Microvoiding and constitutive damage modeling with artificial neural networks

Abstract

Continuum models of porous media have revolutionized computational fracture mechanics for traditional ductile materials, but the inherent assumptions have limited generalizability to other target materials or loading conditions. Here, we adopt a series of artificial neural networks (ANNs) to predict both the microscopic voiding characteristics (void shape, porosity) and macroscopic stress–strain constitutive response of porous elasto-plastic materials under various deformation states. We train the ANNs on a dataset generated from finite element models of 3D representative volume elements (RVEs), each containing a discrete spherical void, subjected to combinations of loading states. Results show that the data-driven model is capable of interpolative predictions as well as some levels of extrapolative predictions across a wide range of initial porosities (0–20%) and loading states outside of the training dataset, even at high deformation strains which induce extensive material softening and void growth. Through transfer learning, we further demonstrate that the ANNs, originally trained on a specific porous material dataset, can be readily adapted to other porous materials with substantially different properties through a significantly reduced training dataset. We discuss the implications of this machine learning approach vis-à-vis the extensively-developed Gurson model for porous material damage and failure predictions.