Machine Learning Enabled Parametric Multiscale Modeling of Metals & Composites: From Fatigue Crack Nucleation to Damage Sensing

Professor Somnath Ghosh

Civil & Systems Engineering, Mechanical Engineering, and Materials Science & Engineering

Johns Hopkins University

The rapid surge of machine learning (ML) tools in developing efficient surrogate models for solving challenging problems has drawn significant attention from the Mechanics of Materials community. However, ML techniques rely on extensive training datasets and often lack physical interpretability. Also, exclusively data-driven models can result in ill-posed problems or non-physical solutions. Alternatively, the notion of ML-enhanced parametric upscaling has been introduced for multi-scale analysis of fatigue failure in metallics materials, damage and failure of unidirectional and woven composites, and damage sensing in multifunctional composites. The Parametrically Upscaled Constitutive Model (PUCM) for metallic materials like Ti alloys and the Parametrically Upscaled Continuum Damage Mechanics Model (PUCDM) for composites are thermodynamically-consistent constitutive models that bridge multiple spatial scales through the explicit representation of representative aggregated microstructural parameters (RAMPs), representing statistical distributions of morphological and crystallographic descriptors of the microstructure. ML tools, viz. genetic programming-based symbolic regression (GPSR) and artificial neural networks (ANN) are implemented for generating PUCM/PUCDM coefficients as functions of lower-scale RAMPs, using data sets of homogenized micromechanical response variables. For damage sensing in piezocomposite structures, the Parametrically Upscaled Coupled Constitutive Damage Model (PUCCDM) is developed coupling mechanical, damage, and electrical fields. An advanced machine learning model (ConvLSTM) based on the combination of a convolutional neural network and a recurrent neural network is developed to predict microstructural damage mechanisms from macroscopic electric signal and RAMPs. The computational tool chain outputs the highly efficient PUCM/PUCDM/PUCCDM, which are invaluable tools for multiscale analysis with implications in location-specific design.