A Neural-Network Framework to Learn History-Dependent Constitutive Laws and Identifiability of Internal Variables

TL;DR

This paper introduces a physics-consistent neural network framework for learning history-dependent constitutive laws, ensuring thermodynamic and stability properties, and demonstrates its effectiveness on polycrystalline magnesium data.

Contribution

It proposes a causal, energetic formulation for learning internal variables that are unique up to a linear transform, advancing constitutive modeling with neural networks.

Findings

Achieved 2% relative error in predicting Taylor-averaged response of magnesium.

Framework ensures thermodynamic consistency and stability in learned models.

Internal variables are identifiable up to a linear transformation.

Abstract

The identification of constitutive laws is ubiquitous in engineering: in modeling of materials where experimental data are fitted to mathematical models or learning surrogate models to beat the FE\textsuperscript{2} computational cost of multiscale numerical simulations. However, these models of constitutive laws, unless equipped with a potential formulation, are not necessarily consistent with (a) the second law of thermodynamics; (b) stability of the material under extreme applied strain; and (c) the mathematical theory underpinning the existence of solutions of the governing equation. In this work, we present a causal and energetic formulation, consistent with aforementioned properties, of learning a history-dependent constitutive law. This characterization of the class of internal variables sheds light on the equivalence class of equivalent surrogate models for the constitutive law. We show that the internal variables that are learned from the data are unique up to a linear transform. The framework is deployed to learn the Taylor-averaged response of a polycrystalline magnesium unit cell. We achieve 2\% relative error in the prediction of the Taylor-averaged response.