Scripta Materialia: Anup Pandey, Reeju Pokharel

Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Scripta Materialia 193 (2021) 1–5

Contents lists available at ScienceDirect

Scripta Materialia
journal homepage: www.elsevier.com/locate/scriptamat

Viewpoint set

Machine learning based surrogate modeling approach for mapping


crystal deformation in three dimensions
Anup Pandey∗, Reeju Pokharel
Material Science and Technology (MST-8), Los Alamos National Laboratory Los Alamos, NM 87544 United States

a r t i c l e i n f o a b s t r a c t

Article history: We present a machine learning based surrogate modeling method for predicting spatially resolved 3D
Received 8 July 2020 crystal orientation evolution of polycrystalline materials under uniaxial tensile loading. Our approach is
Revised 8 October 2020
orders of magnitude faster than the existing crystal plasticity methods enabling the simulation of large
Accepted 13 October 2020
volumes that would be otherwise computationally prohibitive. This work is a major step beyond existing
ML-based modeling results, which have been limited to either 2D structures or only providing average,
Keywords: rather than local 3D full-field predictions. We demonstrate the speed and accuracy of our surrogate model
Deep neural network approach on experimentally collected data from a face-centered cubic copper sample undergoing tensile
Microstructure evolution deformation.
Reorientation
Crystal plasticity © 2020 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.
Surrogate models

Understanding polycrystalline material behaviors requires the without requiring several hours or even days of computation in
study of local hot spots that develop due to complex hetero- high-performance computing clusters.
geneities and non-linearities at the mesoscale. Studying such com- The computational cost of these CP models can be lowered
plex phenomenon based on models requires computationally ex- through either reduced-order modeling or surrogate modeling. The
pensive crystal plasticity (CP) simulations. However, both finite el- past decade has seen a significant increase in machine learning
ement methods-based [1–3] and fast Fourier transform (FFT)-based (ML) methods, such as deep learning, convolution neural network,
CP methods [4,5] are computationally prohibitive when millions etc., for surrogate modeling with significant computational gain
of material simulations are needed to enable advanced material [13–16], relate microstructure to mechanical behavior and perfor-
screening and design [6,7]. On the other hand, homogenized meth- mance [17–19], and to solve inverse problems such as reconstruct-
ods having lower computational cost [3,8,9] provide only average ing microstructure from diffraction patterns [20,21]. Many ma-
material property predictions, hence local material heterogeneity chine learning approaches for predicting polycrystalline plasticity
information at the grain scale is lost. have been reported in the literature [22–26]. In these works,
Recently, advances in experimental characterization tech- microstructure is mapped to effective material properties, where
niques such as high-energy X-ray diffraction microscopy (HEDM) the input is either a 2D image, a low dimensional representation
[10,11] have enabled in situ observation of microstructure and of the 3D microstructure, or a 3D image of the microstructure
micro-mechanical field evolution of polycrystalline materials in which is then mapped to a single scalar output. However, thus
three-dimensions (3D). However, both the data acquisition and far, in attempting to predict material response to imposed load-
microstructure reconstruction from the measured diffraction pat- ing conditions, the ML frameworks either predict average prop-
terns are extremely slow which can be improved by coupling erties, or only single or a few crystal properties, and only in
with CP frameworks [12]. CP predictions under imposed load- two-dimensions.
ing and boundary conditions resembling that of the experiment We present a proof-of-concept method using deep neural net-
would eliminate the need for recording redundant data; therefore, work (DNN) for solving a regression problem, to predict the mi-
enabling guided experiments on limited available beamtime. crostructure evolution of polycrystalline materials under tensile
However, to provide real-time feedback, CP models will have strain. We demonstrate a ML-based surrogate modeling method
to make instantaneous predictions of microstructure evolution, that directly maps 3D microstructure input to the 3D full-field ma-
terial response. The surrogate model is as accurate as the con-
ventional CP model, as well as general enough to predict mi-

crostructure evolution of arbitrary 3D representative volume ele-
Corresponding author.
ment (RVE).
E-mail addresses: anup@lanl.gov (A. Pandey), reeju@lanl.gov (R. Pokharel).

https://doi.org/10.1016/j.scriptamat.2020.10.028
1359-6462/© 2020 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.
A. Pandey and R. Pokharel Scripta Materialia 193 (2021) 1–5

Fig. 1. (a) Synthetic 3D microstructure generated from Dream.3D. Neighboring 3 × 3 × 3 voxels for the center voxel for a representative microstructure are determined.
Evolution of the center voxel and its neighbors are tracked on the application of strains, which are then used for training the surrogate model. (b) Deep neural network
architecture used for mapping 3D microstructure and 3D full-field material response.

The DNN model is trained on synthetic RVEs generated using The hidden layers used the rectified linear unit activation func-
Dream.3D [27] as input to the FFT-based full-field crystal plasticity tion (ReLU), while the output layer used the linear activation func-
simulations. We generated 100 instances of representative 3D mi- tion. We used logcosh as a loss function between each Euler an-
crostructures discretized on 16 × 16 × 16 grid, with average num- gle component of the model predictions and the ground truth. The
bers of grains of 25 (Fig. 1). The full-field elasto-viscoplastic fast Adam [29] optimizer with a learning rate of 0.001 was used in
Fourier transform (EVPFFT) crystal plasticity simulations [5] are training the models. The details on the DNN can be found in refer-
carried out to generate microstructure evolution data under uni- ences [30,31]. EVPFFT simulated microstructures at strain levels up
axial tensile loading. The details of the EVPFFT model and model to 12% were considered as the ground truth. There were 7,372,800
instantiation with 3D microstructure for numerical simulations are and 1572864 no. of data points in the training and test sets (3D
given elsewhere [5,12]. Single crystal elastic stiffness for Cu and microstructures not included in the training), respectively. The dis-
Voce hardening parameters calibrated using the stress-strain curve orientation angle, defined as the smallest possible rotation angle
reported in [28] used in our simulations were: out of all symmetrically equivalent misorientations that lie within
C11 [GPa] = 168.4, C12 [GPa] = 121.4, C44 [GPa] = 75.4, a fundamental zone (FZ), is used as a final metric for comparing
the model predictions with the ground truth. Disorientation angle
τ0 [MPa] = 45.0, τ1 [MPa] = 30.0, θ0 = 800.0, θ1 = 130.0. is given by:
Each simulation was carried out for up to 12% total strain at 0.02%
gAB = OB gB (OA gA )−1 ,
strain increment.
We formulated a method that learned the local physics by where O denotes one of the cubic symmetry operators, and gA and
considering only the local neighborhood’s effect on each mate- gB are the orientations from the ‘ground truth’ and the model pre-
rial point (voxel), eliminating the need for large input volumes for dictions, respectively. The model is referred to as DNN-CP from
model training. For predicting the evolution of a single voxel, we hereinafter. The DNN-CP predictions of three orientation angles
used only the information from a 33 cube at which the voxel was (φ 1 , , φ 2 ) for the test set and the validation set at the end of
centered (up to the third nearest neighbor), as shown in Fig. 1 (a). the training process are in excellent agreement with the ground
For voxels on the edges of the RVE, zero-padding was done to truth, with only a few outliers, as shown in Fig. 2.
account for the missing neighbors. This approach drastically in- We predicted full 3D microstructures evolution at 0.5% strain
creased the number of data points for model training while sig- increments using the DNN-CP model for 163 and 643 synthetic
nificantly reducing the number of expensive crystal plasticity sim- structures.
ulations of 3D RVEs, and also eliminated the use of large volumes Voxel-wise disorientation angle is calculated between the ML
as inputs during training. prediction and the EVPFFT simulation. Fig. 3 shows the voxel-wise
Next, we prepared a sequence of 25 strain steps (including the disorientation distribution for the full 3D volume as well as spa-
original undeformed state) at an interval of 0.5% strain increment; tial maps of 2D slices extracted from the 3D microstructure, to
where each voxel in the microstructure had 25 steps in its se- demonstrate the comparison of crystallographic orientation predic-
j tion at the local scale. The color code at each point on the 2D
quence (Fig. 1 (a)). For each voxel Vt at a given position j and time
step t, a 3 × 3 × 3 subset of its neighboring voxels Xt was de-
j spatial map corresponds to the disorientation angle at that voxel.
j j j For the 163 3D synthetic structure, the disorientation angle be-
termined. Therefore, in a given sequence, V0 , V0.5 , ...., V12 are the
tween the ground truth and the DNN-CP model predicted orien-
j j j
central voxel and X0 , X0.5 ,....,X12
are the corresponding neighboring tations are calculated, and the distributions are shown in Fig. 3(a).
voxels at different strain levels going from 0% to 12% tensile strains. For the predictions from no strain (0%) to 0.5% strain, 100% dis-
j j j j j j
The data are arranged in (X,Y)={(X0 ,V0.5 ), (X0.5 ,V1 ), (X1.5 ,V2 )... ... orientation angles are within 5◦ . The mean of the distribution is
j j
(X11.5 ,V12 )} format. A crystal orientation represented by three Eu- 0.75◦ and the standard deviation is 0.535◦ . The smaller mean dis-
ler angles (φ 1 , , φ 2 ) was assigned to each voxel. Therefore, each orientation angle implies high fidelity prediction. The model can
X is an 81 component (3 × 3 × 3 × 3) vector and each Y is an predict the evolution of a microstructure by 0.5% strain from any
output vector with 3 Euler angles. given strain state up to 12% strain. This feature is tested for two
We trained a feed-forward deep neural network (DNN) model more strain steps. For the model predictions from 1.5% and 5.5%
using the (X,Y) data for solving a regression problem. Our model (EVPFFT simulated strained structures) to the 2% and 6%, respec-
structure comprises an input layer with 81 variables, three fully tively, 100% and 99.93% of disorientation angles are within 5◦ with
connected hidden layers with 60, 40, and 40 nodes, respectively 0.76◦ and 0.78◦ mean and 0.54◦ and 0.60◦ standard deviation, re-
and an output layer with three nodes, as shown in Fig. 1(b). spectively. The spatially resolved misorientation map between the

2
A. Pandey and R. Pokharel Scripta Materialia 193 (2021) 1–5

6 Training (φ1) 3.0 Training (Φ) 6 Training (φ2)


Validation (φ1) Validation (Φ) Validation (φ2)
prediction [c] 5 y=x 2.5 5
y=x y=x
4 2.0 4
3 1.5 3
2 1.0 2
1 0.5 1
0 0 0

6 Test (φ1) 3.0 Test (Φ) 6 Test (φ2)


y=x y=x y=x
prediction [c]

5 2.5 5
4 2.0 4
3 1.5 3
2 1.0 2
1 0.5 1
0 0 0
0 1 2 3 4 5 6 0 0.5 1 1.5 2 2.5 3 0 1 2 3 4 5 6
ground truth [c] ground truth [c] ground truth [c]

Fig. 2. The DNN-CP predicted Euler angles versus the ground truth (FFT) for training dataset. Model is trained from the 16 × 16 × 16 structures.

(a) 1e2 1e2 1e2


0.5% 2% 6%
3

2 2
2
Count

1 1
1

0 0 0
(b) 1e4 1e4 1e4
2
0.5% 2% 2 6%
1.75
Count

1 1
1

0 0 0
0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5
Misorientation [deg] Misorientation [deg] Misorientation [deg]
5
(c) Layer 4 Layer 8 Layer 12
(d) Layer 16 Layer 32 Layer 48
Misorientation [deg]

15 4 60
3

1
0 0
0 15 0 15 0 15 0 60 0 60 0 60
0

Fig. 3. Voxel-wise disorientation distribution between the DNN-CP predictions and the EVPFFT simulations from no strain (0%) to 0.5% strain (left), 1.5% strain to 2% strain
(middle) and 5.5% strain to 6% strain (right) for(a) 16 × 16 × 16 and (b) 64 × 64 × 64 synthetic 3D structures. The spatial map for the point-wise disorientation calculated
between the ground truth and DNN-CP model prediction from 0% to 0.5% strain for the different layers: (c) 4, 8 and 12 of the 16 × 16 × 16 and (d) (layer 16, 32 and 48)
of the 64 × 64 × 64 3D microstructures.

3
A. Pandey and R. Pokharel Scripta Materialia 193 (2021) 1–5

(a) 1e4 (b)


0.5% 3% 6%
8

6
Count

0 0.25 mm
0 1 2 3 4 5 0 1 2 3 4 5 0 1 2 3 4 5
Misorientation [deg] Misorientation [deg] Misorientation [deg] Misorientation [deg]
(c) (d)
Layer 25 Layer 50 Layer 75 0 1 2 3 4 5

Grain 1 Grain 2
Grain 4

Grain 3

Fig. 4. (a) Disorientation distribution between the DNN-CP predictions and the EVPFFT simulations of experimental Cu structure from no strain (0%) to 0.5% strain (left),
2.5% strain to 3% strain (middle) and 5.5% strain to 6% strain (right). (b) Distribution of disorientation between 0.5% EVPFFT simulated experimental Cu microstructure and
the DNN-CP model prediction of crystal orientations. Color scheme shows the disorientaion angle in degrees. The grid size of measured microstructure is 420 × 420 × 100.
The sample is cylindrical and the buffer zone has zero misorientation. (c) Three different layers (layes 25, 50 and 75) shown in (b). (d) Disorientation distribution in random
grains extracted from the 3D volume shown in (b).

ground truth and the DNN-CP model predicted from 0% to 0.5% zero is assigned to the buffer region. The misorientation maps con-
strain for the different layers (layers 4, 8, and 12) of the 163 3D firms the low disorientation angles distribution for both the mod-
structure are shown in Fig. 3(c). All three layers corroborate the els. The accuracy of the predictions at the grain scale are also eval-
overall low disorientation angle distribution. uated for three randomly selected grains from the 3D microstruc-
Similarly, we predicted the full-field orientations evolution of ture (Fig. 4(d)). All 3D grains show low disorientation angle distri-
the 643 3D synthetic microstructure, from no strain (0%) to 0.5% bution.
strain, 1.5% to 2% strain and 5.5% to 6% strain using the DNN- Thus, the surrogate model accurately predicts microstructure
CP model. The disorientation angle between the orientations from evolution of arbitrary 3D RVEs (643 synthetic structure and exper-
the ground truth (EVPFFT simulated strained structures) and model imental Cu structure), with previously unseen microstructure pa-
predictions are calculated, and the distributions are shown in rameters such as texture, grain size distribution, and grain mor-
Fig. 3(b). For all three strain steps, 99.98% of the disorientations are phology, which elucidates the generality of the model.
within 5◦ . For 0.5% prediction, the mean of the distribution is 0.80◦ The maximum average rotation angle for a grain in fcc Cu as
with the standard deviation is 0.54◦ . For 2% and 6% predictions, the measured by Pokharel et al. [28] is 2.7 ◦ . Our calculated average
mean is 0.81% and the standard deviation is 0.55◦ . The correspond- disorientation angle between the unstrained (0%) Cu and the 6%
ing 2D maps of three different cross sections (layers 16, 32 and 48) predicted Cu is 2.25◦ , with 97.81% of disorientation angle within
of a synthetic microstructure are shown in Fig. 3(d). The 2D maps 5◦ . This shows that the microstructures deformation predicted by
corroborate narrow disorientation angle distributions. the model are well within the experimental error. For the 643 syn-
For the practical use of DNN-CP in the real-time experiments, thetic structure, the average disorientation angle between the un-
the model should predict crystal reorientation in the experimen- strained structure and 6% predicted structure is 2.34◦ , which is
tally measured structures. We predicted the 0.5% strain evolution comparable to 6% EVPFFT simulated structure with an average of
of the experimentally measured 3D microstructure obtained from 2.35◦ . This shows that the DNN-CP model is comparable to the
high-energy Xray diffraction microscopy (HEDM) measurements of EVPFFT model and the predicted microstructures are well within
Cu reported by [28] using the DNN-CP model. EVPFFT simulated the experimental error.
structures of the experimentally measured microstructures from We have also trained the models with larger grid sizes
HEDM are used as the ‘ground truth’ for comparison. The orienta- (64 × 64 × 64 and 128 × 128 × 128), but the accuracy of the
tion evolution from no strain (0%) to 0.5% strain, 2.5% (EVPFFT sim- model does not change. This supports our approach of considering
ulated structure) to 3% strain and 5.5% (EVPFFT simulated struc- only the local interactions to capture the required physics. How-
ture) to 6% strain are predicted using the DNN-CP model. The dis- ever, if needed, one can also adopt an adaptive local neighborhood
orientation angle distribution between predicted orientations and approach as demonstrated on 2D images by Donegan et al. [25], by
the ground truth are shown in Fig. 4(a). For the three predictions, employing a 3D convolutional neural network directly on 3D data
99.85%, 99.87%, and 99.78% of the disorientation angles are within and changing the RVE size by adjusting the filter size used in the
5◦ , respectively. The mean of the distributions are 0.82◦ , 0.82◦ and network.
0.81◦ , and the standard deviations are 0.69◦ , 0.65◦ and 0.68◦ , re- To compare the speed of DNN-CP relative to EVPFFT, we ran
spectively. both on a single 2.5 GHz Intel Xenon W processor. The DNN-CP
The disorientation map for the full 3D microstructure is shown model showed a speed up of > 300 × (3.1 min versus 1000
in Fig. 4(b), and the 2D misorientation slice for three different mins.) in predicting polycrystalline microstructure evolution under
layers (layers 25, 50, and 75) are shown in Fig. 4(c). The exper- plastic deformation in comparison with EVPFFT for 0.5% strain evo-
imental Cu sample is cylindrical, and the misorientation angle of lution. For accurate predictions and better convergence, a maxi-

4
A. Pandey and R. Pokharel Scripta Materialia 193 (2021) 1–5

mum strain step size of EVPFFT cannot exceed ~ 0.02%. Because This work was supported by the U.S. Department of Energy
our trained model can instantly evolve a microstructure by 0.5% through the Los Alamos National Laboratory. Los Alamos National
strain, it eliminates the need for large computing resources which Laboratory is operated by Triad National Security, LLC, for the Na-
are required for the iterative methods used by conventional CP cal- tional Nuclear Security Administration of U.S. Department of En-
culations, which must take large numbers of very small strain steps ergy (Contract No. 89233218CNA0 0 0 0 01). This work was supported
(~ 0.02%) in order to ensure convergence. We have found that the by Los Alamos National Laboratory Directed Research and Devel-
DNN-CP model can take single steps as large as 1% before the pre- opment (LDRD) project #20190571ECR. We are grateful to Dr. A.
diction accuracy starts to degrade. This feature makes the model Scheinker and Dr. A.D. Rollett for fruitful discussions. The data and
applicable for in operando using only a single local desktop ma- code will be made available upon reasonable request.
chine.
Although the DNN-CP approach is orders of magnitude faster Declaration of Competing Interest
than the traditional iterative CP method, there are errors (average
disorientation angle ~ 0.8) in the evolved microstructures relative The authors declare that they have no known competing finan-
to the CP method. However, a hybrid approach combining DNN- cial interests or personal relationships that could have appeared to
CP with traditional iterative CP methods has the potential to pro- influence the work reported in this paper.
vide extremely large speed ups, without compromising accuracy
relative to a traditional CP method alone. For example, in order References
to quickly calculate the evolution of a sample state S0 through
[1] D. Peirce, R.J. Asaro, A. Needleman, Acta Metall. 31 (12) (1983) 1951–1976.
a 1% strain step to state S1 , we can use the DNN-CP model to [2] D. Raabe, F. Roters, Int. J. Plast. 20 (3) (2004) 339–361.
almost instantly provide an accurate estimate Sˆ1 of S1 and then [3] F. Roters, P. Eisenlohr, L. Hantcherli, D.D. Tjahjanto, T.R. Bieler, D. Raabe, Acta
use that as a starting point for a traditional iterative CP calcula- Mater. 58 (4) (2010) 1152–1211.
[4] R.A. Lebensohn, Acta Mater 49 (14) (2001) 2723–2737.
tion (or reconstruction techniques), which will fine tune Sˆ1 to cre- [5] R.A. Lebensohn, A.K. Kanjarla, P. Eisenlohr, Int. J. Plast. 32 (2012) 59–69.
ate S1 in a much smaller number of steps than would have been [6] A. Franceschetti, A. Zunger, Nature 402 (6757) (1999) 60–63.
required if it started from S0 . Note that the ML model is inde- [7] K.T. Butler, D.W. Davies, H. Cartwright, O. Isayev, A. Walsh, Nature 559 (7715)
(2018) 547–555.
pendent of how the training or test data set is generated and if
[8] R.A. Lebensohn, C. Tomé, Acta Metall. Mater. 41 (9) (1993) 2611–2624.
the training data can be generated from a predictive model with [9] C. Miehe, J. Schröder, J. Schotte, Comput. Methods Appl. Mech. Eng. 171 (3–4)
more complex hardening laws or boundary conditions, that will (1999) 387–418.
[10] H.F. Poulsen, Three-dimensional X-ray diffraction microscopy: Mapping poly-
be well captured by the ML framework presented here. Further-
crystals and their dynamics, 205, Springer Science & Business Media, 2004.
more, the DNN-CP model provides the flexibility to incorporate ex- [11] R. Pokharel, Materials Discovery and Design, Springer, 2018, pp. 167–201.
perimental data to improve its predictive capability by re-training [12] R. Pokharel, R.A. Lebensohn, Scr. Mater. 132 (2017) 73–77.
and transfer learning, as recently demonstrated in reference [13] R. Liu, A. Kumar, Z. Chen, A. Agrawal, V. Sundararaghavan, A. Choudhary, Sci.
Rep. 5 (1) (2015) 1–12.
[21]. [14] Y. Zhang, A.H. Ngan, Int. J. Plast. 115 (2019) 18–28.
This proof-of-concept study was limited to orientation evolu- [15] P.V. Balachandran, D. Xue, J. Theiler, J. Hogden, T. Lookman, Sci. Rep. 6 (2016)
tion at a given strain step under uniaxial tension, and local interac- 19660.
[16] G. Capuano, J.J. Rimoli, Comput. Methods Appl. Mech. Eng. 345 (2019)
tions were shown to be sufficient to describe such a phenomenon 363–381.
accurately. In future studies, we will extend the model to predict [17] D. Reimann, K. Chandra, N. Vajragupta, T. Glasmachers, P. Junker, A. Hartmaier,
local mechanical properties such as stress and strain fields together et al., Front. Mater. 6 (2019) 181.
[18] A. Mangal, E.A. Holm, Int. J. Plast. 114 (2019) 1–14.
with the full microstructure evolution under general loading con- [19] U. Ali, W. Muhammad, A. Brahme, O. Skiba, K. Inal, Int. J. Plast. 120 (2019)
ditions. 205–219.
We employed deep neural network and successfully demon- [20] M.J. Cherukara, Y.S. Nashed, R.J. Harder, Sci Rep 8 (1) (2018) 1–8.
[21] Y.-F. Shen, R. Pokharel, T.J. Nizolek, A. Kumar, T. Lookman, Acta Mater. 170
strated that the ML model trained using a representative synthetic
(2019) 118–131.
data can predict the spatially and temporally resolved crystal re- [22] A. Cecen, H. Dai, Y.C. Yabansu, S.R. Kalidindi, L. Song, Acta Mater. 146 (2018)
orientation under uniaxial tensile loading on measured 3D mi- 76–84.
[23] S.R. Kalidindi, J. Appl. Phys. 128 (4) (2020) 041103.
crostructure data from HEDM to the level of accuracy of EVPFFT.
[24] C. Herriott, A.D. Spear, Comput. Mater. Sci. 175 (2020) 109599.
The model is demonstrated to be general and can be used for ar- [25] S.P. Donegan, E.J. Schwalbach, M.A. Groeber, Mater Charact 161 (2020) 110123.
bitrary 3D polycrystalline microstructure. This was achieved by im- [26] A. Rovinelli, M.D. Sangid, H. Proudhon, W. Ludwig, npj Comput. Mater. 4 (1)
plementing a novel method, which accounts for each point’s inter- (2018) 1–10.
[27] M.A. Groeber, M.A. Jackson, Integrating materials and manufacturing innova-
actions with only its local neighborhood, thus learning the local tion 3 (1) (2014) 5.
physics from the training data. The computational time is reduced [28] R. Pokharel, J. Lind, S.F. Li, P. Kenesei, R.A. Lebensohn, R.M. Suter, A.D. Rollett,
by several orders of magnitude, as the model can predict deforma- Int. J. Plast. 67 (2015) 217–234.
[29] D.P. Kingma, J. Ba, arXiv preprint arXiv:1412.6980 (2014).
tion for large strain steps and avoids numerical iterations per strain [30] Y. LeCun, Y. Bengio, G. Hinton, Nature 521 (7553) (2015) 436–444.
step otherwise needed by the conventional CP methods. [31] I. Goodfellow, Y. Bengio, A. Courville, Deep learning, MIT press, 2016.

You might also like