A Physics-Informed Data-Driven Discovery for Constitutive Modeling of Compressible, Nonlinear, History-Dependent Soft Materials under Multiaxial Cyclic Loading

TL;DR

This paper introduces a hybrid physics-informed machine learning framework combining Gaussian Process Regression and RNNs to model complex, history-dependent soft material behavior under multiaxial cyclic loading, ensuring physical validity and robustness.

Contribution

The work develops a novel physics-informed surrogate model integrating tensor-based response functions with machine learning, capturing nonlinear, rate-dependent viscoelastic behavior of soft materials.

Findings

Accurately predicts complex nonlinear viscoelastic responses.

Demonstrates robustness to synthetic noise and variable loading conditions.

Ensures thermodynamic consistency in predictions.

Abstract

We propose a general hybrid physics-informed machine learning framework for modeling nonlinear, history-dependent viscoelastic behavior under multiaxial cyclic loading. The approach is built on a generalized internal state variable-based visco-hyperelastic constitutive formulation, where stress is decomposed into volumetric, isochoric hyperelastic, and isochoric viscoelastic components. Gaussian Process Regression (GPR) models the equilibrium response, while Recurrent Neural Networks (RNNs) with Long Short-Term Memory (LSTM) units capture time-dependent viscoelastic effects. Physical constraints, including objectivity, material symmetry, and thermodynamic consistency, are enforced to ensure physically valid predictions. After developing the general form of the surrogate model based on tensor integrity bases and response functions, we employed the nonlinear Holzapfel differential viscoelastic model to generate training data. Two datasets, one for short-term and another for long-term relaxation, are constructed to span a wide range of material memory characteristics. The model is trained and tested under diverse multiaxial loading conditions, including different stretch levels applied independently in the longitudinal and transverse directions, varying strain rates, and both tension and compression states, even beyond the training domain. Energy dissipation is explicitly analyzed at different strain rates for both datasets to verify thermodynamic consistency through the second law. The results show that the proposed framework accurately captures complex, nonlinear, and rate-dependent material responses. Moreover, it demonstrates strong robustness to synthetic noise, enabling generalizable and physically consistent predictions under realistic and variable loading scenarios.