Predicting fatigue and failure in metals and composites: a machine learning enabled multiscale modeling perspective

This article provides a perspective on a novel parametrically upscaled constitutive and damage modeling (PUCM/PUCDM) platform for multiscale modeling of fatigue and failure in metallic and composite structures. It integrates physics and thermodynamics-based modeling with machine learning, integrated computational materials engineering (ICME), temporal acceleration, and uncertainty quantification in a unified platform development. The platform enables uncertainty-quantified efficient scale bridging with explicit representation of lower-scale descriptors (RAMPs) in higher-scale response functions. The PUCM/PUCDMs are constitutive relations that can be readily incorporated in UMAT-type user windows of commercial finite element software for easy translation to a broader community. This is a significant benefit, which can provide the user community with a pathway for highly efficient computations without compromising the fidelity of very complex mechanisms at multiple scales. The embedding of machine learning-generated functions of RAMPs in the constitutive parameters is a unique advantage. It naturally renders them amenable to location-specific material design by coupling with design algorithms. Additionally, through strategic integration with emerging deep learning methods, the platform creates virtual damage sensing digital twins that interface with actual surface sensor measurements for efficient prediction of full-field (both surface and subsurface) temporally evolving deformation and damage variables. This facilitates robust structural health monitoring capabilities.