fig. 4D, the L2 norm map is used instead of the L2 norm for
(unit consistent with table S1). Within the time period, the ampli- comparing high-resolution field image. To derive a L2 norm map,
tude of displacement of the rigid shell follows “Smooth step” type we first slice images of 512 × 512 into 128 × 128 patches each with a
embedded in Abaqus. After loading and unloading, images of strain size of 4 × 4. Second, for each pair of corresponding patches in two
or stress fields are postprocessed and collected for the dataset. images, L2 norm is calculated. As a result, we end up with a matrix
of 128 × 128 with each element corresponding to the difference be-
Data distribution tween a pair of patches in two images. The matrix is named as L2
According to our graphical representation, there are 232 possible norm map.
combinations in total. Among all potential geometries, we randomly When evaluating the performance of ML models, we calculate
select 2000 data to constitute the dataset. The distribution of all the L2 norm or the L2 norm map between the ground truth and
2000 data is plotted over the ratio of soft units (fig. S1A). As the prediction (“predicted contour”). The ground truth is also com-
figure indicates, the distribution generally fits a Gaussian curve with pared with two field images for direct reference. These two field im-
the mean value lying at approximately 0.5. The curve reveals a uni- ages are labeled as “random contour” and “clean contour” (fig. S2).
formly distributed dataset without skewed features. As a consequence, “Random contour” is a contour image randomly selected from test
we suggest that our dataset is representative of all possible data in dataset. “Clean contour” is a contour image without any deforma-
the search space. With 2000 data randomly generated, we further tion and field (all values across the image is 0).

Yang et al., Sci. Adv. 2021; 7 : eabd7416 9 April 2021 7 of 10

SCIENCE ADVANCES | RESEARCH ARTICLE

In addition, the relative errors comparing the residual stress values Nonsquare features
of ML predictions and the FEM results are computed simply follow- Hexagonal and triangular representations are used to testify wheth-
ing the expression er our model can be applied to composites with different shapes.
The sizes of composites in hexagonal and triangular representations

​S​ FEM​​
 ​​ |
​S​ ML​​  − ​S​ FEM​​
​​RE = ​​​ ​ ─  
​​  ×  100%​​ | (2)
are 312 × 315 and 320 × 332 (x × y). At the boundaries of composite,
both hexagonal and triangular units are cut half. The field images
for both cases were generated after one-cycle compressive loading
where SML is the average von Mises stress from ML predictions, and unloading. The magnitude of compressive strain is 10% for the
and SFEM is the average von Mises stress from FEM simulation results. hexagonal representation and 5% for the triangular representation.
The reason why a smaller strain is used for triangular cases is to
DL approach avoid singularity at the edges of triangles when FEM simulations
The ML calculations are performed using TensorFlow (51), a general are carried out.
purpose ML framework. A GAN is implemented for translating
composite geometries to mechanical fields (44). GAN is a type of Distinct loading conditions
deep neural network consisting of a Generator and a Discriminator To explore the capacity of our ML model to predict strain or stress
based on game theory (43). Particularly in our model, U-Net is the fields under various loads, we use two mechanical tests, nanoinden-
Generator and PatchGAN is the Discriminator (44). U-Net is a deep tation and compression. The general setup of those mechanical tests
neural network model that has been used for biomedical image seg- using FEM is the same as the setup mentioned above (see the
mentation in earlier work (27). Here, we use U-Net to generate fake “FEM” section). However, we do not unload the composite after
field images based on composite geometries. PatchGAN evaluates loading. The reason is that the field image after unloading in
the generated field images from U-Net by comparing with real field nanoindentation contains very little information as the composite
images. The model was trained using a single NVIDIA Quadro RTX has most of its region stress free. A compressive strain with magni-
4000 with 8-gigabyte memory. The numbers of training epochs range tude of 5% is used again to again avoid singularity during the
from 150 to 400 in different cases to achieve convergence, and the nanoindentation simulation. To include loading condition in images,
batch size of the training is 1. Although GANs are notorious for the green lines are used to represent analytical rigid bodies, which
training instability, cGANs are relatively more stable with the con- induce deformation (Fig. 4D).
strain of the input labels (52). The numbers of training epochs were As for the nanoindentation test, rather than a straight-line rigid
carefully tuned so that the training process is not too short for the body, a spherical indenter is used to implement loading (movie S3).
generator to learn enough information and also not too long that The indenter is laid at the top center of the composite with radius of

Downloaded from https://www.science.org on February 10, 2022

the discriminator always evaluates the generated images to be fake 40 (unit consistent with table S1). In addition to the boundary con-
(52). We checked the predictions of our ML model on the test set to dition mentioned in the “FEM” section, the motion at the bottom of
determine appropriate training epochs. To show the stability of our the composite is disabled along y direction for both compression
training process, we provided two datasets (datasets S1 and S2), and nanoindentation. With consistent boundary conditions, we are
each containing predicted images over 400 geometries of the test set able to combine more than one loads in one mechanical test.
for multiple fields.
The Generator U-Net consists of two components encoder and Composite with high-resolution geometry
decoder with skip connections between mirrored layers (27). Both To capture more complex patterns, additional composite designs
encoder and decoder contain eight layers including input and out- with high-resolution geometry (32 × 32) are generated and added to
put layers. The complete architecture is shown in table S4. The the dataset. Keeping the length of composites consistent, smaller
trainable number of parameters in total is approximately 5,000,000. square units with lengths of 10 are used. All settings for FEM are the
The Generator loss function is defined as same as those considering low-resolution geometry (8 × 8), except
for two minor differences. First, the element size decreases from 6
​gen _ loss = gan _ loss + LAMBDA × L1 _ loss ​ (3) to 2 for calculation accuracy and convergence. Second, the magni-
tude of compressive strain is set to 5% instead of 10%. The reason to
Here, gan_loss is a sigmoid cross-entropy loss of the generated use a smaller strain is to avoid singularity when running FEM as the
images and an array of ones, and L1_loss is the mean absolute error size of soft/brittle units is smaller.
between the generated image and the target image. LAMBDA is For the GAN model, we use original images with a size of
proportion coefficient that is equal to 100. 512 × 512 as the input as the geometries are more complicated with
The Discriminator PatchGAN consists of eight layers with ap- smaller block units. As a consequence, one extra convolutional layer
proximately 3,000,000 trainable weights. The complete architecture is added ahead of all layers in Generator to “downsample” images
is shown in table S3. PatchGAN evaluates the generated field imag- from 512 × 512 to 256 × 256. Similarly, there is one more layer after
es by classifying individual patches (slice the image into 30  ×  30 all layers in Discriminator to “upsample” images from 256 × 256 to
patches in our case) in the image as “real” or “fake.” The Discrimi- 512 × 512.
nator loss function is defined as
Analogous stress fields prediction around cracks
​
gen _ loss = real _ loss + generated _ loss ​ (4)
Soft materials act like cracks when being surrounded by brittle ma-
Here, real_loss is a sigmoid cross-entropy loss of real images and terials as they create large stress concentration and take majority of
an array of ones, and generated_loss is a sigmoid cross-entropy loss deformations. On the basis of this idea, we generate composites
of the generated images and an array of zeros. with soft materials in narrow shapes and sharp edges embedded in

Yang et al., Sci. Adv. 2021; 7 : eabd7416 9 April 2021 8 of 10

SCIENCE ADVANCES | RESEARCH ARTICLE

brittle materials. We use the model that is trained for high-resolution 28. S. L. Franjou, M. Milazzo, C.-H. Yu, M. J. Buehler, Sounds interesting: Can sonification help
geometry to predict the stress field around soft materials. Two types us design new proteins? Expert Rev. Proteomics 16, 875–879 (2019).
29. C.-H. Yu, M. J. Buehler, Sonification based de novo protein design using artificial
of cracks with different shapes and positions are investigated (fig. S4).
intelligence, structure prediction, and analysis using molecular modeling. APL Bioeng. 4,
016108 (2020).
SUPPLEMENTARY MATERIALS 30. C. H. Yu, Z. Qin, F. J. Martin-Martinez, M. J. Buehler, A self-consistent sonification method
Supplementary material for this article is available at http://advances.sciencemag.org/cgi/ to translate amino acid sequences into musical compositions and application in protein
content/full/7/15/eabd7416/DC1 design using artificial intelligence. ACS Nano 13, 7471–7482 (2019).
31. R. Gómez-Bombarelli, J. N. Wei, D. Duvenaud, J. M. Hernández-Lobato, B. Sánchez-
Lengeling, D. Sheberla, J. Aguilera-Iparraguirre, T. D. Hirzel, R. P. Adams, A. Aspuru-Guzik,
REFERENCES AND NOTES Automatic chemical design using a data-driven continuous representation of molecules.
1. K. K. Chawla, Composite Materials: Science and Engineering (Springer Science & Business
ACS Cent. Sci. 4, 268–276 (2018).
Media, 2012).
32. D. K. Johnson, J. Karanicolas, Ultra-high-throughput structure-based virtual screening
2. A. R. Studart, Biological and bioinspired composites with spatially tunable heterogeneous
for small-molecule inhibitors of protein-protein interactions. J. Chem. Inf. Model. 56,
architectures. Adv. Funct. Mater. 23, 4423–4436 (2013).
399–411 (2016).
3. C. T. Chen, F. J. Martin-Martinez, S. Ling, Z. Qin, M. J. Buehler, Nacre-inspired design
33. K. T. Schütt, F. Arbabzadah, S. Chmiela, K. R. Müller, A. Tkatchenko, Quantum-chemical
of graphene oxide–polydopamine nanocomposites for enhanced mechanical properties
insights from deep tensor neural networks. Nat. Commun. 8, 13890 (2017).
and multi-functionalities. Nano Futur. 1, 011003 (2017).
34. G. R. Schleder, A. C. M. Padilha, C. M. Acosta, M. Costa, A. Fazzio, From DFT to machine
4. M. J. Buehler, H. Gao, Dynamical fracture instabilities due to local hyperelasticity at crack
learning: Recent approaches to materials science–A review. J. Phys. Mater. 2, 032001
tips. Nature 439, 307–310 (2006).
(2019).
5. K. K. Chawla, Ceramic Matrix Composites (Springer Science & Business Media, 2013).
35. K. T. Schütt, P. J. Kindermans, H. E. Sauceda, S. Chmiela, A. Tkatchenko, K. R. Müller,
6. C. H. Yu, M. W. Bird, C. W. Huang, C. S. Chen, Y. F. Gao, K. W. White, C. H. Hsueh,
SchNet: A continuous-filter convolutional neural network for modeling quantum
Micromechanics modeling of creep fracture of zirconium diboride-silicon carbide
interactions. Adv. Neural Inf. Process. Syst. 2017, 992–1002 (2017).
composites at 1400-1700°C. J. Eur. Ceram. Soc. 34, 4145–4155 (2014).
36. J. Wang, S. Olsson, C. Wehmeyer, A. Pérez, N. E. Charron, G. De Fabritiis, F. Noé,
7. Z. M. Huang, Y. Z. Zhang, M. Kotaki, S. Ramakrishna, A review on polymer nanofibers by
C. Clementi, Machine learning of coarse-grained molecular dynamics force fields. ACS
electrospinning and their applications in nanocomposites. Compos. Sci. Technol. 63,
Cent. Sci. 5, 755–767 (2019).
2223–2253 (2003).
37. J. Behler, M. Parrinello, Generalized neural-network representation of high-dimensional
8. X. Xia, A. Afshar, H. Yang, C. M. Portela, D. M. Kochmann, C. V. Di Leo, J. R. Greer,
potential-energy surfaces. Phys. Rev. Lett. 98, 146401 (2007).
Electrochemically reconfigurable architected materials. Nature 573, 205–213 (2019).
38. F. Noé, A. Tkatchenko, K.-R. Müller, C. Clementi, Machine learning for molecular
9. M. S. Pham, C. Liu, I. Todd, J. Lertthanasarn, Damage-tolerant architected materials
simulation. Annu. Rev. Phys. Chem. 71, 361–390 (2020).
inspired by crystal microstructure. Nature 565, 305–311 (2019).
39. G. X. Gu, C. T. Chen, D. J. Richmond, M. J. Buehler, Bioinspired hierarchical composite
10. F. Barthelat, Z. Yin, M. J. Buehler, Structure and mechanics of interfaces in biological
design using machine learning: Simulation, additive manufacturing, and experiment.
materials. Nat. Rev. Mater. 1, 16007 (2016).
Mater. Horizons. 5, 939–945 (2018).
11. H. Gao, B. Ji, I. L. Jäger, E. Arzt, P. Fratzl, Materials become insensitive to flaws at
40. F. Martínez-Martínez, M. J. Rupérez-Moreno, M. Martínez-Sober, J. A. Solves-Llorens,
nanoscale: Lessons from nature. Proc. Natl. Acad. Sci. U.S.A. 100, 5597–5600 (2003).
D. Lorente, A. J. Serrano-López, S. Martínez-Sanchis, C. Monserrat, J. D. Martín-Guerrero, A

Downloaded from https://www.science.org on February 10, 2022

12. P. K. Mallick, Fiber-Reinforced Composites: Materials, Manufacturing, and Design (CRC press,
finite element-based machine learning approach for modeling the mechanical behavior
2007).
of the breast tissues under compression in real-time. Comput. Biol. Med. 90, 116–124 (2017).
13. G. X. Gu, M. Takaffoli, A. J. Hsieh, M. J. Buehler, Biomimetic additive manufactured
41. F. E. Bock, R. C. Aydin, C. J. Cyron, N. Huber, S. R. Kalidindi, B. Klusemann, A review
polymer composites for improved impact resistance. Extrem. Mech. Lett. 9, 317–323
of the application of machine learning and data mining approaches in continuum
(2016).
materials mechanics. Front. Mater. 6, 110 (2019).
14. Z. Quan, A. Wu, M. Keefe, X. Qin, J. Yu, J. Suhr, J. H. Byun, B. S. Kim, T. W. Chou, Additive
42. Z. Yang, Y. C. Yabansu, D. Jha, W. K. Liao, A. N. Choudhary, S. R. Kalidindi, A. Agrawal,
manufacturing of multi-directional preforms for composites: Opportunities
Establishing structure-property localization linkages for elastic deformation of three-
and challenges. Mater. Today 18, 503–512 (2015).
dimensional high contrast composites using deep learning approaches. Acta Mater. 166,
15. P. Parandoush, D. Lin, A review on additive manufacturing of polymer-fiber composites.
335–345 (2019).
Compos. Struct. 182, 36–53 (2017).
43. S. Mahdizadehaghdam, A. Panahi, H. Krim, Sparse generative adversarial network, in
16. G. X. Gu, L. Dimas, Z. Qin, M. J. Buehler, Optimization of composite fracture properties:
Proceedings of 2019 International Conference on Computer Vision Workshop (ICCVW) (Seoul,
Method, validation, and applications. ASME J. Appl. Mech. Trans. 83, 071006 (2016).
Korea, 27 to 28 October 2019), pp. 3063–3071.
17. R. Chawla, S. Sharma, A molecular dynamics study on efficient nanocomposite formation
44. P. Isola, J. Y. Zhu, T. Zhou, A. A. Efros, Image-to-image translation with conditional
of styrene–butadiene rubber by incorporation of graphene. Graphene Technol. 3, 25–33
adversarial networks, in Proceedings of the 30th IEEE Conference on Computer Vision
(2018).
Pattern Recognition, (CVPR 2017) (Honolulu, USA, 21 TO 26 July 2017), pp. 5967–5976.
18. T. Kairn, P. J. Daivis, I. Ivanov, S. N. Bhattacharya, Molecular-dynamics simulation of model
45. L. S. Dimas, M. J. Buehler, Modeling and additive manufacturing of bio-inspired
polymer nanocomposite rheology and comparison with experiment. J. Chem. Phys. 123,
composites with tunable fracture mechanical properties. Soft Matter 10, 4436–4442
194905 (2005).
(2014).
19. G. X. Gu, C. Chen, M. J. Buehler, De novo composite design based on machine learning
46. G. X. Gu, I. Su, S. Sharma, J. L. Voros, Z. Qin, M. J. Buehler, Three-dimensional-printing
algorithm. Extrem. Mech. Lett. 18, 19–28 (2018).
of bio-inspired composites. J. Biomech. Eng. 138, 021006 (2016).
20. C.-H. Yu, Z. Qin, M. J. Buehler, Artificial intelligence design algorithm for nanocomposites
47. G. X. Gu, S. Wettermark, M. J. Buehler, Algorithm-driven design of fracture resistant
optimized for shear crack resistance. Nano Futur. 3, 035001 (2019).
composite materials realized through additive manufacturing. Addit. Manuf. 17, 47–54
21. K. T. Butler, D. W. Davies, H. Cartwright, O. Isayev, A. Walsh, Machine learning
(2017).
for molecular and materials science. Nature 559, 547–555 (2018).
48. Y. Masso-Moreu, N. J. Mills, Impact compression of polystyrene foam pyramids. Int.
22. Y. Lecun, Y. Bengio, G. Hinton, Deep learning. Nature 521, 436–444 (2015).
J. Impact Eng. 28, 653–676 (2003).
23. T. W. Hughes, I. A. D. Williamson, M. Minkov, S. Fan, Wave physics as an analog recurrent 49. V. S. Deshpande, N. A. Fleck, Isotropic constitutive models for metallic foams. J. Mech.
neural network. Sci. Adv. 5, eaay6946 (2019). Phys. Solids 48, 1253–1283 (2000).
24. Z. Qin, Q. Yu, M. J. Buehler, Machine learning model for fast prediction of the natural 50. G. Bradski, The OpenCV Library. Dr. Dobb’s J. Softw. Tools 120, 122–125 (2000).
frequencies of protein molecules. RSC Adv. 10, 16607–16615 (2020). 51. J. J. Weinman, A. Lidaka, S. Aggarwal, TensorFlow: Large-scale machine learning. GPU
25. Z. Jensen, E. Kim, S. Kwon, T. Z. H. Gani, Y. Román-Leshkov, M. Moliner, A. Corma, Comput. Gems Emerald Ed. 2011, 277–291 (2011).
E. Olivetti, A machine learning approach to zeolite synthesis enabled by automatic 52. I. Goodfellow, NIPS 2016 Tutorial: Generative Adversarial Networks (2016); http://arxiv.org/
literature data extraction. ACS Cent. Sci. 5, 892–899 (2019). abs/1701.00160.
26. M. Aykol, V. I. Hegde, L. Hung, S. Suram, P. Herring, C. Wolverton, J. S. Hummelshøj,
Network analysis of synthesizable materials discovery. Nat. Commun. 10, 2018 Acknowledgments: We acknowledge support from the Google Cloud platform and MIT
(2019). Quest for providing computational resources. Funding: We acknowledge support by the
27. O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image Army Research Office (W911NF1920098), the Office of Naval Research (N000141612333), and
segmentation. Lect. Notes Comput. Sci. 9351, 234–241 (2015). AFOSR-MURI (FA9550-15-1-0514). Author contributions: M.J.B., Z.Y., and C.-H.Y. conceived

Yang et al., Sci. Adv. 2021; 7 : eabd7416 9 April 2021 9 of 10

SCIENCE ADVANCES | RESEARCH ARTICLE

the idea. Z.Y., C.-H.Y., and M.J.B. developed the model and carried out the simulations. Z.Y. and Submitted 9 July 2020
C.-H.Y. curated the training and testing data. M.J.B. supervised the project, analyzed the Accepted 24 February 2021
results, and interpreted it with Z.Y. and C.-H.Y. Z.Y., C.-H.Y., and M.J.B. wrote the manuscript. Published 9 April 2021
Competing interests: The authors declare that they have no competing interests. Data and 10.1126/sciadv.abd7416
materials availability: All data needed to evaluate the conclusions in the paper are present in
the paper and/or the Supplementary Materials. Additional data related to this paper may be Citation: Z. Yang, C.-H. Yu, M. J. Buehler, Deep learning model to predict complex stress and
requested from the corresponding author. strain fields in hierarchical composites. Sci. Adv. 7, eabd7416 (2021).

Downloaded from https://www.science.org on February 10, 2022

Yang et al., Sci. Adv. 2021; 7 : eabd7416 9 April 2021 10 of 10

 Deep learning model to predict complex stress and strain fields in hierarchical
composites
Zhenze YangChi-Hua YuMarkus J. Buehler

Sci. Adv., 7 (15), eabd7416. • DOI: 10.1126/sciadv.abd7416

View the article online

https://www.science.org/doi/10.1126/sciadv.abd7416
Permissions
https://www.science.org/help/reprints-and-permissions

Downloaded from https://www.science.org on February 10, 2022

Use of think article is subject to the Terms of service

Science Advances (ISSN 2375-2548) is published by the American Association for the Advancement of Science. 1200 New York Avenue
NW, Washington, DC 20005. The title Science Advances is a registered trademark of AAAS.
Copyright © 2021 The Authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim
to original U.S. Government Works. Distributed under a Creative Commons Attribution NonCommercial License 4.0 (CC BY-NC).