Acta Materialia 190 (2020) 105
112

Contents lists available at ScienceDirect

Acta Materialia
journal homepage: www.elsevier.com/locate/actamat

Full length article

A machine learning approach to fracture mechanics problems

Xing Liua, Christos E. Athanasioua, Nitin P. Padturea, Brian W. Sheldona,*, Huajian Gaoa,b,*
a
School of Engineering, Brown University, Providence, RI 02912, USA
b
College of Engineering, College of Science, Nanyang Technological University, Singapore

A R T I C L E I N F O A B S T R A C T

Article History: Analytical and empirical solutions to engineering problems are usually preferred because of their conve-
Received 3 February 2020 nience in applications. However, they are not always accessible in complex problems. A new class of solu-
Revised 9 March 2020 tions, based on machine learning (ML) models such as regression trees and neural networks (NNs), are
Accepted 12 March 2020
proposed and their feasibility and value are demonstrated through the analysis of fracture toughness meas-
Available online 18 March 2020
urements. It is found that both solutions based on regression trees and NNs can provide accurate results for
the specific problem, but NN-based solutions outperform regression-tree-based solutions in terms of their
Keywords:
Machine learning simplicity. This example demonstrates that ML solutions are a major improvement over analytical and
Analytical methods empirical solutions in terms of both reliable functionality and rapid deployment. When analytical solutions
Mechanical properties testing are not available, the use of ML solutions can overcome the limitations of empirical solutions and substan-
Fracture tially change the way that engineering problems are solved.
© 2020 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.

1. Introduction characterization, taking fracture toughness measurements of micro-

fabricated brittle ceramic microcantilevers as a relevant example.
Engineers often seek analytical solutions for simplicity and reli-
ability, which bring great convenience to engineering applications
2. Experimental
such as materials characterization, structural analysis and design.
However, analytical solutions cannot always be obtained. An accept-
This investigation is based on a mechanical test that is used to
able compromise is the empirical solution which relies on engineers’
measure the mode-I fracture toughness, KIC, in small specimens
understanding and generalization of experimental and numerical
[4
7]. In this method, a pentagonal cross-section microcantilever is
data. For example, in ASTM standard C1421 for determination of frac-
cut with focused ion beam (FIB) micromachining in Helios NanoLab
ture toughness of advanced ceramics, empirical solutions are pro-
450 system (FEI, Oregon, USA), as shown in Fig. 1b. To induce a well-
vided to evaluate the plane-strain stress intensity factor at the crack
defined controlled fracture event, a sharp pre-notch is milled at a dis-
tip. Both analytical and empirical solutions are known for their rapid
tance, L0, from the fixed end of the cantilever. The pre-notch is
deployment and reliability. However, some complicated engineering
required to be sufficiently sharp to guarantee the validity of the frac-
problems may involve a nonlinear and complex relationship among
ture toughness measurement. Therefore, a two-step strategy is
higher-dimensional data, and empirical solutions cannot be readily
adopted, i.e., a coarse cut with a moderate current followed by a sec-
obtained. In the case where neither analytical nor empirical solutions
ond finer cut with a lowest possible current, 1 pA. This strategy
are feasible the following question can be raised: is there any other
restricts the notch radius well below 50 nm and eliminates the arti-
possible way to obtain a solution? Machine learning (ML) algorithms
facts caused by ion-implantation damage around the notch tip during
can be informed directly by experiments and simulations [1
3], and
FIB milling. In this way, the notch tip is sharp enough to behave like
thus provide “machine learning solutions”. These ML solutions are a
an ideal crack and ensure the validity of the measurement [8,9]. After
promising substitute for analytical and empirical solutions if they can
the microcantilever fabrication, a nanoindenter (Hysitron TI 900, Tri-
provide rapid and accurate results. In this context, the initial study
boindenter, Minneapolis, USA) equipped with a Berkovich tip is used
presented here, summarized in Fig. 1, uses ML solutions for materials
to apply a controlled load at the free end of the cantilever and the
load-displacement response is recorded. Because the fracture of the
specimen geometry is inherently unstable, a catastrophic crack
* Corresponding author. growth would be observed, i.e., a peak load followed by a displace-
E-mail addresses: [email protected] (X. Liu),
[email protected] (C.E. Athanasiou),
ment burst. Therefore, the fracture toughness is correlated with the
[email protected] (N.P. Padture), [email protected] (B.W. Sheldon), critical load at the onset of fracture. An accurate evaluation of KIC
[email protected] (H. Gao). requires theoretical solutions or numerical simulations.

https://doi.org/10.1016/j.actamat.2020.03.016
1359-6454/© 2020 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.
106 X. Liu et al. / Acta Materialia 190 (2020) 105
112

Fig. 1. ML solutions to engineering problems. (a) Analytical solutions are developed based on a physical understanding of the problem, which can be inaccessible in complicated
cases. An acceptable compromise that has been used traditionally relies on fitting abundant experimental and/or numerical data to an empirical solution. This approach can have
limited accuracy when dealing with a nonlinear and complex relationship among high-dimensional data. Therefore, machine learning (ML) solutions are proposed to overcome this
weakness and provide accurate results rapidly. (b) The specific engineering problem addressed in this work: determination of fracture toughness by loading (using a nanoindenter)
a pre-notched pentagonal cross-section microcantilever at its end. The microcantilever is milled out of the bulk material, whose dimensions are {w, b, a, L0, L1}. (c) The ratio of the
crack tip plane-strain stress intensity factor, KI, to the indentation load, P, is evaluated from empirical solutions [4,5], and compared to finite element method (FEM) simulations and
ML solutions for different cantilever lengths, L1, and fw; b; a; L0 g ¼ f5:20; 4:80; 0:90; 1:00g mm. The ML solution can achieve comparable accuracy to FEM results under conditions
where the empirical solution is inaccurate.

In this study, three polycrystalline silicon specimens ( » 1 mm practical considerations related to the use of the FIB instrument and
grain size, American Elements, California, USA) were tested as by the need to obtain a well-defined controlled fracture event during
reference samples. A scanning electron microscope (SEM) image of a loading. With the geometries that are typically used, it is impossible
microcantilever is shown in Fig. 2a and the fracture toughness meas- to derive analytical or empirical solutions that are highly accurate
urements are summarized in Fig. 2b. The specimen sizes are suffi- over the full range of relevant sample dimensions. An early attempt
ciently larger than the grain sizes, thus a homogeneous polycrystalline [4,5], based on an oversimplified functional form and limited numeri-
microstructure is sampled during the measurement. cal results, fails as it deviates significantly from the results of finite
element method (FEM), whose accuracy is guaranteed as shown in
3. Theory/calculation Fig. 1c. It is certainly possible to obtain accurate solutions by running
an FEM simulation for each specimen. However, this is time-consum-
3.1. Problem description ing and impractical if there are a large number of specimens with var-
iations in the specimen dimensions, which are essential to capture
An accurate analysis of the fracture toughness measurement is the statistical distribution of fracture toughness. In this context, a ML
based on linear elastic fracture mechanics and requires a solution solution can provide a valuable tool to accelerate the implementation
that describes the plane-strain stress intensity factor at the crack tip of a specific fracture toughness test. More generally, this provides a
for a given load. The exact microcantilever geometry is limited by tool with both portable deployment (comparable to analytical/

Fig. 2. Fracture toughness tests on polycrystalline silicon specimens. (a) The scanning electron microscope (SEM) image shows the exact geometry of the polysilicon specimens.
(b) Fracture toughness values of these three specimens were evaluated from empirical solutions, FEM simulations and ML solutions for comparison. The dimensions of the polysili-
con specimens are listed in Table A.1.
 X. Liu et al. / Acta Materialia 190 (2020) 105
112 107

empirical solutions) and high accuracy (comparable to FEM simula- More than 10,000 full integration elements (C3D20) were employed
tions). For the example investigated here, the ML solution should with the mesh being refined near the crack tip. The average wall time
capture the relationship among the plane-strain stress intensity fac- for each simulation was 120 s when running in parallel on 16 CPUs.
¢ 120 ¢ 16
tor, KI, at the notch tip, the indentation load, P, and the specimen The generation of the whole data set took roughly 439;956
320 ¢ 60 ¢ 60 ¢ 24 31
dimensions, {w, b, a, L0, L1}. days, with 20 jobs running simultaneously on 320 CPUs. The elapsed
time can be further shortened if more CPUs are utilized.
There is a wide variety of ML models and algorithms, and one
3.2. Data set preparation based on FEM simulations
needs to find the most appropriate one for their specific problems.
After screening the popular ML models, it is found that regression
Developing a ML solution requires a data set that includes desired
trees (RTs) and their ensembles, neural networks (NNs) are most suit-
inputs and outputs for a large number of samples. These data are
able for this problem.
then used to train appropriate ML models, which can evolve to opti-
mize performance. The quality and quantity of the data dictate the
accuracy that can be achieved. Therefore, it is of great importance to
4. Results
build a good data set, comprising a combination of both input and
target data. The way that these data are defined can greatly affect the
4.1. Regression-tree-based solutions
ML process. The fracture toughness measurements used here are
assumed to strictly follow the assumptions of linear-elastic fracture
The tree-based method for regression can capture the highly non-
mechanics, such that the calculation of the plane-strain stress inten-
linear relationship between the input variables, x ¼ ðx1 ; x2 ; x3 ; x4 Þ,
sity factor, KI, is a boundary value problem in linear elasticity. The
and the target variable, y. The regression tree is built through dividing
boundary conditions of the problem imply that the crack tip plane-
the input parametric space into distinct, non-overlapping regions
strain stress intensity factor, KI, is directly proportional to the inden-
known as leaf nodes. Each leaf node is labeled with the average value
tation load, P, and the ratio between them depends only on the speci-
of targets of the training samples that fall into the region, as shown in
men dimensions, {w, b, a, L0, L1}, and is independent of elastic
Fig. 3. A given input flows from the root node, through the internal
properties (Young’s modulus, E, and Poisson’s ratio, n). Dimensional
branch nodes, into the terminal leaf node, and produces a predicted
analysis indicates that five independent dimensionless variables,
n o target variable (by the label of the leaf node). An optimized CART
a w L0 L1 KI
; ; ; ;
b b b b PL a0:5 b
1 w
2
, are relevant. Based on this, the following algorithm [14] based on information gain is implemented in the
1

structure of the ML solutions is employed: open-source package Scikit-learn [15] to grow the regression trees.
8   The complexity and size of a grown regression tree is controlled by
< Input variables : x ¼ ðx1 ; x2 ; x3 ; x4 Þ ¼ a ; w ; L0 ; L1
>
the maximum tree depth and the minimum leaf size. The tree depth
b b b b ð1Þ
> KI represents the maximum number of edges from the root node to the
: Target variable : y ¼
0:5
PL1 a b w
1
2 leaf node and is limited to a moderate interval 4»8. The leaf size rep-
resents the number of samples required for each leaf node and a rea-
Based on the specimen sizes used for typical measurements, the
sonable lower bound is set to 50. Using these parameters overfitting
parameter space of the input variables is defined as:
can be effectively eliminated.
a w L0 L1 The structure and performance of the regression tree solutions are
2 ½0:1; 0:8; 2 ½1:0; 3:0; 2 ½0:1; 0:4; 2 ½2:0; 5:0: ð2Þ
b b b b summarized in Fig. 3 and Table A.2. The accuracy of the solutions is
In order to fully explore the parameter space of input variables, a measured by the maximum value of the absolute percentage error:
grid-search strategy is adopted when the data  set for ML is generated.  
e ¼ j ytrue
ypred =ytrue j  100%; ð3Þ
Each input domain xi in ðx1 ; x2 ; x3 ; x4 Þ ¼ ab ; wb ; Lb0 ; Lb1 , is discretized
into mi uniform intervals and a grid of all possible input variables is over all the samples in the data set, where ypred is the target value
then constructed inside the parameter space. This gives M ¼ ðm1 þ 1Þ predicted by the ML solution and ytrue is the target value of the data
ðm2 þ 1Þðm3 þ 1Þðm4 þ 1Þ different samples. A dense grid of ðm1 ; m2 ; set which is initially obtained from FEM simulations. As a limit is set
m3 ; m4 Þ ¼ ð35; 100; 10; 10Þ is then used to generate a data set for ML on the minimum leaf size, the accuracy of these regression trees is
with M ¼ 439; 956 samples. Since the target variable in this problem saturated before reaching an acceptable low value, as shown in Fig. 3
is continuous and bounded over the whole input parametric space, and Table A.2.
the accuracy of the ML solutions over the continuous space can be Although a single regression tree is not powerful enough to
estimated in a reliable manner by sampling these discrete points. address the complex problem, it can be used as a building block
FEM is used to evaluate the target variable y for each sample. The to construct a better model with stronger learning ability, i.e., a
direct domain J-integral [10,11] and elastic compliance [12] are the regression tree ensemble, as shown in Fig. 3. Given an input
two conventional methods to evaluate the energy release rate, G, and
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi of x ¼ ðx1 ; x2 ; x3 ; x4 Þ, each regression tree in the ensemble provides a
the plane-strain stress intensity factor, KI ¼ EG=ð1
n2 Þ, at the crack response which is then synthesized into a final prediction of the tar-
tip. This first evaluates G from the direct domain J-integral. The elastic get variable y. The Gradient Tree Boosting algorithm [16,17] is imple-
compliance method calculates the rate of change of the compliance mented to build the regression tree ensembles which allows for the
with crack extension for a given indentation load, and hence the rate optimization of arbitrary differentiable loss functions.
of potential energy loss, which is equivalent to G. Both methods pro- A series of gradient-boosted regression trees (GBRT) with least
KI
vide an accurate value of the target, y ¼ PL1 a0:5 b
1 w
2 , for a given input, squares loss are constructed with different numbers and sizes of the
a w L0 L1
x ¼ b ; b ; b ; b . However, the computational cost of the J-integral basic regression trees, and their structures and performance are sum-
method is significantly lower than the elastic compliance method marized in Fig. 3 and Table A.2. Interestingly, the accuracy of predic-
since the latter requires multiple simulations for different crack tion is greatly improved. An accuracy of less than 5% of relative error
lengths. It is crucial to improve the computational efficiency of the can be achieved by combining 512 regression trees of depth 6.
data generation, especially when the size of the data set is huge, and Thus, the learning capability of the regression tree ensemble
thus the J-integral method is clearly the best choice. This was imple- method in producing an accurate ML solution is demonstrated. It is
mented through user-defined subroutines in FEAP [13]. An encastre worth mentioning that a regression tree ensemble typically contains
boundary condition was applied to the fixed end of the cantilever a large number of basic trees which drastically increases the com-
and a displacement-controlled boundary condition to the free end. plexity of the corresponding ML solution.
108 X. Liu et al. / Acta Materialia 190 (2020) 105
112

Fig. 3. Regression-tree-based solutions. (a) A regression tree is a hierarchically organized structure with each branch node splitting the input parametric space, x. Eventually the
parametric space is divided into distinct and non-overlapping regions, i.e., the leaf nodes at the end of each branch. The value of target variable, y, is then predicted by the region
that the input falls in. (b) A regression-tree ensemble combines a set of weak regression trees and make predictions by aggregating responses from each single regression tree. (c)
Ensemble methods, such as GBRT, can achieve much higher accuracy (i.e., smaller absolute percentage error) than a single regression tree. (For interpretation of the references to
colour in this figure legend, the reader is referred to the web version of this article.)

4.2. Neural-network-based solutions while keeping a small portion for validation and test. Therefore, the
data set is split into training data set (70%), validation data set (15%),
NNs are computing systems with interconnected nodes inspired by and test data set (15%). By collecting and reviewing the accuracy of
biological neural networks in the human brain. NNs are known for their these three data sets during the training process as shown in Fig. A.1, it
power in approximating the nonlinear relationships among high- is demonstrated that all the NNs are converged and not overfitted.
dimensional data (inputs and outputs) according to the universal The structure and performance of the NN-based solutions are
approximation theorem [18]. A simple NN consists of an input layer, an summarized in Fig. 4 and Table A.3. It turns out that solutions with
output layer, and hidden layers in-between. These layers are intercon- high accuracy can be achieved even with simple NNs: a NN with only
nected to form a network architecture, as shown in Fig. 4. Feedforward 29 neurons (4/16/8/1) can produce an accuracy of less than 5% of rela-
NN is adopted, in which each node in one layer is connected to all the tive error. As presented in Fig. 4, one can choose appropriate NN-
nodes in the adjacent layers and information is fed between layers only based solutions to tune the accuracy according to the application
in the forward direction. As the input data x ¼ ðx1 ; x2 ; x3 ; x4 Þ is fed into needs. Surprisingly, by using a large data set, i.e., M ¼ 439; 956, even
the NN through the input layer, each node in the next hidden layer pro- with a small number of neurons, i.e., 29, the strong learning ability of
cesses data from the input layer and feeds the next layer through an a NN is demonstrated (Fig. 4).
activation function. Finally, the output layer collects data from the last
hidden layer and produce the target data, y. The architecture of a NN is 4.3. Portable deployment of ML solutions
controlled by the number of hidden layers and the number of nodes in
each hidden layer. Simple NNs with 1 or 2 hidden layers with the recti- It is concluded that both the regression-tree-based solutions and
fied linear activation function (ReLU), i.e., single-layer and multilayer NN-based solutions can provide accurate results for the specific prob-
perceptrons, are employed in this study. lem. However, NN-based solutions outperform regression-tree-based
Due to the difficulty in their learning process, NNs must be comple- solutions in terms of their simplicity (Fig. 3 and Fig. 4). Indeed, a regres-
mented by a robust ML algorithm, such that, they can learn and evolve sion-tree-based solution with less than 5% of relative error involves
continuously as new data is fed to them. Nadam algorithm [19] with hundreds of regression trees and each regression tree consists of hun-
Log-Cosh loss function is adopted for the training of NNs using the dreds of nodes. Therefore, it is challenging to obtain an overall view of
open-source platform TensorFlow r2.0 [20]. To fully exploit the learning the redundant structure of the solution. In contrast, a NN-based solution
capability of NNs, a large portion of the data set is utilized for training can achieve the same level of accuracy only with a simple
 X. Liu et al. / Acta Materialia 190 (2020) 105
112 109

 
Fig. 4. NN-based solutions. A NN is composed of an input layer, hidden layers, and an output layer. In this context, the input layer consists of 4 nodes corresponding to x ¼ ab ; wb ; Lb0 ; Lb1 ;
KI
the output layer has a single node for y ¼ PL a0:5 b
1 w
2 . NNs with various hidden layer structures are examined: (a) single-layer perceptrons denotated “ni/n1/no”, and (b) multilayer percep-
1
trons denotated “ni/n1/n2/no”, where ni, n1, n2, no are the number of nodes in the input layer, hidden layers, and output layer, respectively. (c) ML solutions with high accuracy (i.e., small
absolute percentage error) can be obtained even with simple NNs. One can choose appropriate NN-based solutions to tune the accuracy according to the application needs.

interconnected network. As simplicity of a ML solution is essential for user interface can be further developed for accessing and visualizing the
rapid and robust deployment, the NN-based solutions are preferred. problem and the established standard ML solution. In this context, the
The optimal ML solution can be exported in an open-standard file ML solution for the fracture toughness measurements has been created
format such as Javascript object notation (JSON), and shared among and shared through a web-based application, named “SIF Calculator”
researchers and engineers in the community. The standard data-inter- [21]. The ML solution can achieve comparable accuracy to FEM results
change format sets the scene for researchers to operate in various devel- as shown in Figs. 1c and 2b. The overview of the deployment and posi-
opment platforms. The optimal ML solution can also be established as tioning of the ML solutions is illustrated in Fig. 5. The ML solution can
an open-access standard for specific engineering problems. A graphical achieve the functionalities of an empirical solution.

Fig. 5. Functionality and positioning of a ML solution to engineering problems. The optimal ML solution can be exported in an open-standard file format, such as JSON, and
shared among researchers for further operations and development. It can also be established as an engineering standard and stored online for public access. Cloud or web-based
applications can be developed to interactively provide fast and accurate solutions to engineers.
110 X. Liu et al. / Acta Materialia 190 (2020) 105
112

5. Discussion 6. Conclusions

In this work, an innovative class of solutions based on ML models ML provides an alternative way to solve complex engineering
are proposed to solve engineering problems, especially materials problems, due to its advantages in dealing with the nonlinear and
characterization measurements, where they can serve as a promising complex relationship among high-dimensional physical quantities.
substitute when analytical and empirical solutions are not accessible. More specifically in this work, it is demonstrated that ML greatly
The feasibility and advantage of the ML solutions are demonstrated accelerates data interpretation in fracture toughness measurements.
through the application to small-scale fracture toughness measure- The simplicity of employing ML solutions also creates a foundation
ments of pre-notched pentagonal cross-section cantilevers. This ML for standardizing these types of novel experimental methods. Besides
approach can be easily adopted for other experimental characteriza- materials characterization, ML solutions have a wide range of possi-
tion methods with complex specimen geometries, such as cantilever- ble applications in materials science. Possible examples include pre-
based methods [22
25] and indentation-based methods [26
28]. dictions of surface roughness and temperature rise (target variables)
The development process consists of 4 basic steps, in metal cutting under different machining conditions (input varia-
bles), and predictions of the mechanical properties of 3D printing
1. Definition of a well-posed problem. Solving a complex engineering products (target variables) under different manufacturing conditions
problem requires the development of a mathematical description (input variables). In summary, ML solutions can substantially change
and the understanding of what physical quantities are involved. the way that engineering problems are solved.
Dimensional analysis can assist in finding all the relevant dimen-
sionless numbers from these quantities, and the essence of the solu-
tion is to understand the relationships among these quantities. Declaration of Competing Interest
2. Preparation of a data set. The dimensionless quantities involved
can be assorted into input variables and target variables based on The authors declare that they have no known competing financial
their physical meanings. It is necessary to explicitly define the interests or personal relationships that could have appeared to influ-
parameter space of the input variables, which covers the whole ence the work reported in this paper.
region of interest (ROI). Since the reliability of the ML solutions is
of high priority, these solutions are restrained from doing extrap-
olation and the performance within ROI is emphasized. The qual-
Acknowledgements
ity of the data set directly controls the accuracy of the solutions.
Reliable experimental and numerical data are essential.
We acknowledge the financial support from the US Department of
3. Model selection and training. There are a large variety of ML models
Energy Basic Energy Sciences Grant # DE-SC0018113 for this research.
that can be employed. Based on the data set and type of problem,
This work was conducted using the computational resources and serv-
the appropriate ML models should be selected. The optimal solu-
ices at the Brown University Center for Computation and Visualization.
tion is achieved by evaluating their accuracy and simplicity.
4 Deployment of ML solutions. The optimal ML solution can be shared
among the community in an open-standard format. Engineers can
obtain rapid and accurate results by interacting with cloud or web- Appendix A
based applications which can be integrated with the ML solution.

Table A.1
Fracture toughness measurements of polysilicon specimens.

b (mm) a (mm) w (mm) L0 (mm) L1 (mm) Pcr (mN) KIC (MPa ¢ m0.5)

Empirical1 FEM ML

PolySi 1 5.00 0.90 5.00 1.00 10.00 2.30 1.73 0.97 0.97
PolySi 2 4.80 0.90 5.20 1.00 10.00 2.35 1.77 0.99 0.99
PolySi 3 5.20 1.00 5.50 1.00 11.00 2.50 1.78 0.98 0.98
1
The empirical solution was developed in [4,5].

Table A.2
Construction and performance of regression-tree-based solutions.

No. of trees Tree depth Avg No. of leaf nodes Min No. of samples at a leaf node Max. APE1

RT 1 4 16 8580 29.706%
RT 1 5 32 3240 23.828%
RT 1 6 64 1296 19.982%
RT 1 7 128 240 17.440%
RT 1 8 256 88 17.146%
GBRT 256 4 16 430 20.348%
GBRT 256 5 32 106 17.019%
GBRT 256 6 64 51 14.327%
GBRT 256 7 128 50 11.483%
GBRT 256 8 255 50 9.724%
GBRT 512 4 16 93 7.873%
GBRT 512 5 32 50 6.018%
GBRT 512 6 64 50 4.122%
GBRT 512 7 127 50 3.334%
GBRT 512 8 251 50 2.171%
1
Max. APE: maximum absolute percentage error.
 X. Liu et al. / Acta Materialia 190 (2020) 105
112 111

Table A.3
Construction and performance of NN-based solutions.

NN architecture No. of trainable parameters Training Max. APE1 Validation Max. APE Test Max. APE Fine-tuned Max. APE1

4/16/1 97 9.003% 8.438% 8.858% 8.866%

4/32/1 193 6.046% 5.169% 5.357% 5.693%
4/64/1 385 2.907% 2.638% 1.556% 2.764%
4/128/1 769 1.755% 1.762% 1.675% 1.590%
4/256/1 1537 1.691% 1.266% 1.402% 1.476%
4/512/1 3073 1.359% 0.905% 1.101% 1.150%
4/8/8/1 121 7.589% 7.208% 7.177% 7.192%
4/8/16/1 201 5.993% 5.072% 5.299% 5.742%
4/16/8/1 225 4.076% 3.359% 3.561% 3.789%
4/16/16/1 369 4.109% 3.457% 3.989% 3.785%
4/32/32/1 1249 1.618% 1.394% 1.512% 1.558%
4/64/64/1 4545 1.130% 1.043% 1.119% 1.111%
1
Max. APE: maximum absolute percentage error.

Fig. A.1. Training history of NNs. The maximum absolute percentage error (max. APE) on the training and validation data sets are recorded over the number of epochs for some
simple NNs: (a) 4/16/1, (b) 4/32/1, (c) 4/64/1, (d) 4/128/1, (e) 4/256/1, (f) 4/512/1, (g) 4/8/8/1, (h) 4/8/16/1, (i) 4/16/8/1, (j) 4/16/16/1, (k) 4/32/32/1 and (l) 4/64/64/1. Each plot shows
that the training and validation accuracy reach the same plateau, which indicates that the NN is converged and not overfitted.
112 X. Liu et al. / Acta Materialia 190 (2020) 105
112

References [14] L. Breiman, J.H. Friedman, R.A. Olshen, C.J. Stones, Classification and Regression
Trees, CRC, Boca Raton, 1984.
[1] M. Mozaffar, R. Bostanabad, W. Chen, K. Ehmann, J. Cao, M.A. Bessa, Deep learning [15] F. Pedregosa, et al., Scikit-learn: machine learning in Python, J. Mach. Learn. Res.
predicts path-dependent plasticity, Proc. Natl. Acad. Sci. 116 (2019) 26414– 12 (2011) 2825–2830.
26420, doi: 10.1073/pnas.1911815116. [16] J.H. Friedman, Greedy function approximation: a gradient boosting machine, Ann.
[2] L. Liang, M. Liu, C. Martin, W. Sun, A deep learning approach to estimate stress Stat. 29 (2001) 1189–1232.
distribution: a fast and accurate surrogate of finite-element analysis, J. R. Soc. [17] J.H. Friedman, Stochastic gradient boosting, Comput. Stat. Data Anal. 38 (2002)
Interface 15 (2018) 20170844, doi: 10.1098/rsif.2017.0844. 367–378, doi: 10.1016/S0167-9473(01)00065-2.
[3] G.X. Gu, C.T. Chen, M.J. Buehler, De novo composite design based on machine [18] K. Hornik, Approximation capabilities of multilayer feedforward networks, Neural
learning algorithm, Extreme Mech. Lett. 18 (2018) 19–28, doi: 10.1016/j. Netw. 4 (1991) 251–257, doi: 10.1016/0893-6080(91)90009-T.
eml.2017.10.001. [19] T. Dozat, Incorporating Nesterov momentum into Adam, in: Proceedings of the
[4] D. Di Maio, S.G. Roberts, Measuring fracture toughness of coatings using focused- International Conference on Learning Representation, 2018.
ion-beam-machined microbeams, J. Mater. Res. 20 (2005) 299–302, doi: 10.1557/ [20] M. Abadi et al., TensorFlow: large-scale machine learning on heterogeneous dis-
JMR.2005.0048. tributed systems, Preprint at http://arxiv.org/abs/1603.04467 (2016).
[5] D. Di Maio, S.G. Roberts, Measuring fracture toughness of coatings using focused- [21] X. Liu, SIF calculator. https://hint1412.github.io/XLiu.github.io/SIF/, 2019
ion-beam-machined microbeams
ERRATUM, J. Mater. Res. 32 (2017) 1211, doi: (accessed 30 December 2019).
10.1557/jmr.2017.84. [22] D.E. Armstrong, A.J. Wilkinson, S.G. Roberts, Measuring anisotropy in Young’s
[6] C.E. Athanasiou, et al., High toughness carbon-nanotube-reinforced ceramics via modulus of copper using microcantilever testing, J. Mater. Res. 24 (2009) 3268–
ion-beam engineering of interfaces, Carbon 163 (15) (2020) 169–177, doi: 3276, doi: 10.1557/jmr.2009.0396.
10.1016/j.carbon.2020.02.075. [23] M.G. Mueller, et al., Fracture toughness testing of nanocrystalline alumina and
[7] B.N. Jaya, V. Jayaram, Fracture testing at small-length scales: from plasticity in Si fused quartz using chevron-notched microbeams, Acta Mater. 86 (2015) 385–
to brittleness in Pt, JOM 68 (2016) 94–108, doi: 10.1007/s11837-015-1489-2. 395, doi: 10.1016/j.actamat.2014.12.016.

[24] G. Zagar, V. Pejchal, M.G. Mueller, L. Michelet, A. Mortensen, Fracture toughness
[8] S. Wurster, C. Motz, R. Pippan, Characterization of the fracture toughness of
micro-sized tungsten single crystal notched specimens, Philos. Mag. 92 (2012) measurement in fused quartz using triangular chevron-notched micro-cantile-
1803–1825, doi: 10.1080/14786435.2012.658449. vers, Scr. Mater. 112 (2016) 132–135, doi: 10.1016/j.scriptamat.2015.09.032.
€ ken, K. Durst, In situ micro-cantilever tests to study fracture
[9] F. Iqbal, J. Ast, M. Go [25] A.D. Norton, S. Falco, N. Young, J. Severs, R.I. Todd, Microcantilever investigation
properties of NiAl single crystals, Acta Mater. 60 (2012) 1193–1200, doi: 10.1016/ of fracture toughness and subcritical crack growth on the scale of the microstruc-
j.actamat.2011.10.060. ture in Al2O3, J. Eur. Ceram. Soc. 35 (2015) 4521–4533, doi: 10.1016/j.jeurceram-
[10] C.F. Shih, B. Moran, T. Nakamura, Energy release rate along a three-dimensional soc.2015.08.023.
crack front in a thermally stressed body, Int. J. Fract. 30 (1986) 79–102, doi: [26] X. Wang, et al., High damage tolerance of electrochemically lithiated silicon, Nat.
10.1007/BF00034019. Commun. 6 (2015) 8417, doi: 10.1038/ncomms9417.
[11] T.D. Nguyen, S. Govindjee, P.A. Klein, H. Gao, A material force method for inelastic [27] M. Sebastiani, K.E. Johanns, E.G. Herbert, F. Carassiti, G.M. Pharr, A novel pillar
fracture mechanics, J. Mech. Phys. Solids 53 (2005) 91–121, doi: 10.1016/j. indentation splitting test for measuring fracture toughness of thin ceramic coat-
jmps.2004.06.010. ings, Philos. Mag. 95 (2015) 1928–1944, doi: 10.1080/14786435.2014.913110.
[12] T.L. Anderson, Fracture Mechanics: Fundamentals and Applications, CRC Press, [28] M. Sebastiani, K.E. Johanns, E.G. Herbert, G.M. Pharr, Measurement of fracture
Boca Raton, 2005. toughness by nanoindentation methods: recent advances and future challenges,
[13] R.L. Taylor, FEAP - A Finite Element Analysis Program, University of California, Current Opin. Solid State Mater. Sci. 19 (2015) 324–333, doi: 10.1016/j.
Berkeley, 2014. cossms.2015.04.003.