Data-Driven Modeling of 1D Cyclic Plasticity: A Comparative Study of Recurrent and Feedforward Neural Networks

Constitutive modeling of materials under cyclic loading remains computationally demanding in solid mechanics. Classical combined isotropic--kinematic hardening models, such as Voce and Armstrong--Frederick formulations, require an iterative return-mapping algorithm at each integration point, which is costly in large-scale finite element (FE) simulations. Data-driven machine learning (ML) surrogates offer an alternative, but standard memoryless feedforward networks (multilayer perceptrons, MLPs) cannot represent path-dependent elastoplastic behavior unless internal state variables (e.g., backstress and accumulated plastic strain) are supplied as inputs; these variables are often unavailable in experiments. This work compares recurrent neural network architectures---long short-term memory (LSTM) networks and gated recurrent units (GRUs)---with feedforward MLPs for history-dependent cyclic plasticity. Synthetic datasets are generated with a custom Abaqus UMAT implementation of combined isotropic--kinematic hardening under non-monotonic, multi-amplitude strain paths. A sliding-window formulation lets the recurrent hidden states act as implicit surrogates for internal variables. Hyperparameter studies show that recurrent models reduce validation mean squared error (MSE) by more than an order of magnitude relative to the MLP baseline (from \(3.960 \times 10^{-3}\) to \(2.410 \times 10^{-4}\) for a two-layer LSTM with \(W=15\)). The recurrent models reproduce hysteresis loops and sharp reversal vertices without the path-dependent drift seen in the MLP. Among the tested configurations, the single-layer GRU (\(W=15\)) gives the lowest mean absolute error and reversal error, while the two-layer LSTM (\(W=15\)) gives the lowest validation MSE. The results provide practical guidance for using history-aware neural surrogates in nonlinear FE solvers.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-19
DOI
https://doi.org/10.21203/rs.3.rs-11626771/v1
Primary Topic
Machine Learning in Materials Science
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Data-Driven Modeling of 1D Cyclic Plasticity: A Comparative Study of Recurrent and Feedforward Neural Networks

Seyed Hadi Ghaderi, Seyed Mojtaba Varedi-Koulaei, Mohammad Ashraf-Ganjouei
Research Square
Machine Learning in Materials Science
preprint

Data-Driven Modeling of 1D Cyclic Plasticity: A Comparative Study of Recurrent and Feedforward Neural Networks

Seyed Hadi Ghaderi, Seyed Mojtaba Varedi-Koulaei, Mohammad Ashraf-Ganjouei
preprint en

Abstract

Constitutive modeling of materials under cyclic loading remains computationally demanding in solid mechanics. Classical combined isotropic--kinematic hardening models, such as Voce and Armstrong--Frederick formulations, require an iterative return-mapping algorithm at each integration point, which is costly in large-scale finite element (FE) simulations. Data-driven machine learning (ML) surrogates offer an alternative, but standard memoryless feedforward networks (multilayer perceptrons, MLPs) cannot represent path-dependent elastoplastic behavior unless internal state variables (e.g., backstress and accumulated plastic strain) are supplied as inputs; these variables are often unavailable in experiments. This work compares recurrent neural network architectures---long short-term memory (LSTM) networks and gated recurrent units (GRUs)---with feedforward MLPs for history-dependent cyclic plasticity. Synthetic datasets are generated with a custom Abaqus UMAT implementation of combined isotropic--kinematic hardening under non-monotonic, multi-amplitude strain paths. A sliding-window formulation lets the recurrent hidden states act as implicit surrogates for internal variables. Hyperparameter studies show that recurrent models reduce validation mean squared error (MSE) by more than an order of magnitude relative to the MLP baseline (from \(3.960 \times 10^{-3}\) to \(2.410 \times 10^{-4}\) for a two-layer LSTM with \(W=15\)). The recurrent models reproduce hysteresis loops and sharp reversal vertices without the path-dependent drift seen in the MLP. Among the tested configurations, the single-layer GRU (\(W=15\)) gives the lowest mean absolute error and reversal error, while the two-layer LSTM (\(W=15\)) gives the lowest validation MSE. The results provide practical guidance for using history-aware neural surrogates in nonlinear FE solvers.

Research Square
University of Shahrood (IR)
Machine Learning in Materials Science
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon