www.nature.

com/npjcompumats

ARTICLE OPEN

Learning two-phase microstructure evolution using neural

operators and autoencoder architectures
1 3✉ 1,2 ✉
Vivek Oommen , Khemraj Shukla2, Somdatta Goswami 2
, Rémi Dingreville and George Em Karniadakis

Phase-field modeling is an effective but computationally expensive method for capturing the mesoscale morphological and
microstructure evolution in materials. Hence, fast and generalizable surrogate models are needed to alleviate the cost of
computationally taxing processes such as in optimization and design of materials. The intrinsic discontinuous nature of the physical
phenomena incurred by the presence of sharp phase boundaries makes the training of the surrogate model cumbersome. We
develop a framework that integrates a convolutional autoencoder architecture with a deep neural operator (DeepONet) to learn the
dynamic evolution of a two-phase mixture and accelerate time-to-solution in predicting the microstructure evolution. We utilize the
convolutional autoencoder to provide a compact representation of the microstructure data in a low-dimensional latent space. After
DeepONet is trained in the latent space, it can be used to replace the high-fidelity phase-field numerical solver in interpolation tasks
or to accelerate the numerical solver in extrapolation tasks.
npj Computational Materials (2022)8:190 ; https://doi.org/10.1038/s41524-022-00876-7
1234567890():,;

INTRODUCTION techniques solve such coupled models but to the detriment of a

The phase-field method has emerged as a powerful, heuristic tool higher computational cost.
for modeling and predicting mesoscale microstructural evolution Modern ML models have paved the way for the development of
in a wide variety of material processes1–5. This method models fast emulators for solving parametric partial differential equations
interfacial dynamics without the overhead of resorting to (PDEs)23–36. There are strategies for accelerating the simulation of
advanced interfacial tracking algorithms such as level-set6 or PDEs. A promising approach for accelerating the predictions of
adaptive meshing7. Scalar, auxiliary, continuous field variables phase field-based microstructure evolution problems consists of
(so-called phase-field variables) are used to represent the using recurrent neural networks (RNNs) to learn the time-
evolutionary state of the microstructure dynamics such as in dependent, microstructure evolution in latent space37,38. Within
crack growth and propagation8,9, thin-film deposition10,11, and this framework, statistical functions combined with linear and
dislocation dynamics12 to name a few. The Cahn-Hilliard, nonlinear embedding techniques are used to represent the
nonlinear diffusion equation1,13,14, is one of the most commonly microstructure evolution in latent space. Such RNN-based
used governing equations in phase-field models. It describes the surrogate models demonstrated success in generating rapid
process of phase separation, by which a two-phase mixture predictions of the time evolution of the microstructural auto-
spontaneously separates and form domains pure in each correlation function. The microstructure reconstructed from these
component. The Cahn-Hilliard equation finds applications in statistical functions, using for instance a phase recovery algo-
diverse fields ranging from complex fluids to soft matter and rithm39, was then used as an input for a high-fidelity solver that
serves as the starting point of many phase-filed models for marches ahead in time. The developed approach reported a 5%
microstructure evolution. loss in accuracy against the high-fidelity phase-field solvers.
Traditional numerical approaches to solve the fourth-order However, this class of models also comes with challenges. First,
parabolic Cahn-Hilliard equation include finite differences15, the training and inference using RNNs as a surrogate model can
spectral approximation16, finite element analysis with mixed be relatively slow due to the temporal dependence of the current
methods17, and isogeometric analysis18,19. The coupled stiff predicted field on fields predicted at previous time steps,
equation simultaneously captures a quick phase separation and prohibiting the efficiency of the algorithm for large datasets.
a very slow coalescence. Evidently, the two sub-processes operate Second, the RNN-based architecture learns the underlying
on significantly different spatial and temporal scales, making it evolutionary dynamics in terms of statistical functions (non-
challenging to solve efficiently and accurately within realistic time primitive variables) of the microstructure. Reconstructing a
constraints and reasonable computational capabilities20. Improve- microstructure from these statistical functions is a non-trivial
ments in computational complexity have been enabled by the and ill-posed problem40. This reconstruction step can incur
growing interest in data-driven models using machine learning additional errors especially for interfacial dynamics problems
(ML) methods. However, striking a balance between computa- where resolving intricate spatial length scales such as in dendrite
tional efficiency and accuracy has often been a challenge while growth phase-field problems is key.
employing these methods. Indeed, for complex and multi-variate In this work, we propose an alternative approach to circumvent
phase-field models, the efficient Green’s function21 does not the aforementioned challenges. We formulate the microstructure
ensure an accurate solution, while Bayesian optimization22,23 evolution problem as being equivalent to learning a mapping

1
School of Engineering, Brown University, Providence, RI, USA. 2Division of Applied Mathematics, Brown University, Providence, RI, USA. 3Center for Integrated Nanotechnologies,
Sandia National Laboratories, Albuquerque, NM 87185, USA. ✉email: [email protected]; [email protected]

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 V. Oommen et al.
2
1234567890():,;

Fig. 1 Schematic representation of DeepONet with convolutional autoencoder. Step 1 involves training of the convolutional autoencoder
to minimize Lae . The encoder learns a suitable transformation from the high-dimensional microstructure to a low-dimensional latent space
through a series of convolution (blue layers) and MaxPooling (green layers) operations. The decoder remaps the latent representation of the
microstructure back to the original, real space by performing transpose convolution (orange layers) operations. A detailed description of the
architecture is provided in Table 1. In step 2, we train the DeepONet in the latent space to minimize Ld . The entire history of 80 steps is
encoded by the pre-trained convolutional encoder as Φ. ~ DeepONet learns to predict ϕðtÞ
~ at any desired time t, fed to the trunk network. The
latent representation of the microstructure predicted by DeepONet is then re-mapped back to the primitive space by the transpose
convolutional decoder.

function G : u ! ϕ such that, (ordinary differential equations (ODEs) and PDEs) of such problems
by learning the underlying operator, a mapping from functions to
Gðuðx; y; tÞÞ ¼ ϕðx; y; tÞ; (1)
functions, from the available datasets for a broad range of
where u is the history of the microstructure evolution and ϕ(x, y, t) problems28,42. We show that such an architecture is more robust
is the state of the microstructure at time t. We develop a than the RNN-based architecture in terms of training, computational
framework that integrates a convolutional autoencoder architec- efficiency, and sensitivity to noise. The decoder part of the
ture with a Deep Operator Network41 (DeepONet) to learn this convolutional autoencoder can efficiently reconstruct the time-
mapping. Figure 1 illustrates the complete end-to-end workflow of evolved microstructure from the DeepONet predictions bypassing
the proposed algorithm. We utilize a convolutional autoencoder to the challenges associated with reconstruction-induced errors when
provide a compact representation of the microstructure data in a using statistical functions to represent the microstructure for
low-dimensional, latent space. This convolutional autoencoder instance. Overall, the trained autoencoder–DeepONet framework
approach is then combined with the DeepONet architecture to can then be used to replace the high-fidelity phase-field numerical
learn the dynamics of two-phase microstructures in the auto- solver in interpolation tasks for parameters inside the distribution of
encoder latent space. The DeepONet architecture has demon- inputs used during training or to accelerate the numerical solver in
strated its ability to model the governing differential equations extrapolation tasks for parameters outside this distribution.

npj Computational Materials (2022) 190 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 V. Oommen et al.
3
RESULTS information about the history through the branch network. In
Training and optimization of neural operators and addition, due to the lack of temporal dependencies, the fully
autoencoder architectures connected layers in the trunk network and convolutional layers in
We first investigated the impact of the size of the latent dimension the branch network of the DeepONet architecture can be easily
of the autoencoder, ld, on the model performance. To this end, we parallelized, unlike LSTMs. This makes training and inference of
trained five autoencoder models with ld = 9, 25, 64, 100, and 196 DeepONet significantly faster than the RNN architectures.
respectively. Details of the hyper-parameters used in these five We carried out additional simulations to analyze the sensitivity of
convolutional autoencoders are provided in Table 1. For any given the proposed approach to the number of samples used for training.
time step during the evolution of the microstructure, the encoder We considered training datasets with 25%, 50%, 75% and all the
reduced a 128 × 128 microstructure ϕ(x, y, t) to a latent vector of 1600 training samples. We adopted the same methodology
size ld. The decoder mapped the microstructural latent space proposed in Methods and trained separate autoencoder–DeepONet
representation back to a 128 × 128 microstructure ϕðx; ^ y; tÞ (see models on each of these datasets. Details can be found in
Methods for more details). Each autoencoder training took Supplementary Note 2. The model performance was evaluated on
approximately 33 h on one NVIDIA GeForce RTX 3090 GPU. Next, the basis of forecasting errors on test data, shown in Supplementary
we trained the DeepONet model for 120,000 epochs on the latent Figure 2. As expected, we observe better accuracy in the model
space learned by the convolutional encoder for each of the five predictions when increasing the number of training samples. The
trained autoencoder models. The last layer of the branch and model trained with 1200 data samples shows similar accuracy to the
trunk networks for all the models uses a linear activation function. best model trained with 1,600 data samples, indicating convergence
The output of the DeepONet model was then sent to the trained of the training procedure.
convolutional decoder, which performed a mapping from the Finally, we also evaluated the effect of using different loss
latent space back to the original microstructure space, ϕðx;^ y; tÞ. functions on training the autoencoder models. Specifically, we
We evaluated the effect of the size of the latent dimension of trained various autoencoders by minimizing L1 loss, relative L1
each of the models on the basis of the relative L2 norm computed loss, L2 loss, relative L2 loss, and mixed loss (L2 loss for the initial
across the training and testing dataset for all the time steps, 5000 epochs and L1 loss for the remaining epochs). The choice of a
including the forecasting time frames, t = {t90, … , t99} not seen by loss function determines the landscape in a hyperspace for the
the surrogate model (Note that one time frame is equal to 500,000 optimizer to traverse in pursuit of global minima/best local
time steps, t = 500,000Δt, see Methods for additional details). All minima and avoiding the saddle points. For this task, we used a
the details of this survey analysis, including the DeepONet DeepONet model with architecture 1 (Table 2) on each of the
architecture, the L2 norm of relative error on train and test learned latent microstructure data and re-transformed the
datasets, and the computational time taken for training the DeepONet predictions using a pre-trained decoder to retrieve
DeepONet model are reported in Table 2. From this survey, we the microstructure. We analyzed the model performance by
observe that the model predictions improve when we increase the computing the forecasting error, Dtest ðtÞ, on unseen test data, as
size of the latent dimension. In general, DeepONet models with shown in Supplementary Note 3. We observed that models for
tanh and sin activation functions performed better compared to which the autoencoder is trained on L2 loss performed better than
models with a ReLU activation for this particular class of problems. the one which used L1 loss. When there are no outliers as
As such, our best model consists of a convolutional autoencoder solutions, L2 loss is expected to perform better than L1. In the
with ld = 196 and a DeepONet model with architecture 1 and sin presence of outliers, L2 squares them as compared to linear
activation function (shown in Table 2). Although the training contribution in the L1 norm. Similarly, mean values of relative L1
dataset consists of 1,600 different microstructure-evolution and L2 are a better choice for autoencoder loss, Lae , than the
trajectories, each represented by over 80 snapshots from t = {t10, mean of L1 and L2, respectively. The relative loss values are always
t11, … , t89}, the DeepONet training is faster compared to popular of Oð1Þ and help in achieving convergence faster as the learning
RNN architectures such as the Long Short-Term Memory (LSTM) or rate is of Oð103 Þ. We have observed such an improvement in
Gated Recurrent Unit (GRU) networks38,42,43. Since DeepONet does convergence in other problems, e.g. electro-convection, where we
not have recurrent connections, there are no temporal depen- had sharp interfaces and multiscale dynamics44. Overall, all the
dencies during the training or at the inference stage. Instead, the models performed consistently well. As such, our autoencoder
network relies on the convolution operations that encode architecture of choice is an autoencoder with ld = 196 using a

Table 1. Details of the hyper-parameters used in the convolutional autoencoder.

Layer Kernel Size Width Activation Output

1 Conv2D 3×3 16 ReLU 128 × 128 × 16

2 Max-Pool 2×2 64 × 64 × 16
3 Conv2D 3×3 8 ReLU 64 × 64 × 8
4 Max-Pool 2×2 32 × 32 × 8
5 Conv2D 3×3 4 ReLU 32 × 32 × 4
6 Max-Pool 2×2 Reshaped to 1024
7 Fully connected ld Linear ld
8 Fully connected 1024 ReLU Reshaped to 16 × 16 × 4
9 Transpose Conv2D 2×2 ReLU 32 × 32 × 4
10 Transpose Conv2D 2×2 ReLU 64 × 64 × 8
11 Transpose Conv2D 2×2 ReLU 128 × 128 × 16
12 Transpose Conv2D 3×3 ReLU 128 × 128 × 1
ld is the dimension of latent space.

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2022) 190
 V. Oommen et al.
4
Table 2. Detailed survey of different latent dimension size, ld, network architecture, and non-linear activation functions.

ld DeepONet Architecture Activation Dtrain Dtest DeepONet training time

per 1000 epochs (s)

ReLU 0.03621 0.06803 56

196 Architecture 1 tanh 0.02177 0.06233 56
sin 0.01408 0.01620 57
ReLU 0.04076 0.05991 31
100 Architecture 2 tanh 0.03196 0.04699 30
sin 0.02684 0.03679 31
ReLU 0.06708 0.07791 31
64 Architecture 3 tanh 0.04773 0.06013 35
sin 0.04781 0.05739 32
ReLU 0.16527 0.20097 19
25 Architecture 4 tanh 0.16507 0.20134 19
sin 0.16551 0.20167 20
ReLU 0.31536 0.3186 10
9 Architecture 5 tanh 0.31523 0.31903 11
sin 0.31539 0.31876 11

Architecture Branch Network Trunk Network

1 3 × [conv(32,(3,3))] + [1960] 2 × [100] + [1960]

2 2 × [conv(32,(3,3))] + [1100] 2 × [100] + [1100]
3 2 × [conv(32,(3,3))] + [512] 2 × [100] + [512]
4 1 × [conv(64,(3,3))] + [500] 2 × [100] + [500]
5 1 × [conv(128,(3,3))] + [180] 2 × [100] + [180]

relative L2 loss function. Taken together, these results demonstrate high for the initial time steps, where features span multiple length
not only the ability of our framework to accurately provide a scales and evolve rapidly with time. However, the predictions
compact representation of the microstructure data in a low- improve over time when the evolution process slows down and
dimensional latent space, but they also illustrate the robustness of the microstructure features coarsen. The time steps shown in
the training of this framework. Fig. 3a were used during the training of the model.
Next, we evaluated the capability of the model to forecast time
Performance accuracy and forecasting ability frames t = {t90, t91, … , t99}. From Fig. 3b, the error is seen to
increase gradually when the model extrapolates at unseen time
A comparison of the predictions from our accelerated framework
instances. A closer look at the forecasting predictions offers
with that from high-fidelity phase-field simulations for a represen-
further insights into the DeepONet predictive performance. We
tative case of microstructure evolution at three different time steps
computed the mean of DðtÞ across the training and testing
is shown in Fig. 2. During the initial time steps, the microstructure is
datasets for all the models given in Table 2. We also plotted these
rich with multiple features and evolves rapidly with respect to time.
values in Supplementary Fig. 4 with additional details in
Our autoencoder–DeepONet as a surrogate model is able to
Supplementary Note 4. We observe that the mean relative L2
successfully predict the larger features and the overall morphology
error reduces when increasing the latent dimension of the
of the microstructure. The point-wise error snapshots suggest that
autoencoder model. In other words, the model with a larger
the model fails to identify the relatively smaller features in the
latent space is able to predict the evolution of the microstructure
microstructure and contains significant errors along the sharp
in forecasting mode. This is intuitive because a larger dimension of
boundaries. In other words, the spatial gradient of the phase
the latent space implies that there are more basis functions to
concentration is not as sharp as that of the true microstructure
express the encoded information about the microstructure and its
obtained from high-fidelity, phase-field simulations.
evolution, and therefore the network has an improved represen-
From Fig. 2, we qualitatively get the intuition that the predicted
tation capability. However, this trend seems to saturate beyond
microstructures contain errors at the earlier time steps because of
ld = 100. For the model with ld = 100 or ld = 196, the forecasting
the missing, small-size features and at the later time steps due to
error is always less than 6%. The logarithm of the relative L2 error
smoother boundaries predicted by the model. To confirm and
linearly increases for ld = 64,100 and 196 for the forecasting time
quantify this notion, we computed the L2 norm of the relative
error at each time step, DðtÞ, defined as: step, whereas for ld = 9 and ld = 25, the error is high and remains
constant.
PP 2
^
x y ϕðx; y; tÞ  ϕðx; y; t; θÞ
DðtÞ ¼ PP 2 ; t 2 ½t10 ; t 11 ; ¼ ; t99 : Robustness of the surrogate DeepONet framework: Sensitivity
x y ϕð x; y; t Þ
to noise
(2) We evaluated the accuracy and robustness of the predictions from
To analyze the accuracy of the prediction at each time step, we our surrogate model by systematically increasing the noise levels
calculated DðtÞ across the samples in the training and testing in the model input. For this analysis, we considered the best
datasets, and created a boxplot as shown in Fig. 3. The error is model with ld = 196 and DeepONet architecture 1 (see Table 2)

npj Computational Materials (2022) 190 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 V. Oommen et al.
5

Fig. 2 Predictions of microstructure evolutions. The true (top row), predicted (middle row), and point-wise error (bottom row) for a
microstructure realization evolving in time. The snapshots at time frames t = t10, t30, t99 are shown here. The network used for this simulation
has ld = 196, and uses the following network architecture: Branch network -- 2 × [conv(128, (3, 3))] + [3920]; Trunk network -- 2 × [100] + [3920].

with sin activation functions. We added Gaussian white noise to are sufficient to capture the salient features of the microstructure
our microstructure data with zero mean and standard deviations, in latent space, such representation does not uniquely map back
σ = 0.5%, 1%, 2%, 3%, 4%, 5%, 10%. To evaluate the model to the true microstructure as it is an ill-posed, inverse problem.
performance, we used the relative L2 norm, D, as defined in Eq. Here, our results essentially bypass such a challenge by taking
(8). The forecasting error, D, is calculated across the samples advantage of the fact that autoencoders are robust to corruption
present in the test dataset. Details can be found in Supplementary in the representations they learn. The denoising nature of a
Note 5. trained autoencoder enables the encoder to learn a stable and
From Supplementary Table 1 and Supplementary Figure 5, the consistent mapping to the latent space. This makes training of the
relative L2 norm does not increase noticeably when noise is added DeepONet much more stable and results in accurate predictions
to the model input. In fact, the surrogate is almost invariant to of the microstructure at any desired time step.
noise up to 10% Gaussian white noise, as presented in
Supplementary Figure 5. Previous studies45–47 illustrated the
Effect of time resolution
capability of autoencoders to denoise noisy images. The
transformation to a low-dimensional latent space forces the The high-fidelity phase-field forward numerical solver (MEMPHIS)
autoencoder to retain the dominant features alone while discretizes the time with a time step Δt = 1 × 10−4 (see Methods
discarding unnecessary noise. The convolutional autoencoder for additional details). The stability of this numerical integration
used in our approach does exactly that by denoising the noisy scheme can be achieved by strictly following the Courant-
microstructure input. The encoder filters out noise and only Friedrichs-Lewy (CFL) condition to solve the Cahn-Hilliard
retains the dominant energy modes of the microstructure data. equation for 50M time steps. The solver saves snapshots of the
The output of the convolutional encoder is almost in its pure form, solution at every 500,000th time step resulting in 100 micro-
free from noise and therefore it enables the DeepONet to make structure time frames for each realization. We initially utilized 80
stable predictions. The decoder accurately reconstructs the equally spaced time frames between the 10th and 90th time
microstructure from the predictions made by DeepONet in the frames for training the surrogate model. Therefore, for the
latent space. This performance illustrates the robustness and surrogate DeepONet model, each time step was 500000 × Δt = 50.
efficacy of the present framework as compared to other machine- To investigate the effect of different spacing of physical time on
learned frameworks that use statistical functions to encode the the surrogate DeepONet model, we performed a sensitivity study
microstructure representation40. Indeed, it has been illustrated by using data that are spaced differently in time to train the model.
Herman and coworkers40 and others48 that, while statistical Specifically, we trained DeepONet models on datasets with a time
functions such as the microstructure auto-correlation functions spacing of 500kΔt, 2 × 500kΔt, 5 × 500kΔt, and 10 × 500kΔt. The

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2022) 190
 V. Oommen et al.
6

Fig. 4 Variation of forecasting error on test data for models

trained on data with different spacing in time. Spacing in time
tested are: 500kΔt, 2 × 500kΔt, 5 × 500kΔt, 10 × 500kΔt. Dtest ðtÞ
represents the relative L2 error computed across the samples in test
dataset at different time steps.

realization in the dataset. By placing more emphasis on early

snapshots during training, we endow DeepONet with an inductive
bias to learn the fast dynamic accurately. Practically speaking, we
are forcing the DeepONet model, GðΦÞðt; ~ θd Þ, to predict earlier
time steps repeatedly by creating a new training dataset with
~
repeated ϕðtÞ for each realization. Since the DeepONet model is
trained to minimize the mean squared error between the true and
predicted microstructures, the model is driven to give greater
emphasis to microstructures developing at earlier time steps. The
results from this training procedure are depicted in Fig. 5. Indeed,
Fig. 3 L2 norm of the relative error between true and predicted from the comparison of the predicted microstructure without and
microstructures at each time step. a Box plot with respect to DðtÞ with an emphasis on earlier time steps in Fig. 5b and c,
computed over the training time steps over the train and test respectively, we observe that increasing the weight given to the
datasets. Error bars are equivalent throughout the figure. b Same earlier time steps of evolution for each realization results in a
error metric, but in future time steps never seen during the training DeepONet model capable of recovering smaller, high-frequency
phase. Error bars are equivalent throughout the figure. components. This ability to accurately resolve multiple length
scales is particularly important in dynamic problems such as
DeepONet predictions were then remapped to the primitive space dendrite or grain growth problems, for instance, where the
using the pre-trained convolutional decoder to recreate the simulated microstructure dynamics can be extremely sensitive to
microstructure at the required time step. We plot the mean the development of multiple length scales concurrently.
relative L2 forecasting error corresponding to the models trained
on differently spaced datasets in Fig. 4. As expected, just like with
Integration of DeepONet with a numerical high-fidelity phase-
any other time-integration scheme, we observe that the forecast-
field solver
ing error increases for larger spacing in physical time. However,
the computational efficiency of the DeepONet predictions The results above show that a pre-trained autoencoder–DeepONet
remained the same for predicting consecutive time frames model can be used as a robust and efficient surrogate of the
regardless of the time spacing used. numerical solver when inference is requested for initial microstruc-
ture and parameters within the distributions of the training datasets
(interpolation task). Our proposed framework can also be used for
Training strategy for learning concurrently multi-scale extrapolation tasks and be integrated into the phase-field numerical
features solver to accelerate the predictions for initial microstructure and
We showed in a previous work37,38 that the microstructure parameters that are outside the aforementioned distributions
evolution in latent space is non-linear. Indeed, for early time steps (extrapolation task). To demonstrate this point, we devised a hybrid
the microstructure evolves rapidly and then later on it evolves approach that integrates the autoencoder–DeepONet framework
more slowly once the phase separation dynamics have taken with our high-fidelity phase-field Mesoscale Multiphysics Phase Field
effect. Simulator (MEMPHIS solver). This hybrid model unites the efficiency
We noted that the DeepONet model architecture presented and computational speed of the autoencoder–DeepONet framework
above was not able to resolve the small-scale microstructural with the accuracy of high-fidelity phase-field numerical solvers.
features as shown in Fig. 2. In the early time steps, which represent The hybrid framework consists of alternating between predic-
fast dynamics, small wavelength features are hard to capture due tions from the high-fidelity phase-field simulations and those from
to spectral bias of neural networks; Fig. 3 quantifies this difficulty. the autoencoder–DeepONet model. The high-fidelity phase-field
To circumvent this issue, during training of the DeepONet simulation step provides accuracy in the description of the
model, we increased the weight given to earlier time steps of each dynamics, while the autoencoder–DeepONet model enables us to

npj Computational Materials (2022) 190 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 V. Oommen et al.
7

Fig. 5 Capturing small microstructural features at earlier time steps. a True microstructure at t = t10. b Predicted microstructure without
emphasizing earlier time steps during training. c Microstructure predicted by DeepONet trained on a dataset where the earlier time steps
where repeated to increase the importance given to earlier time steps.

‘leap in time’. The algorithm is presented in Algorithm 1. Here we For instance, one could easily consider to use very short time steps
choose to split the time evolution predicted between the high- within the high-fidelity phase-field solver window to course-
fidelity simulation and those of the autoencoder–DeepONet to be correct the physical predictions and much longer time steps when
equal to one another. Each solver within this integrated scheme using DeepONet to accelerate the time evolution predictions. This
sequentially predicts 10-time frames, which corresponds to type of time splitting integration scheme would dramatically
10 × 500k = 5M time steps for the high-fidelity phase-field solver increase the speedup even further. Additionally, although we
alone. A schematic of the approach and results are shown in Fig. 6 showed the microstructure evolution only until t = t105, such a
for the training data. Results for the test data are provided in hybrid time integration strategy can be adopted to forecast time
Supplementary Fig. 6. The discontinuity along the centerlines of evolution of the microstructure for time windows that can be
the predictions made by the autoencoder–DeepONet model arises much longer, for instance as long as the input time history used to
from splitting each realization into four different realizations, as train the DeepONet model while still keeping a good accuracy,
discussed in the ‘Microstructure-evolution dataset’ subsection. leading to substantial savings in CPU hours.

Algorithm 1. Integration of DeepONet with high-fidelity phase

field simulator (MEMPHIS) DISCUSSION
Require: ϕ0 : Initial condition: ϕ(x, y, 0) In this work, we investigated the effectiveness of a convolutional
Require: NT : Number of total time steps autoencoder–DeepONet approach for modeling the evolution
Require: nt : Initial number of time steps to be simulated by dynamics of mesoscale microstructures. The proposed framework
MEMPHIS consists of two parts. First, learning a non-linear mapping to a
Require: DONnt : Number of time steps to be leaped by latent manifold using convolutional autoencoders, and second,
DeepONet learning the dynamics in the latent space (from the first step)
n←0 ⊳ Initialize using DeepONet. We trained our model on high-fidelity, phase-
While n ! = NT do field data generated by solving the Cahn-Hilliard equation. The
ϕnt MEMPHISðϕ0 ; nt Þ ⊳ Solution from MEMPHIS results presented above show that the trained DeepOnet
ϕnt þDONnt DeepONetðϕnt Þ ⊳ Prediction from DeepONet architecture can be used robustly to replace the high-fidelity
ϕ0 ϕnt þDONnt ⊳ Update the input for MEMPHIS phase-field numerical solver in interpolation tasks or to speed up
n ← nt + DONnt ⊳ Leaping n: MEMPHIS + DeepONet the numerical solver for extrapolation tasks. We showed that
end while increasing the latent dimension used to describe the micro-
structure evolution and putting more emphasis on earlier time
We see in Fig. 6 that the forecasting from 10 time frames using steps during the training improve the overall representation
the high-fidelity phase-field solver MEMPHIS running on 32 CPU- capability of the framework. Given its performance, this framework
cores (Intel® Xeon®, e5-2670) takes approximately 90 min. The offers several advantages as compared to other machine-learned
subsequent 10 time frames predicted by the architectures used for accelerating the prediction of the phase-
autoencoder–DeepONet model take only 2 s. A comparison of field-based microstructure evolution.
the computational cost between the high-fidelity phase-field First, unlike existing methods37,40 that train machine-learning-
solver alone and the hybrid approach is reported in Table 3. Here, based surrogate models using low-dimensional representations of
we achieve a speed-up of 29% . This performance can be microstructures based on statistical functions (e.g. auto-correlation
improved by a much greater factor with more extensive offline function), our autoencoder–DeepONet approach learns a suitable
training with a richer dataset of operating conditions, which will low-dimensional latent space using a convolutional autoencoder.
lead to better generalization. For each evolution, our hybrid We showed that this approach bypasses any post-processing
approach saves 135 min, without loss of accuracy as shown steps (e.g. a phase-recovery algorithm) necessary to reconstruct
previously in the Results section. The choice of a specific time step the microstructure from statistical functions40. The advantage of
splitting, for which the system is evolved using the high-fidelity training DeepONet in the autoencoder latent space is two-fold. On
phase-field framework and then by the autoencoder–DeepONet one hand, training DeepONet in a low-dimensional space is
model, is arbitrary and can be considered as a hyper-parameter. computationally efficient. On the other hand, the presence of

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2022) 190
 V. Oommen et al.
8

Fig. 6 Schematic of our hybrid approach for integrating DeepONet with the numerical solver MEMPHIS to accelerate phase-field
predictions. The computational time corresponding to the autoencoder--DeepONet model and the MEMPHIS solver for one realization in the
training dataset is reported in this figure. The error is shown on the third column.

npj Computational Materials (2022) 190 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 V. Oommen et al.
9
high-gradient regions in the microstructure data (see Fig. 7a) can Second, even though the workflow presented in this work is
make the training of the model challenging. However, the encoder focused on two-dimensional (2D) microstructure data, it can easily
transforms microstructure data to a latent space (Fig. 7b), where be extended to three-dimensional (3D) microstructure data. For
the gradients are not as high and more gradual . In other words, the 3D microstructure evolution case, each realization can be
the encoder learns a non-linear mapping of the microstructure represented by a sequence of 3D tensor data structures. Hence,
data coming from an untrainable distribution in the primitive we could use an autoencoder with 3D convolution layers in the
space, as shown in Fig. 7c, to a trainable distribution in the low- encoder and 3D transpose convolution layers in the decoder and
dimensional latent space, as shown in Fig. 7d. Such a transforma- learn a suitable non-linear mapping to a low-dimensional latent
tion from data with a high gradient to data with gradual, manifold. Following the same approach, a DeepONet model can
smoother gradient facilitates the training of the DeepONet model. be trained to learn the dynamics in the latent space, and be then
Although we have trained our own autoencoder model in this remapped to primitive space using the already trained decoder.
work, we believe that fine-tuning any autoencoder pre-trained on As shown in a recent theoretical paper, DeepOnet can tackle the
existing image datasets with similarities to our microstructure curse of dimensionality in the input space, so training it in
images will be suitable for this task. Reusing such readily available highdimensions is not a prohibitive issue49.
Third, the autoencoder and DeepONet are trained solely from
pre-trained autoencoders can further save on the computational
data, making the proposed approach purely data-driven, inde-
cost of our workflow.
pendent of the boundary conditions. The boundary conditions are
never explicitly assumed as input data to the framework at any
stage. They are implicitly fed through the latent representation of
Table 3. A comparison between computational time for high-fidelity
phase-field simulations (MEMPHIS) and proposed hybrid model
the microstructure history data inputted to the branch network of
(Hybrid) for a single microstructure evolution realization from time
DeepONet. Therefore, any information regarding a change in the
frame t1 to time frame t105. boundary condition will be available in the latent microstructure
history fed to the branch network, enabling the DeepONet model
Computational time to predict the dynamics accordingly in the latent manifold. In this
manner, the purely data-driven nature makes the proposed
Time frames MEMPHIS Hybrid autoencoder–DeepONet framework agnostic to any changes in
t1 to t30 135 mins 135 mins boundary conditions within the considered history. We also need
to clarify that periodic boundary conditions were imposed at all
t30 to t40 45 mins 2 sec
four boundaries of the computational domain while generating
t40 to t60 90 mins 90 mins the data from the numerical solver MEMPHIS. We note that
t60 to t70 45 mins 2 sec DeepONet can be trained to map boundary conditions to an
t70 to t90 90 mins 90 mins output field if so desired for a very general set of variable
t90 to t100 45 mins 2 sec boundary conditions.
t100 to t105 23 mins 23 mins There are several extensions to the present framework that can
be implemented in order to improve the accuracy, predictability,
Total time 473 mins 338.1 mins
and acceleration performance. These improvements are related to

Fig. 7 Representation of microstructure data and statistical insights. a represents a 3D visualization of the function to be approximated by
the surrogate model. The presence of several high-gradient regions at every time-step, makes it challenging for neural network models to
learn the evolution dynamics of microstructures. Panel (b) represents a smoother latent-microstructure learned by the encoder during the
autoencoder training. c The microstructure data, ϕ(x, y, t), is predominantly represented by 1s or 0s. d The encoder transforms ϕ to a latent
~ where deep neural networks can learn easily. The curves in (c) and (d) represent the smoothed density estimates of the histogram.
space, ϕ,

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2022) 190
 V. Oommen et al.
10
the training of the model and extension to a multi-fidelity present model, both the mobility and the interfacial energy are taken to
implementation. The first topic is related to improving the be isotropic and ωc and κc are stet to unity for simplicity. The evolution of
accuracy of the model with physics constraints in order to better one phase is expressed as a symmetric double-well potential, with minima
capture non-linearities in the model evolution. The second topic is at ϕ ± 1.
related to fusing different sources of data within our dataset, e.g.,
‘low’ fidelity from simulation and ‘high’ fidelity from physical Microstructure-evolution dataset
experiments of the same nature. The phase-field model described above is implemented using Sandia’s in-
Regarding a physics-informed implementation, Wang et al.50 for house multi-physics phase-field modeling code MEMPHIS10,54. In order to
instance put forward a physics-informed DeepONet, where the generate a diverse and large set of simulation results exhibiting a rich
PDE of the underlying system is added as a soft constraint to the variety of microstructure features, we independently sampled the phase
loss functions. In the present study, the training framework is fraction ϕA, such that each phase has at least a minimum concentration of
purely data-driven and we are learning the dynamical system in 0.15 (note that ϕB = 1 − ϕA), and the phase mobilities MA and MB of
species ‘A’ and ‘B’. Phase mobilities are sampled independently to vary in
the latent space defined by non-primitive coordinates except for
the range [0.01, 100]. In total we generated 500 triplets (ϕA, MA, MB) using
time, which is fed as an input to the trunk net. However, similar to Latin Hypercube Sampling. In the simple case of the spinodal decomposi-
Wang et al., some physical constraints could be injected into the tion, only the tuple (ϕA, MA/MB) is necessary. As demonstrated in other
current framework by using, for instance, the fact that mass ϕ is studies10,40 that share similar microstructure evolution as the spinodal
conserved at all times. decomposition, it is, however, necessary to handle (ϕA, MA, MB) separately
Regarding the multi-fidelity implementation, the present since the ratio MA/MB by itself will not be sufficient anymore to characterize
approach can be extended to incorporate experimental data the dynamics of the microstructure evolution. Herein, we frame the
coming from similar processes. For instance, recently, several present work in a broader context for generality. All the simulations were
researchers51–53 explored diverse ways of exploiting the inherent performed using a two-dimensional (2D) square domain Ω = [0, 1] × [0, 1],
correlations between datasets coming from different sources of discretized with 512 × 512 grid points, with a dimensionless spatial
data with different levels of fidelity and obtained optimal discretization of unity in either direction, and a temporal discretization
of Δt = 1 × 10−4. The simulation domain’s composition field is initialized
predictions. In this context, the data obtained from phase-field using truncated random Gaussian distribution in the range [ − 1, 1] with
models using a numerical solver can be considered as a low- μ = ϕA, and σ = 0.35. The microstructure was allowed to evolve and grow
fidelity dataset and the limited amounts of experimental for 50,000,000 time steps, saving the state of the microstructural domain
microstructure image data from similar processes can be treated every 500,000 time steps, hence a total of 100-time frames were saved
as a high-fidelity dataset. The autoencoder–DeepONet framework from each simulated case.
proposed here can be extended to generate accurate predictions In order to use the data in the proposed algorithm, we down-sampled
from a limited number of high-fidelity experimental microstruc- each snapshot of our 512 × 512 domain into four images of 256 × 256, and
ture microscopy image data, by utilizing the high correlation with later used cubic interpolation55 to further reduce the resolution to
the surplus low-fidelity phase field data. The assimilation of 128 × 128. Hence, from the 500 microstructure evolution samples, we were
experimental data in the present DeepONet architecture can be able to generate 2000 microstructure evolution samples of 128 × 128
resolution. From this dataset, we have used 1600 cases for training the
concatenated with numerical data as another realization and the DeepONet and 400 cases for testing the network accuracy. Since the
proposed workflow will remain unchanged. Merging and taking compositional field is randomly distributed spatially, the microstructure
advantage of both experimental and modeling efforts is a future has no recognizable features at the first frame t0. The quick development
direction of our research. of subdomains is then observed between frames t0 and t10, followed by a
To summarize, we developed and applied a machine-learned smooth and steady coalescence and growth of the microstructure from
framework based on neural operators and autoencoder architec- time frames t10 to t100. We have trained our proposed model based on this
tures to efficiently and rapidly predict complex microstructural observation, starting at time frame, t10, when the microstructure had
evolution problems. Such an architecture is not only computa- reached a slow and steady development regime.
tionally efficient and accurate, but it is also robust to noisy data.
The demonstrated performance makes it an attractive alternative Training the autoencoder: learning the latent microstructure
to other existing machined-learned strategies to accelerate the representation
predictions of microstructure evolution. It opens up a computa- In this work, each microstructure evolution is represented by
tionally viable and efficient path forward for discovering, under- (NT, Nx, Ny) = (80 × 128 × 128), with NT representing the number of snap-
standing, and predicting materials processes, where evolutionary shots and Nx × Ny denoting the spatial resolution along x− and y−
mesoscale phenomena are critical, such as in the optimization and direction, respectively. To handle the entire feature space (R128 ´ 128 )
design of materials problems. 16,384 distinct features are required to represent the microstructure at
each time step. Subsequently, to compute the prediction for all 1,600
microstructure evolutions, we will have 1600 × 80 × 16, 384 (≈2.5 Billion)
32-bit floating data points. Learning microstructure dynamics from such a
METHODS
high-dimensional dataset is challenging.
Phase-field model of the spinodal decomposition of a two- To circumvent issues pertaining to the data dimensionality and
phase mixture preparing the phase-field microstructure data for DeepONet training, we
We illustrate our accelerated phase-field workflow on the simplest case of explored a couple of options. First, we tried using Principal Component
the spinodal decomposition of a two-phase mixture. This model is highly Analysis (PCA) with a linear kernel, for reducing the dimensionality of the
relevant to many phase-field models. In the spinodal decomposition of a data21,56,57. The low-dimensional representation of the data obtained from
two-phase mixture uses a single order parameter, ϕ(x, t) to describe the PCA is a linear transformation of the high-dimensional data and discards
atomic fraction of solute diffusing within a matrix. The free energy of the the insignificant modes (eigen/singular) corresponding to the lower eigen/
system is expressed by the Cahn-Hilliard equation based on the Onsager singular values (λi). However, the system considered here is non-diffusive,
force-flux relationship such that which is confirmed by cumulative explained variance and energy
  distribution of the system over principal modes. Therefore, using PCA for
∂ϕ
¼ ∇  Mc ðϕÞ∇½ωc ðϕ3  ϕÞ þ κ c ∇2 ϕ ; (3) reducing the dimensionality of the microstructure description could result
∂t in the loss of valuable information, if only a convenient low number of
where ωc is the height of the energy barrier between the two phases, κc is principal components are considered. A detailed explanation on PCA of
the gradient energy coefficient, and Mc denotes the concentration the microstructure dataset is presented in Supplementary Note 1. Learning
dependent mobility, with Mc = s(ϕ)MA + (1 − s(ϕ))MB. The function s a non-linear mapping from a high-dimensional to a low-dimensional latent
defines a smooth interpolation to switch from phase ‘A’ to phase ‘B’. This space is one way to compress data without losing as much information as
interpolation function is defined as sðϕÞ ¼ 14 ð2  ϕÞð1 þ ϕÞ2 . In the in the PCA. An autoencoder precisely does this by learning a non-linear

npj Computational Materials (2022) 190 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 V. Oommen et al.
11
transformation to a low-dimensional latent space using an encoder. The DeepONet, θd, are obtained by minimizing a loss function, ↕d, defined as:
decoder learns the mapping to retrieve initial high-dimensional data from  2
Ld ¼ min ϕðtÞ ~
~  GðΦÞðt; θd Þ2 ; (7)
its latent representation. θd
In this study, we have used a convolutional autoencoder58 with
~
where ϕðtÞ is the ground truth for the low-dimensional phase field
convolutional layers in the encoder and transpose convolutional layers in
the decoder as shown in Fig. 1. The encoder learns a nonlinear mapping of representation at time, t obtained from the convolutional encoder. The
trained DeepONet is used to predict ϕðtÞ ~ 2 Rld . The output of the
the high-dimensional microstructure data, ϕ(x, y, t), to a low-dimensional
~
latent space represented by ϕðtÞ and is expressed as DeepONet is fed into the transposed convolutional decoder to predict,
^ 2 R128 ´ 128 . The DeepONet is trained using the Adam optimizer59. The
ϕðtÞ
~
αθenc : ϕðx; y; tÞ ! ϕðtÞ; (4) implementation has been carried out using the TensorFlow framework60.
We use Xavier Initialization61 to initialize the weights of all the models.
~ ! ϕðx;
βθdec : ϕðtÞ ^ y; tÞ; (5)
Error metrics
where α and β represent the mappings performed by the encoder and the The L2 norm of relative error, D, is used as the evaluation metric to analyze
decoder, respectively. In Eq. 4, the encoder takes ϕðx; y; tÞ 2 R128 ´ 128 as the performance of each model considered in this study. D is defined as:
~ 2 Rld , where ld is the dimension of the latent
input, and maps it to ϕðtÞ
space. θenc represents the trainable parameters of the convolutional P P P P  ðnÞ 2
^ ðnÞ
n x y t ϕ ðx; y; tÞ  ϕ ðx; y; t; θÞ
encoder. Equation 5 represents the decoder network, which takes the D¼ P P P P ðnÞ ; (8)
~ 2 Rld as the input and predicts the
latent dimensional representation, ϕðtÞ 2
n x y t ϕ ðx; y; tÞ
^ y; tÞ 2 R128 ´ 128 , using transpose convolu-
primitive microstructure, ϕðx;
tional operations. The details of the autoencoder architecture are provided where n corresponds to the nth sample of the given dataset.
in Table 1.
θae represents the trainable parameters of the autoencoder. These
parameters are learned by minimizing the loss function, Lae , which reads DATA AVAILABILITY
The data that support the findings of this study are available from the corresponding
Lae ¼ min ^ y; t; θae Þk :
kϕðx; y; tÞ  ϕðx;
2
(6) authors upon reasonable request.
2
θae ¼fθenc ;θdec g

Alternatively, the autoencoder provides a low-dimensional representa-

tion of the microstructure by learning a non-linear transformation to a CODE AVAILABILITY
latent space with ld features. We also observe that it is easier to learn the The codes developed (in Python3) to perform the computational experiments in the
microstructure dynamics in the latent space representation, learned by the present study are freely available at this git repository.
autoencoder, than the original primitive form of the microstructure in real
space. This is due to the presence of several high gradient regions in the Received: 20 April 2022; Accepted: 17 August 2022;
original form of the microstructure as shown in Fig. 7a. These high gradient
regions in the solution are due to the nature of the governing Cahn-Hilliard
equation. The latent microstructure representation in Fig. 7b is smoother
and does not have high gradient regions. The latent representation of the
data offers higher regularity and therefore, we achieve faster convergence
REFERENCES
during the training of the surrogate neural network model.
1. Chen, L.-Q. Phase-field models for microstructure evolution. Annu. Rev. Mater. Res.
32, 113–140 (2002).
Training the DeepONet: learning the microstructure dynamics 2. Moelans, N., Blanpain, B. & Wollants, P. An introduction to phase-field modeling
in lower dimensions of microstructure evolution. Calphad 32, 268–294 (2008).
Neural operators generate nonlinear mappings across infinite-dimensional 3. Yang, Z., Yu, C.-H. & Buehler, M. J. Deep learning model to predict complex stress
function spaces on bounded domains, giving a simulation framework for and strain fields in hierarchical composites. Sci. Adv. 7, eabd7416 (2021).
multidimensional complex dynamics prediction in real time. Once properly 4. Pokharel, R., Pandey, A. & Scheinker, A. Physics-informed data-driven surrogate
trained, such models are discretization invariant, which means they share modeling for full-field 3D microstructure and micromechanical field evolution of
the same network parameters regardless of how the underlying functional polycrystalline materials. JOM 73, 3371–3382 (2021).
data is parameterized. DeepONet, originally proposed by Lu and 5. Chakraborty, S., Goswami, S. & Rabczuk, T. A surrogate assisted adaptive frame-
coworkers41, allows the mapping between infinite-dimensional functions work for robust topology optimization. Computer Methods Appl. Mech. Eng. 346,
using deep neural networks. This subsection provides a detailed 63–84 (2019).
description of the training of DeepONet to model the evolution of the 6. Olsson, E. & Kreiss, G. A conservative level set method for two phase flow. J.
microstructure in the latent dimension. Computational Phys. 210, 225–246 (2005).
The unstacked DeepONet architecture is made up of two concurrent 7. Yue, P., Zhou, C., Feng, J. J., Ollivier-Gooch, C. F. & Hu, H. H. Phase-field simulations
deep neural networks: one encodes the input function at fixed sensor of interfacial dynamics in viscoelastic fluids using finite elements with adaptive
meshing. J. Computational Phys. 219, 47–67 (2006).
locations (branch network), while the other represents the domain of the
8. Bharali, R., Goswami, S., Anitescu, C. & Rabczuk, T. A robust monolithic solver
output function (trunk network). Time, t 2 R1 , is given as input to the
trunk network while Φ ~ ¼ fϕðt
~ 10 Þ; ϕðt
~ 11 Þ; ¼ ; ϕðt
~ 89 Þg 2 R80 ´ ld is the input for phase-field fracture integrated with fracture energy based arc-length
fed to the branch network. Φ ~ represents the phase field in the latent method and under-relaxation. Computer Methods Appl. Mech. Eng. 394, 114927
(2022).
dimension, ld, for all the 80-time steps available in the given dataset. The
goal of the DeepONet is to learn the solution operator, ϕðtÞ ~  ϕðtÞ ^
~ ¼ 9. Goswami, S., Anitescu, C. & Rabczuk, T. Adaptive fourth-order phase field analysis
~
GðΦÞðtÞ from the 1600 microstructure evolutions provided in the training
for brittle fracture. Computer Methods Appl. Mech. Eng. 361, 112808 (2020).
10. Stewart, J. A. & Dingreville, R. Microstructure morphology and concentration
dataset. The output of the DeepONet is a vector 2 Rld and is expressed as
~
GðΦÞðt; θd Þ, where θd ¼ fWd ; bd g includes the trainable weights, Wd, and
modulation of nanocomposite thin-films during simulated physical vapor
deposition. Acta Materialia 188, 181–191 (2020).
biases, bd, of the DeepONet model. The framework of the
11. Powers, M., Stewart, J. A., Dingreville, R., Derby, B. K. & Misra, A. Compositionally-
DeepONet allows the branch network to have a flexible architecture. To
driven formation mechanism of hierarchical morphologies in co-deposited
model the microstructure evolution, we have considered a fully connected
immiscible alloy thin films. Nanomaterials 11, 2635 (2021).
neural network for the trunk network. Due to the high-dimensional nature
12. Beyerlein, I. & Hunter, A. Understanding dislocation mechanics at the mesoscale
of the branch network input, R80 ´ ld , a convolutional neural network is
using phase field dislocation dynamics. Philos. Trans. R. Soc. A: Math., Phys. Eng.
used as the branch network because it utilizes the same kernels across the Sci. 374, 20150166 (2016).
time axis and enables the branch network to encode the entire history in a 13. Elliott, C. M. & French, D. A. Numerical studies of the Cahn-Hilliard equation for
memorypffiffiffi efficient
pffiffiffi manner. Hence, the input has to be reshaped to phase separation. IMA J. Appl. Math. 38, 97–128 (1987).
R80 ´ ld ´ ld before feeding it to the branch network. The network 14. Cahn, J. W. & Hilliard, J. E. Free energy of a nonuniform system. I. Interfacial free
architecture is presented in Fig. 1. The trainable parameters of the energy. J. Chem. Phys. 28, 258–267 (1958).

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2022) 190
 V. Oommen et al.
12
15. Sun, Z. Z. A second-order accurate linearized difference scheme for the two- 40. Herman, E., Stewart, J. A. & Dingreville, R. A data-driven surrogate model to
dimensional Cahn-Hilliard equation. Math. Comput. 64, 1463–1471 (1995). rapidly predict microstructure morphology during physical vapor deposition.
16. Liu, C. & Shen, J. A phase field model for the mixture of two incompressible fluids Appl. Math. Model. 88, 589–603 (2020).
and its approximation by a Fourier-spectral method. Phys. D: Nonlinear Phenom. 41. Lu, L., Jin, P., Pang, G., Zhang, Z. & Karniadakis, G. E. Learning nonlinear operators
179, 211–228 (2003). via DeepONet based on the universal approximation theorem of operators. Nat.
17. Barrett, J. W., Blowey, J. F. & Garcke, H. Finite element approximation of the Cahn- Mach. Intell. 3, 218–229 (2021).
Hilliard equation with degenerate mobility. SIAM J. Numer. Anal. 37, 286–318 42. Lin, C. et al. Operator learning for predicting multiscale bubble growth dynamics.
(1999). J. Chem. Phys. 154, 104118 (2021).
18. Gómez, H., Calo, V. M., Bazilevs, Y. & Hughes, T. J. Isogeometric analysis of the 43. Osorio, J. D. et al. Forecasting solar-thermal systems performance under transient
Cahn-Hilliard phase-field model. Computer Methods Appl. Mech. Eng. 197, operation using a data-driven machine learning approach based on the deep
4333–4352 (2008). operator network architecture. Energy Convers. Manag. 252, 115063 (2022).
19. Chan, C. L., Anitescu, C. & Rabczuk, T. Strong multipatch C1-coupling for iso- 44. Cai, S., Wang, Z., Lu, L., Zaki, T. A. & Karniadakis, G. E. Deepm&mnet: Inferring the
geometric analysis on 2D and 3D domains. Computer Methods Appl. Mech. Eng. electroconvection multiphysics fields based on operator approximation by neural
357, 112599 (2019). networks. J. Computational Phys. 436, 110296 (2021).
20. Alikakos, N. D., Bates, P. W. & Chen, X. Convergence of the Cahn-Hilliard equation 45. Vincent, P., Larochelle, H., Bengio, Y. & Manzagol, P.-A.Extracting and composing
to the Hele-Shaw model. Arch. Ration. Mech. Anal. 128, 165–205 (1994). robust features with denoising autoencoders. In Proceedings of the 25th Inter-
21. Brough, D. B., Kannan, A., Haaland, B., Bucknall, D. G. & Kalidindi, S. R. Extraction of national Conference on Machine Learning, 1096–1103 (2008).
process-structure evolution linkages from X-ray scattering measurements using 46. Vincent, P. et al. Stacked denoising autoencoders: Learning useful representa-
dimensionality reduction and time series analysis. Integrating Mater. Manuf. tions in a deep network with a local denoising criterion. Journal of Machine
Innov. 6, 147–159 (2017). Learning Research 11 (2010).
22. Pfeifer, S., Wodo, O. & Ganapathysubramanian, B. An optimization approach to 47. Gondara, L.Medical image denoising using convolutional denoising auto-
identify processing pathways for achieving tailored thin film morphologies. encoders. In 2016 IEEE 16th International Conference on Data Mining Workshops
Computational Mater. Sci. 143, 486–496 (2018). (ICDMW), 241–246 (IEEE, 2016).
23. Teichert, G. H. & Garikipati, K. Machine learning materials physics: Surrogate 48. Bostanabad, R. et al. Computational microstructure characterization and recon-
optimization and multi-fidelity algorithms predict precipitate morphology in an struction: Review of the state-of-the-art techniques. Prog. Mater. Sci. 95, 1–41 (2018).
alternative to phase field dynamics. Computer Methods Appl. Mech. Eng. 344, 49. Lanthaler, S., Mishra, S. & Karniadakis, G. E. Error estimates for DeepONets: A deep
666–693 (2019). learning framework in infinite dimensions. Trans. Math. Its Appl. 6, tnac001 (2022).
24. Lin, C., Maxey, M., Li, Z. & Karniadakis, G. E. A seamless multiscale operator neural 50. Wang, S., Wang, H. & Perdikaris, P. Learning the solution operator of parametric
network for inferring bubble dynamics. J. of Fluid Mechanics. 929 (2021). partial differential equations with physics-informed deeponets. Sci. Adv. 7,
25. Zhang, X. & Garikipati, K. Machine learning materials physics: Multi-resolution eabi8605 (2021).
neural networks learn the free energy and nonlinear elastic response of evolving 51. De, S., Hassanaly, M., Reynolds, M., King, R. N. & Doostan, A.Bi-fidelity modeling of
microstructures. Computer Methods Appl. Mech. Eng. 372, 113362 (2020). uncertain and partially unknown systems using deeponets. Preprint at https://
26. Goswami, S., Yin, M., Yu, Y. & Karniadakis, G. E. A physics-informed variational arxiv.org/abs/2204.00997 (2022).
deeponet for predicting crack path in quasi-brittle materials. Computer Methods 52. Howard, A. A., Perego, M., Karniadakis, G. E. & Stinis, P.Multifidelity deep operator
Appl. Mech. Eng. 391, 114587 (2022). networks. Preprint at https://arxiv.org/abs/2204.09157 (2022).
27. Li, Z. et al. Fourier neural operator for parametric partial differential equations. Preprint 53. Lu, L., Pestourie, R., Johnson, S. G. & Romano, G. Multifidelity deep neural
at: https://arxiv.org/abs/2010.08895 (2020). operators for efficient learning of partial differential equations with application to
28. Goswami, S., Anitescu, C., Chakraborty, S. & Rabczuk, T. Transfer learning fast inverse design of nanoscale heat transport. Phys. Rev. Res. 4, 023210 (2022).
enhanced physics informed neural network for phase-field modeling of fracture. 54. Dingreville, R. P. M., Stewart, J. A., Chen, E. Y. & Monti, J. M.Benchmark problems
Theor. Appl. Fract. Mech. 106, 102447 (2020). for the Mesoscale Multiphysics Phase Field Simulator (MEMPHIS). Tech. Rep.,
29. Kunselman, C., Attari, V., McClenny, L., Braga-Neto, U. & Arroyave, R. Semi- Sandia National Laboratories, Albuquerque, NM (United States) (2020).
supervised learning approaches to class assignment in ambiguous micro- 55. Süli, E. & Mayers, D. F. An Introduction to Numerical Analysis (Cambridge University
structures. Acta Materialia 188, 49–62 (2020). Press, 2003).
30. Haghighat, E., Raissi, M., Moure, A., Gomez, H. & Juanes, R. A physics-informed 56. Niezgoda, S. R., Kanjarla, A. K. & Kalidindi, S. R. Novel microstructure quantification
deep learning framework for inversion and surrogate modeling in solid framework for databasing, visualization, and analysis of microstructure data.
mechanics. Computer Methods Appl. Mech. Eng. 379, 113741 (2021). Integrating Mater. Manuf. Innov. 2, 54–80 (2013).
31. Zhang, K. et al. High-throughput phase-field simulations and machine learning of 57. Gupta, A., Cecen, A., Goyal, S., Singh, A. K. & Kalidindi, S. R. Structure–property
resistive switching in resistive random-access memory. npj Computational Mater. linkages using a data science approach: application to a non-metallic inclusion/
6, 1–10 (2020). steel composite system. Acta Materialia 91, 239–254 (2015).
32. Goswami, S., Anitescu, C. & Rabczuk, T. Adaptive fourth-order phase field 58. Lee, K. & Carlberg, K. T. Model reduction of dynamical systems on nonlinear
analysis using deep energy minimization. Theor. Appl. Fract. Mech. 107, 102527 manifolds using deep convolutional autoencoders. J. Computational Phys. 404,
(2020). 108973 (2020).
33. Attari, V. & Arroyave, R. Machine learning-assisted high-throughput exploration of 59. Kingma, D. & Ba, J. Adam: a method for stochastic optimization. International
interface energy space in multi-phase-field model with calphad potential. Mater. Conference on Learning Representations. (2014).
Theory 6, 1–20 (2022). 60. Abadi, M. et al. TensorFlow: Large-scale machine learning on heterogeneous
34. Samaniego, E. et al. An energy approach to the solution of partial differential distributed systems. Software available from tensorflow.org. 1 (2015).
equations in computational mechanics via machine learning: Concepts, imple- 61. Glorot, X. & Bengio, Y. Understanding the difficulty of training deep feedforward
mentation and applications. Computer Methods Appl. Mech. Eng. 362, 112790 neural networks. Proceedings of the Thirteenth International Conference on Artificial
(2020). Intelligence and Statistics, 249–256 (2010).
35. Shukla, K., Jagtap, A. D., Blackshire, J. L., Sparkman, D. & Em Karniadakis, G. A
physics-informed neural network for quantifying the microstructural properties
of polycrystalline nickel using ultrasound data: A promising approach for solving ACKNOWLEDGEMENTS
inverse problems. IEEE Signal Process. Mag. 39, 68–77 (2022). R.D. and G.E.K. acknowledge funding under the BeyondFingerprinting Sandia Grand
36. Perera, R., Guzzetti, D. & Agrawal, V. Graph neural networks for simulating crack Challenge Laboratory Directed Research and Development (GC LDRD) program. V.O.
coalescence and propagation in brittle materials. Computer Methods Appl. Mech. and K.S. acknowledge Dr. Elton Chen from Sandia National Laboratories for their
Eng. 395, 115021 (2022). valuable discussions that enabled us to better understand the MEMPHIS code. V.O. and
37. Montes de Oca Zapiain, D., Stewart, J. A. & Dingreville, R. Accelerating phase-field- K.S. acknowledge Dr. Aniruddha Bora for his insightful suggestions and valuable
based microstructure evolution predictions via surrogate models trained by discussions at various stages of this project. The phase-field framework is supported by
machine learning methods. npj Computational Mater. 7, 1–11 (2021). the Center for Integrated Nanotechnologies (CINT), an Office of Science user facility
38. Hu, C., Martin, S. & Dingreville, R. Accelerating phase-field predictions via recur-
operated for the U.S. Department of Energy. This research was conducted using
rent neural networks learning the microstructure evolution in latent space.
computational resources and services at the Center for Computation and Visualization,
Computer Methods Appl. Mech. Eng. 397, 115128 (2022).
Brown University. Sandia National Laboratories is a multi-mission laboratory managed
39. Fullwood, D. T., Niezgoda, S. R. & Kalidindi, S. R. Microstructure reconstructions
and operated by National Technology and Engineering Solutions of Sandia, LLC., a
from 2-point statistics using phase-recovery algorithms. Acta Materialia 56,
wholly owned subsidiary of Honeywell International, Inc., for the U.S. Department of
942–948 (2008).
Energy National Nuclear Security Administration under contract DE-NA0003525. This

npj Computational Materials (2022) 190 Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences
 V. Oommen et al.
13
paper describes objective technical results and analysis. Any subjective views or Reprints and permission information is available at http://www.nature.com/
opinions that might be expressed in the paper do not necessarily represent the views of reprints
the U.S. Department of Energy or the United States Government.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims
in published maps and institutional affiliations.
AUTHOR CONTRIBUTIONS
R.D. and G.E.K. designed the study and supervised the project. V.O. and K.S.
developed the autoencoder–DeepONet architecture and the hybrid model. V.O., K.S.
and S.G. analyzed the results. V.O., K.S., S.G., R.D. and G.E.K. wrote the manuscript. Open Access This article is licensed under a Creative Commons
Attribution 4.0 International License, which permits use, sharing,
adaptation, distribution and reproduction in any medium or format, as long as you give
appropriate credit to the original author(s) and the source, provide a link to the Creative
COMPETING INTERESTS
Commons license, and indicate if changes were made. The images or other third party
The authors declare no competing interests. material in this article are included in the article’s Creative Commons license, unless
indicated otherwise in a credit line to the material. If material is not included in the
article’s Creative Commons license and your intended use is not permitted by statutory
ADDITIONAL INFORMATION regulation or exceeds the permitted use, you will need to obtain permission directly
Supplementary information The online version contains supplementary material from the copyright holder. To view a copy of this license, visit http://
available at https://doi.org/10.1038/s41524-022-00876-7. creativecommons.org/licenses/by/4.0/.

Correspondence and requests for materials should be addressed to Rémi Dingreville

or George Em Karniadakis. © The Author(s) 2022

Published in partnership with the Shanghai Institute of Ceramics of the Chinese Academy of Sciences npj Computational Materials (2022) 190