Machine Learning-Assisted Modelling of Void Evolution and Ductile Fracture in Anisotropic Materials: A Critical Review

Abstract

Void evolution and ductile failure in anisotropic metals with tension-compression asymmetry are strongly influenced by the coupled effects of stress triaxiality T, Lode parameter L, material orientation, hardening, and void morphology. Three-dimensional unit-cell simulations can capture these interactions in considerable detail but are computationally expensive for extensive calibration, design-space exploration, and uncertainty quantification. This critical review examines four complementary roles of artificial intelligence: surrogate prediction of finite-element responses, inverse identification of constitutive and damage parameters, physics-informed constitutive learning, and microstructure-aware image- or graph-based modelling. A multi-fidelity workflow is proposed that integrates homogenized models, unit-cell calculations, statistically representative porous volume elements, and experiments, while incorporating constraints related to objectivity, material symmetry, thermodynamic dissipation, physically admissible porosity evolution, and uncertainty calibration. Current evidence most strongly supports the use of AI for repeated forward evaluation and inverse parameter identification. Reliable extension toward broader failure prediction requires validation on unseen loading trajectories and material domains, separate assessment of stress and damage-related variables, and an uncertainty-based abstention or fallback mechanism outside the model's training coverage.