Interpretable machine learning for microstructure-dependent models of fatigue indicator parameters

Fatigue indicator parameters (FIPs) are typically computed using crystal plasticity finite element modeling (CPFEM) and used to predict microscale crack initiation. While informative, computing FIPs in this manner can limit their application in engineering use cases due to the computational demand of CPFEM. To address this limitation, an interpretable machine learning approach is developed and used to model FIPs in additive manufactured IN625 single-phase microstructures. Genetic programming based symbolic regression is used to evolve inherently interpretable expressions of FIPs from microstructure features. Once developed, these FIP models act as an efficient surrogate for CPFEM and, due to their symbolic representation, can be readily combined with engineering workflows.