Physics-informed neural networks for inverse length scale identification and forward phase-field fracture modeling

Abstract

This article presents a unified physics-informed neural networks (PINNs) framework for phase-field fracture modeling, capable of simultaneously identifying the length scale parameter and predicting crack evolution. The method embeds local observational data into a variational formulation by minimizing the total potential energy functional, enabling both inverse parameter identification and forward prediction without pre-assigned parameters. A discrete loading strategy with weight inheritance is employed to ensure the irreversibility of crack evolution. Numerical studies on single-edge cracked plates under tension and shear, together with a 3D compact-tension experiment, demonstrate substantial error reductions and superior performance compared to conventional PINNs-based phase-field model.