My-journal manuscript No.

(will be inserted by the editor)

Embedding an ANN-Based Crystal Plasticity Model into the Finite Element

Framework using an ABAQUS User-Material Subroutine

Yuqing He1 · Yousef Heider1,2,∗ · Bernd Markert1

September 25, 2024

Abstract This manuscript presents a practical method for incorporating trained Neural Networks (NNs) into the
Finite Element (FE) framework using a user material (UMAT) subroutine. The work exemplifies crystal plasticity, a
complex inelastic non-linear path-dependent material response, with a wide range of applications in ABAQUS UMAT.
However, this approach can be extended to other material behaviors and FE tools. The use of a UMAT subroutine
serves two main purposes: (1) it predicts and updates the stress or other mechanical properties of interest directly from
the strain history; (2) it computes the Jacobian matrix either through backpropagation or numerical differentiation,
which plays an essential role in the solution convergence. By implementing NNs in a UMAT subroutine, a trained
machine learning model can be employed as a data-driven constitutive law within the FEM framework, preserving
multiscale information that conventional constitutive laws often neglect or average. The versatility of this method
makes it a powerful tool for integrating machine learning into mechanical simulation. While this approach is expected
to provide higher accuracy in reproducing realistic material behavior, the reliability of the solution process and the
convergence conditions must be paid special attention. While the theory of the model is explained in [20], exemplary
source code is also made available for interested readers [https://doi.org/10.25835/6n5uu50y].
Keywords Machine learning · Path-dependent material behavior · Crystal plasticity · Finite element · User material
subroutine · UMAT · ABAQUS

Abbreviations

ANN Artificial Neural Network UMAT User Material Subroutine

FE Finite Element FEM Finite Element Method
CNN Convolutional Neural Network DSA Dynamic Strain Aging
LSTM Long Short-Term Memory RNN Recurrent Neural Network
DDSDDE Material Jacobian Matrix HDF5 Hierarchical Data Format Version 5
ML Machine Learning IBVP Initial Boundary Value Problem
f.c.c. Face-Centered Cubic FFNN Feed-Forward Neural Network

1 Introduction

In solid mechanics, the conventional constitutive laws explicitly formulate the stress-strain relations with symbolic
expressions. Such relations are mostly constructed from simplified assumptions and based on empirical observations.
1 Institute
of General Mechanics, RWTH Aachen University.
Eilfschornsteinstr. 18, 52062 Aachen, Germany
2Institute of Mechanics and Computational Mechanics (IBNM), Leibniz University Hannover.
Appelstr. 9A, 30167 Hannover, Germany
∗ Corresponding author: Yousef Heider.
E-mail: yousef.heider@ibnm.uni-hannover.de
2 Yuqing He1 et al.

Thus, the microscopic information is partially or entirely neglected from the material models [12]. However, data-
driven constitutive models make it possible to describe the stress-strain relations merely based on acquired data, with-
out specifying a certain constitutive model, see, e.g., [17, 19, 20]. Machine learning (ML) techniques can be applied to
capture such constitutive relations. Hence, integrating ML models as a substitution of model-based constitutive laws
into the FE framework offers the potential to incorporate lower-scale information, without explicitly specifying any
material parameter in a model. In this regard, Lefik and Schrefler [29] used ML to model elasto-plasticity and biaxial
non-linear behavior, whereas Oishi and Yagawa [34] used ML for the stiffness matrix in FEM. Tao et al. [39] utilized
ML to fit the engineering constants of the constituents of a fiber-reinforced composite, as well as the progressive dam-
age constitutive law. Xu et al. [41] built the Cholesky-factored symmetric positive definite neural network to predict
constitutive relations in dynamic equations. Pantidis and Mobasher [35] employed a ML model that utilized material
point deformation as input and produced element Jacobian matrix and residual vector as output. In other works, like
by Hashash et al. [16], Huang et al. [24], Li et al. [30], ML models are trained to acquire the material’s constitutive
laws. These models utilized strain components as input and stress components as output. In the work of Koeppe et al.
[26], a hybrid substructuring approach is introduced using the TRUNet deep neural network to develop a nonlinear,
inelastic intelligent meta element that accelerates finite element simulations in history-dependent mechanical prob-
lems with reduced computational effort. While the focus of the current work is on regression models, convolutional
neural networks (CNN), as a class of ANN, are also widely used within multiscale material modeling to generate
ML-based constitutive models. CNNs require mostly data that have a grid-like nature, like images or time-series data,
see, e.g., [2, 3, 6, 9, 10, 14, 22, 38] for review and applications.

Several research groups worked on the implementation of ML material models within the Finite Element Method
(FEM) as an alternative to traditional constitutive models. To mention some, Guan et al. [15] implemented their ML
constitutive model within multi-scale granular materials modeling. In this, the FEM and the Discrete Element Method
(DEM) are coupled to generate training samples. They showed that the FEM-ML framework offers considerable
improvements in terms of computational efficiency and the ability to simulate the mechanical responses of granular
materials. In the work of Suh et al. [37], an open-access PyTorch-ABAQUS deep-learning framework for a group of
plasticity models is presented. In this, an interface code is presented in which the weights and biases of the trained
NNs are automatically converted into a FORTRAN code that is compatible with the UMAT/VUMAT of ABAQUS. In
connection with unsaturated porous media, Heider et al. [21] implemented a trained ML material model of retention
curve within a Python FE framework. They discussed through numerical examples the accuracy improvement due
to ML material models as well as the instability sources. In the work of Weber et al. [40], challenges associated
with the application of ANN elasto-plasticity material models within FE calculations together with implementation
instructions are presented. As a remedy for some challenges, they proposed an approach that includes constrained
neural network training. An application of ANN in modeling Dynamic Strain Aging (DSA) in stainless steel and the
implementation of the developed ANN-based model in the FE formulation is presented by Patra et al. [36]. In the
work of Lourenço et al. [31], the use of ML techniques for improving the accuracy of material constitutive models in
metal plasticity using ABAQUS VUMAT is discussed. In Jang et al. [25], a J2-plasticity constitutive model to predict
elastoplastic behavior is proposed, where the ML model is implemented in ABAQUS UMAT. The ML model is used
only for nonlinear plastic loading while leaving linear elastic loading and unloading to the physics-based model. They
reported that the results of their model are in good agreement with those from the conventional constitutive models for
a given boundary-value problem. For multiscale modeling, Garcia-Teijeiro and Rodriguez-Herrera [11] proposed a
methodology that combines FEM with ANN in the numerical modeling of systems with behavior that involves a wide
span of spatial scales. In the context of biomechanics, Lee et al. [28] discussed the potential of using the ML-based
approach in solving FE-based classification questions.

While feed-forward neural networks (FFNN) can be used to train path-independent material models (e.g., elas-
ticity), recurrent neural networks (RNN) are used to capture responses that depend on the deformation history,
i.e. path- or time-dependent responses. This includes for instance, viscoelasticity and elasto-plasticity, see, e.g.,
[13, 20, 27, 32, 33, 42]. In this context, the Long Short-Term Memory (LSTM) neural network, an integral com-
ponent of the RNN paradigm introduced by Hochreiter and Schmidhuber [23], is considered in this work. In this
manuscript, we build upon the model presented by Heider et al. [20], which introduced a graph-based neural net-
work model focusing on the material point response in crystal plasticity. This model is extended later in [19] to
fulfill in addition to objectivity also the dissipation inequality. Our current work specifically proposes a methodology
that seamlessly integrates LSTM neural networks into the FEM framework of ABAQUS. The goal is to employ the
trained machine learning constitutive model as an alternative to the traditional crystal plasticity material model by
Embedding an ANN-Based Crystal Plasticity Model in ABAQUS UMAT 3

using the user-defined material (UMAT) subroutine of ABAQUS. In summary, the highlights of this manuscript can
be summarized in the following points:
q The primary focus is on the history-dependent or path-dependent material mechanical behaviors, such as those
encountered in crystal plasticity.
q An approach is proposed to facilitate the integration of machine-learning-based material models into FE analysis.
In this work, the simulations are conducted in ABAQUS.
q This paper adopts the UMAT subroutine designed especially for single-grain crystal plasticity materials as the test
platform.
q This subroutine inherently ensures the fulfillment of equilibrium equations and employs the Newton-Raphson
method for convergence, augmenting the accuracy and efficiency of the simulation outcomes.

2 Theoretical fundamentals

2.1 Data-based ML plasticity model and data preparation

2.1.1 Overview and procedures

Normalization of
Define simulation settings model inputs Solve the global equations
Tarined machine (node, element, boundary Predict stress
learning model conditions, etc.) Calculate residual, iterate
(*.hdf5) by Newton-Raphson Output
Compute Jacobian matrix
method
*.odb,
Define weights and biases
Extract weights, Convergence check *.dat
as material properties Reverse scale the
biases, layer info model ouptuts

Input file (*.inp) UMAT (*.f) Implicit solver

Model preparation Model implementation in Abaqus

Fig. 1: Illustration of the process flow: From model preparation to implementation in ABAQUS UMAT. The input
file and the UMAT subroutine file can be modified within the ABAQUS components, while the implicit solver is the
standard integral ABAQUS component.

As illustrated in Fig. 1, integrating a data-driven material model into a FE solver starts with building, training, and
validating the neural networks (NNs) using pre-generated datasets. Details and references regarding dataset genera-
tion and training settings are presented in Section 2.1.2. Subsequently, the trained models are saved in Hierarchical
Data Format version 5 (HDF5) files, which encompass the model architecture (weights, biases) and hyperparame-
ters (number of hidden layers, learning rate, activation function, etc.). These parameters can be transferred to the
ABAQUS interface via user-defined material constants. Notably, this process is implemented in the Python program-
ming language with the libraries of Tensorflow [1] and Keras version 2.12.0 [7].
To configure the simulation, an ABAQUS input file is used to specify or overwrite various parameters within
the simulation environment. These parameters include mesh nodes and elements, loads, boundary conditions, as well
as initial conditions. Within ABAQUS, the UMAT subroutine customizes the material responses by performing two
functions:
1. It predicts and updates the stresses STRESS(NTENS) on a material point based on the normalized strain history
using the forward path of the trained LSTM neural network. For general three-dimensional (3D) problems, the
symmetric stress tensor comprises NTENS=6 components.
4 Yuqing He1 et al.

2. It computes the Jacobian square matrix DDSDDE (NTENS, NTENS) representing the 2D format of the material
∂∆σI
Jacobian matrix with NTENS rows and NTENS columns. Each element DDSDDE (I, J) = signifies the
∂∆ϵJ
change in the I th stress component caused by an infinitesimal perturbation of the J th component of the strain
increment array. DDSDDE can be computed either through NN’s backward propagation or numerical differenti-
ation. Due to the path dependency of the LSTM neural network, DDSDDE is continually updated using previous
stresses and strains.

The material constants PROPS(NPROPS), which represent the previously extracted NN’s settings, are defined in
the ABAQUS input file. These constants enable ABAQUS to incorporate the trained neural network into the solving
process during simulation.
Since the ABAQUS implicit solver uses unscaled values for iteration, both predicted variables must be reverse-
scaled into their original value ranges. Following the approximation of a converged solution with the ABAQUS
implicit solver, the simulation results are exported as *.odb or *.dat files for subsequent postprocessing and plotting.

2.1.2 Generation of the training datasets

The synthetic datasets are generated by subjecting a face-centered cubic (f.c.c.) crystal to stepwise increasing strain-
controlled loadings, resulting in strain-stress material responses. The overall crystal stress is generated by rotating the
slip systems based on Borja’s “ultimate algorithm” [5], which is a stable solution of the micro-scale crystal plasticity
model. In particular, an f.c.c crystal with a maximum of 12 activatable slip systems at 49 different configurations
is subjected to strain-controlled loading, yielding a history of strain-stress components. A total of 1176 experiments
were conducted, comprising 24 numerical experiments for each of the 49 crystal configurations. The crystal was
subjected to monotonic incremental loading of one strain component during each numerical experiment. Specifically,
100 increments were applied for each experiment, which allowed capturing the history-dependent constitutive relation
up to 100 history steps. The dataset generation process should ensure an even distribution of data points throughout
the entire range of interest. The material parameters and the loading cases are listed in the preceding research of the
second author [20]. For training purposes, the database is also divided into training, validation, and testing subsets.
The generated database input and output are also scaled into the range [0,1] using the MinMaxScaler object from the
scikit-learn.preprocessing library.
As illustrated in Fig. 2, two distinct graphs are introduced in [21] to predict the current stress from the strain
history with different input structures. The “black box” graph relies only on the strain history, while the “informed”

LSTM LSTM

direct relation (black box) "informed" direct relation

(BB) (DG1)

Fig. 2: Directed and informed graph representing the information flow in the machine learning-based crystal plasticity
material model based on [21].

approach incorporates additional intermediate inputs, such as the previously predicted stresses. The input dimension
of the neural network for the “black box” approach is (6, lsequence ), while for the “informed” approach it is (12, lsequence ).
When implementing the UMAT subroutine, the solution-dependent state variables STATV(NSTATV) facilitate the
temporary retention of stress and strain history. This enables the historical information to be available for the current
calculation step, thereby reproducing the input of the neural networks. Within the current step n + 1, ABAQUS passes
the strain components from step n − 2 to step n − 2 − lsequence as the state variable, while the strain from the last
two steps can be directly computed from the previous strain ϵn and the strain increment ∆ϵn+1 , where lsequence is the
input’s length of interest.
Embedding an ANN-Based Crystal Plasticity Model in ABAQUS UMAT 5

2.2 Mathematical expression for forward and backward pass in a single LSTM memory cell

Hochreiter and Schmidhuber [23] first introduced long-short-term memory (LSTM) in 1997 as a variation of the
recurrent neural networks (RNNs). The primary objective of LSTM is to mitigate the vanishing gradient problem
commonly encountered in RNNs. This is achieved through the utilization of gates that dynamically control the mem-
ory. To elucidate the functioning of the LSTM layer, a visual representation of a single LSTM cell is provided in
Fig. 3. An LSTM layer comprises multiple such cells. Each cell receives three primary inputs: an external current
input xt from the sequence, the internal hidden state ht−1 , and the cell state Ct−1 from the preceding time step. Sub-
sequently, the cell updates the hidden state ht and cell state Ct . The final output of the memory cell is encapsulated by
the hidden state computed at the last step, which is denoted as hN for N steps in total. Additionally, within the LSTM
cell, there exist three fundamental gates. The input gate (it ) regulates the inflow of information into the memory, the
forget gate (ft ) controls which information is retained in the memory, and the output gate (ot ) determines the extent
to which the output is influenced by each state respectively.

σ σ σ

Fig. 3: Diagram illustrating the backward pass of the information flow within a single LSTM memory cell, with
labeled derivatives on each edge.

This work specifically employed the LSTM layer from the Tensorflow Keras package to construct neural net-
works. The mathematical expressions presented in this context align with Keras’ implementation and should remain
consistent if any other deep learning framework is chosen for the same purpose. For each LSTM cell within the frame-
work of Keras, we use two distinct sets of weight matrices: (1) the kernel weight matrix Wkernel , which is responsible
for weighting the input sequence, and (2) the recurrent weight matrix Wrecurrent , which is responsible for weighting
the preceding time step. These matrices consist of four vertical submatrices corresponding to the input gate, forget
gate, output gate, cell and cell input activation vectors, i.e.
n o
i f C o
Wkernel/recurrent = Wkernel/recurrent |Wkernel/recurrent |Wkernel/recurrent |Wkernel/recurrent

with one set of biases, i.e. b = bi |bf |bC |bo .
(t)
As a starting point of calculation, the forget gate fi for time step t and cell i is computed as introduced in [23]:
 
(t) (t) (t−1)
Wf Wf + bf  .
X X
f = σ
i x + kernel,i,j j h recurrent,i,j j (1)
i
j j

We also have the following expression to update the cell state:

 
(t ) X C (t)
X C (t−1)
C̃i =tanh  Wkernel,i,j xj + Wrecurrent,i,j hj + bC
i
, (2)
j j

(t ) (t ) (t−1) (t ) (t)
Ci =f i ∗ Ci + ii ∗ C̃i . (3)
6 Yuqing He1 et al.

The calculation of the output gate is

 
(t) (t) (t−1)
X o
X o
oi = σ  Wkernel ,i,j xj + Wrecurrent,i,j hj + boi  . (4)
j j

The update of the hidden state is  

(t) (t) (t )
hi = oi ∗ tanh Ci . (5)

2.3 Numerical differentiation

Numerical differentiation is a computational method used for calculating derivatives. Unlike the backpropagation
approach, numerical differentiation does not depend on the network’s architecture, making it more applicable to com-
plex network structures. The underlying concept of numerical differentiation is straightforward, wherein the original
derivative is substituted with the derivative of an interpolation polynomial, which is generated with the help of a Taylor
expansion [8]. In this work, the central difference method is used to determine the optimal discretization distance.
Considering the vectorial representation of the stress tensor σ and the strain tensor ϵ, the central difference method
delivers the partial derivative of the stress component σi (ϵ) with respect to the strain component ϵj as

σi ϵ + 21 ∆ϵj − σi ϵ − 12 ϵj ∆ϵ2j ∂σi3

∂σi (ϵ)
= − . (6)
∂ϵj ∆ϵj 24 ∂ 3 ϵj
| {z } | {z }
Explicit formula Remainder term

The error introduced by numerical differentiation comes from the truncation and rounding errors. Let the absolute
error, which indicates the resolution of approximation, be 10−9 , the optimal discretization size would be then ∆ϵj ≈
10−2 . Notice that the numerical differentiation method employed in this study involves using the scaled input and
scaled output data to calculate the derivative (see Section 2.1.2). Consequently, it is necessary to reverse-scale the
Jacobian matrix back to its original value range:
 
σi − σi,min
∂
∂σi,scaled (ϵ) σi,max − σi,min ϵ − ϵj,min ∂ (σi (ϵ) − σi,min ) ϵj,max − ϵj,min ∂σi (ϵ)
=   = j,max =
∂ϵj,scaled ϵj − ϵj,min σi,max − σi,min ∂ (ϵj − ϵj,min ) σi,max − σi,min ∂ϵj
∂
ϵj,max − ϵj,min
 
∂σi (ϵ) σi,max − σi,min ∂σi,scaled (ϵ)
⇒ = , (7)
∂ϵj ϵj,max − ϵj,min ∂ϵj,scaled

where σi,max/min and ϵj,max/min are the maximal and minimal constant values each of the stress and strain components.
Using the range of data values from the generated database, the error resulting from numerical differentiation is
estimated to be about 3.16 × 10−5 . Since the Jacobian matrix cannot be defined exactly in a numerical context, to
α
obtain an acceptable converged solution, the largest residual in the balance equation rmax from the tolerance definition
should not be greater than the error of the numerical differentiation.

3 Numerical examples

Two numerical examples are presented to examine the feasibility of the introduced approach: A two-dimensional plate
with a hole and a two-dimensional plate with a double notch. These benchmark problems are expected to demonstrate
the stress concentration effect for plasticity at the center of the plate, i.e. at the narrow areas, under monotonic tension
loading conditions. The geometry of the samples is illustrated in Fig. 4.
Both plates are completely fixed on the left edge and subjected to a monotonic displacement of 0.01 on the right
edge along the x-axis. Other degrees of freedom along the y - and z -axis are free. Both plates use a fine mesh with
a global element size of 1.0, composed of triangular plane-stress CPS3 elements and quadrilateral elements CPS4.
Automatic mesh refinement is implemented in the critical regions. The default element thickness is set to 1.0 . For
Embedding an ANN-Based Crystal Plasticity Model in ABAQUS UMAT 7

50
50 70°

20 7.5 15

R7
100 100

Fig. 4: Geometry of the two-dimensional initial-boundary-value problems (IBVPs) defined in ABAQUS v.2022.

the current modeling, it is important to notice that the subjected displacement limit should not exceed the maximum
value considered in the training data. This will help avoiding NN’s inherent low accuracy in the extrapolation (larger
errors for displacement limits out of the training range).
In the case of the plate with a hole, the suggested initial increment size is chosen as 0.0014. The total simulation
time is 92.88 seconds. In the case of the plate with a notch, the suggested initial increment size is chosen as 0.0012.
The simulation time is 96.23 seconds. To evaluate the validity of the proposed implementation, a comparison between
the ML-based approach and a conventional plasticity model is presented. In this, the simulation results based on the
ML model are compared with that of a simple isotropic hardening plasticity model provided in ABAQUS/Standard
and characterized by a monotonic stress-strain curve. The Young’s modulus and Poisson’s ratio of the material are
fitted from the generated database. The fitted Young’s modulus was rounded to 1500, and the fitted Poisson’s ratio
was 0.33 .

3.1 Simulation resutls

Figure 5 illustrates the nodal results of the von Mises stress after deformation for both cases of the data-based material
model and the fitted isotropic hardening plasticity model of ABAQUS/Standard. Results indicate that the LSTM model
effectively captures the stress concentration phenomenon around the hole and the notch. In the case of the plate with
a hole, the maximum von Mises stress was 0.6222 MPa, and the minimum was 0.0266 MPa. In the case of the plate
with a notch, the maximum von Mises stress was 0.4840 MPa, and the minimum was 0.0534 MPa.

3.2 Discussion of results and convergence

3.2.1 Impact of initial increment size

To attain accurate results, it is important to ensure convergence during simulation. It is observed during the simulation
process that the suggested initial value of increment size significantly influences the convergence performance for
different displacement loads. Figure 6 shows the correlation between the critical minimal increment size in displace-
ment and the overall displacement experienced within a single load step. These critical values indicate that, to achieve
converged solutions, the suggested initial increment size for simulation should exceed the critical initial increment
size. An initial increment size below the corresponding critical value leads to divergence and subsequent errors in the
simulation result.

3.2.2 Impact of mesh fineness

To investigate the impact of mesh fineness on the convergence behavior, simulations are conducted on the same
part geometry employing a coarser mesh. The simulation settings, including the initial displacement increment size,
remained unaltered. Figure 7 shows the von Mises stress distribution. The acquired results show that the mesh fineness
notably influences result precision, while its impact on the convergence process itself is comparatively minor.
8 Yuqing He1 et al.

Von Mises Stress

0.6
Fitted isotropic
hardening 0.5

0.4

0.2

Machine 0.1
learning
0.0

(a) Geometry (1): 2D plane-stress IBVP of a plate with a hole.

Von Mises Stress

0.5
Fitted isotropic
hardening 0.4

0.3

0.2

Machine 0.1
learning
0.0

(b) Geometry (1):2D plane-stress IBVP of a plate with a double notch.

Fig. 5: Two initial-boundary-value problems (IBVPs) for comparison of the von Mieses stress components (in MPa)
between the fitted isotropic hardening model and the ML-based material model in ABAQUS/Standard.

0.020 Plate with hole

Critical initial increment size

Plate with notch

0.015

0.010

0.005

0.000
0.00 0.02 0.04 0.06 0.08 0.10
Displacement load
Fig. 6: Critical minimal incremental size related to displacement load within a single load step.

0.0331 0.15 0.20 0.25 0.30 0.3504

(a) 2D plat with hole, coarse mesh.

0.0940 0.15 0.20 0.25 0.30 0.3477

(b) 2D plate with a notch, coarse mesh.

Fig. 7: Von Mises stress for models with coarse mesh. The initial displacement load increment size is identical to the
fine mesh model.
Embedding an ANN-Based Crystal Plasticity Model in ABAQUS UMAT 9

4 Summary and conclusions

This manuscript presents an approach to integrate machine learning-based material models for complex path-dependent
constitutive relationships, such as crystal plasticity, into the finite element simulation framework of ABAQUS UMAT.
The procedure particularly addresses solving nonlinear problems through the iterative Newton-Raphson method
by substituting the material model with a neural network-based structure. The proposed methodology reconstructs
offline-trained LSTM neural network calculation paths within ABAQUS, encapsulating path-dependent stress-strain
relations at individual material points. Thus, it assists in improving the accuracy of describing nonlinear material
responses. Two numerical 2D plane-stress IBVPs are conducted under uniaxial displacement loading conditions to
prove the quality and applicability of the proposed approach. It is found in this study that using an evenly distributed
training dataset across the whole value range can improve the accuracy of NN’s stress prediction. Moreover, using the
direct graph approach instead of the informed graph can lower the training errors. It is also worth mentioning that NN
prediction inherently contains bias and variance, which can lead to convergence challenges during analysis. In this
connection, it is found that a displacement load below the critical value results in divergence.
In conclusion, the integration of ML techniques offers an efficient and accurate method for modeling path-
dependent material behaviors, such as crystal plasticity. This efficiency is primarily achieved because the trained
ML model embedded within the UMAT behaves like an elasticity model during simulations. This eliminates the need
for explicit determination of the material parameters, which is often required in traditional material models. In con-
ventional models, material parameters such as elastic moduli, yield stress, and hardening rules must be determined
either experimentally or through fitting processes, which can be time-consuming and difficult to generalize across
different loading scenarios. However, with the ML approach, these parameters are implicitly learned from the data,
simplifying the simulation setup.
The accuracy of this approach is rooted in the fact that the ML model is trained on a database generated from
lower-scale simulations of crystal plasticity. By relying on lower-scale simulations that capture the detailed micro-
scopic interactions and deformation mechanisms, such as slip behaviors in crystals, the model can make predictions
that reflect the true material response with high precision. This ability to incorporate microstructural information al-
lows the ML model to overcome the limitations of traditional constitutive models, which are often based on simplified
assumptions and overlook the complexities of material behavior at lower scales.
Future research works will explore incorporating more complex material behavior in connection with ML-based
techniques into the FE framework. This could involve, for example, modeling hydraulic fracturing with a phase-field
approach [18] or simulating additive manufacturing processes [4].

Conflict of interest

The authors declare that there is no conflict of interest.

Data Availability

Exemplary source codes are made available as open-access for interested readers at [https://doi.org/10.25835/6n5uu50y]

Acknowledgements The second author, Y. Heider, would like to gratefully thank the German Research Foundation (DFG) for the support
of the project “Multi-field continuum modeling of two-fluid-filled porous media fracture augmented by microscale-based machine-learning
material laws”, grant number 458375627.

References

1. Abadi M, Agarwal A, Barham P, Brevdo E, Chen Z, Citro C, Corrado GS, Davis A, Dean J, Devin M, et al. (2016)
Tensorflow: Large-scale machine learning on heterogeneous distributed systems. arXiv preprint arXiv:160304467
2. Aldakheel F, Elsayed E, Tarek Z, Wriggers P (2023) Efficient multiscale modeling of heterogeneous materials
using deep neural networks. Comput Mech 72:155–171, DOI 10.1007/s00466-023-02324-9
3. Aldakheel F, Soyarslan C, Palanisamy H, Elsayed E (2023) Machine learning aided multiscale magnetostatics.
arXiv preprint -:1–18, DOI arXiv:2301.12782
10 Yuqing He1 et al.

4. Ali B, Heider Y, Markert B (2024) Predicting residual stresses in slm additive manufacturing using a phase-field
thermomechanical modeling framework. Computational Materials Science 231:112576, DOI https://doi.org/10.
1016/j.commatsci.2023.112576
5. Borja RI (2013) Plasticity Modeling & Computation. Springer Heidelberg New York Dordrecht London, DOI
10.1007/978-3-642-38547-6
6. Chaaban M, Heider Y, Sun W, Markert B (2023) A machine-learning supported multi-scale lbm-tpm model of
unsaturated, anisotropic, and deformable porous materials. Int J Numer Anal Methods Geomech pp 1–22, DOI
https://doi.org/10.1002/nag.3668
7. Chollet F, et al. (2015) Keras. https://github.com/fchollet/keras
8. Dahmen W, Reusken A (2008) Numerik für Ingenieure und Naturwissenschaftler. DOI 10.1007/
978-3-540-76493-9
9. Dhillon A, Verma G (2020) Convolutional neural network: a review of models, methodologies and applications
to object detection. Prog Artif Intell 9:85–112, DOI https://doi.org/10.1007/s13748-019-00203-0
10. Eidel B (2023) Deep cnns as universal predictors of elasticity tensors in homogenization. Computer Methods in
Applied Mechanics and Engineering 403:115741
11. Garcia-Teijeiro X, Rodriguez-Herrera A (2021) Combined Machine-Learning and Finite-Element Approach
for Multiscale 3D Stress Modeling. SPE Reservoir Evaluation & Engineering 24(04):827–846, DOI 10.2118/
205493-PA
12. Giovine P (2022) Continua with partially constrained microstructure. Continuum Mechanics and Thermodynam-
ics 34:273–295, DOI 10.1007/s00161-021-01057-5
13. Gorji MB, Mozaffar M, Heidenreich JN, Cao J, Mohr D (2020) On the potential of recurrent neural networks for
modeling path dependent plasticity. Journal of the Mechanics and Physics of Solids 143:103972, DOI 10.1016/j.
jmps.2020.103972
14. Gu J, Wang Z, Kuen J, Ma L, Shahroudy A, Shuai B, Liu T, Wang X, Wang G, Cai J, Chen T (2018) Recent ad-
vances in convolutional neural networks. Pattern Recognition 77:354–377, DOI https://doi.org/10.1016/j.patcog.
2017.10.013
15. Guan S, Qu T, Feng YT, Ma G, Zhou W (2023) A machine learning-based multi-scale computational framework
for granular materials. Acta Geotech 18:1699–1720, DOI https://doi.org/10.1007/s11440-022-01709-z
16. Hashash YMA, Jung S, Ghaboussi J (2004) Numerical implementation of a neural network based material model
in finite element analysis. International Journal for Numerical Methods in Engineering 59:989–1005, DOI 10.
1002/nme.905
17. Heider Y (2021) Multi-field and multi-scale computational fracture mechanics and machine-learning material
modeling. Habilitation, Report No. IAM-13 of Institute of General Mechanics, RWTH Aachen University
18. Heider Y (2021) A review on phase-field modeling of hydraulic fracturing. Engineering Fracture Mechanics
253:107881
19. Heider Y, Sun W (2023) Objectivity and accuracy enhancement within ann-based multiscale material modeling.
PAMM 22(1):e202200203, DOI https://doi.org/10.1002/pamm.202200203
20. Heider Y, Wang K, Sun W (2020) SO(3)-invariance of informed-graph-based deep neural network for anisotropic
elastoplastic materials. Computer Methods in Applied Mechanics and Engineering 363:112875, DOI 10.1016/j.
cma.2020.112875
21. Heider Y, Suh HS, Sun W (2021) An offline multi-scale unsaturated poromechanics model enabled by self-
designed/self-improved neural networks. Int J Numer Anal Methods Geomech 45(9):1212–1237, DOI https:
//doi.org/10.1002/nag.3196
22. Heider Y, Aldakheel F, Ehlers W (2024) Cnn-powered micro- to macro-scale flow modeling in deformable porous
media. Meccanica -:under review
23. Hochreiter S, Schmidhuber J (1997) LONG SHORT-TERM MEMORY. Neural Computation 9:1735–1780, DOI
10.1162/neco.1997.9.8.1735
24. Huang D, Fuhg JN, Weißenfels C, Wriggers P (2020) A machine learning based plasticity model using proper
orthogonal decomposition. Computer Methods in Applied Mechanics and Engineering 365:113008, DOI 10.
1016/j.cma.2020.113008
25. Jang DP, Fazily P, Yoon JW (2021) Machine learning-based constitutive model for j2- plasticity. International
Journal of Plasticity 138:102919, DOI https://doi.org/10.1016/j.ijplas.2020.102919
26. Koeppe A, Bamer F, Markert B (2020) An intelligent nonlinear meta element for elastoplastic continua: deep
learning using a new time-distributed residual u-net architecture. Computer Methods in Applied Mechanics and
Embedding an ANN-Based Crystal Plasticity Model in ABAQUS UMAT 11

Engineering 366:113088, DOI https://doi.org/10.1016/j.cma.2020.113088

27. Lazar A (2009) SORN: A Self-organizing Recurrent Neural Network. Frontiers in Computational Neuroscience
3, DOI 10.3389/neuro.10.023.2009
28. Lee YT, Wu TH, Lin ML, Ko CC (2021) Machine (Deep) Learning and Finite Element Modeling, Springer
International Publishing, Cham, pp 183–188. DOI 10.1007/978-3-030-71881-7 14
29. Lefik M, Schrefler B (2003) Artificial neural network as an incremental non-linear constitutive model for
a finite element code. Computer Methods in Applied Mechanics and Engineering 192:3265–3283, DOI
10.1016/S0045-7825(03)00350-5
30. Li X, Roth CC, Mohr D (2019) Machine-learning based temperature- and rate-dependent plasticity model: Ap-
plication to analysis of fracture experiments on DP steel. International Journal of Plasticity 118:320–344, DOI
10.1016/j.ijplas.2019.02.012
31. Lourenço R, Andrade-Campos A, Georgieva P (2022) The use of machine-learning techniques in material
constitutive modelling for metal forming processes. Metals 12(3), DOI 10.3390/met12030427, URL https:
//www.mdpi.com/2075-4701/12/3/427
32. Maia M, Rocha I, Kerfriden P, Van Der Meer F (2023) Physically recurrent neural networks for path-dependent
heterogeneous materials: Embedding constitutive models in a data-driven surrogate. Computer Methods in Ap-
plied Mechanics and Engineering 407:115934, DOI 10.1016/j.cma.2023.115934
33. Miyazawa Y, Briffod F, Shiraiwa T, Enoki M (2019) Prediction of Cyclic Stress–Strain Property of Steels by
Crystal Plasticity Simulations and Machine Learning. Materials 12:3668, DOI 10.3390/ma12223668
34. Oishi A, Yagawa G (2017) Computational mechanics enhanced by deep learning. Computer Methods in Applied
Mechanics and Engineering 327:327–351, DOI 10.1016/j.cma.2017.08.040
35. Pantidis P, Mobasher ME (2023) Integrated Finite Element Neural Network (I-FENN) for non-local continuum
damage mechanics. Computer Methods in Applied Mechanics and Engineering 404:115766, DOI 10.1016/j.cma.
2022.115766
36. Patra S, Dhar S, Acharyya S (2023) Finite element implementation of ann-based constitutive models for dsa in
ss304. J Inst Eng India Ser D pp –, DOI https://doi.org/10.1007/s40033-023-00475-w
37. Suh HS, Kweon C, Lester B, Kramer S, Sun W (2023) A publicly available pytorch-abaqus umat deep-
learning framework for level-set plasticity. Mechanics of Materials 184:104682, DOI https://doi.org/10.
1016/j.mechmat.2023.104682, URL https://www.sciencedirect.com/science/article/pii/
S016766362300128X
38. Tandale S, Stoffel M (2023) Recurrent and convolutional neural networks in structural dynamics: a modified
attention steered encoder–decoder architecture versus lstm versus gru versus tcn topologies to predict the response
of shock wave-loaded plates. Comput Mech DOI 10.1007/s00466-023-02317-8
39. Tao F, Liu X, Du H, Yu W (2021) Learning composite constitutive laws via coupling Abaqus and deep neural
network. Composite Structures 272:114137, DOI 10.1016/j.compstruct.2021.114137
40. Weber P, Wagner W, Freitag S (2023) Physically enhanced training for modeling rate-independent plasticity with
feedforward neural networks. Comput Mech 72:827–857, DOI https://doi.org/10.1007/s00466-023-02316-9
41. Xu K, Huang DZ, Darve E (2021) Learning Constitutive Relations using Symmetric Positive Definite Neural
Networks. Journal of Computational Physics 428:110072, DOI 10.1016/j.jcp.2020.110072
42. Yuan M, Paradiso S, Meredig B, Niezgoda SR (2018) Machine Learning–Based Reduce Order Crystal Plasticity
Modeling for ICME Applications. Integrating Materials and Manufacturing Innovation 7:214–230, DOI 10.1007/
s40192-018-0123-x