A micromechanics-based machine learning model for evaluating the microstructure-dependent rolling contact fatigue performance of a martensitic steel

Abstract

We designed the prediction system for the rolling contact fatigue (RCF) performance of martensitic steel based on the machine learning (ML) algorithm and micro-mechanical simulations. The emphasis here is not only to propose a low-cost micromechanics-based ML prediction system, but also to establish the linkage between microstructure features and RCF responses. In this prediction system, the hierarchical microstructure modeling approach and size-dependent physics-based crystal plasticity (CP) framework were developed to capture the micromechanical response of the martensitic steel in RCF. The hierarchical microstructure geometry, crystallographic orientation, and the corresponding fatigue indicator parameter (FIP) obtained from CP simulations were selected for database construction. The data filtering method was adopted by considering the loading history of RCF. The random forest (RF) algorithm was trained using the CP simulations to link the microstructure features to RCF performance. After that, the confidence and limitation of the prediction system were explained and analyzed in detail. It is demonstrated that the prediction system is capable of precisely predicting the RCF performance and its relation to microstructure features in a physically sound way. Results give data-driven insights into the microstructure design for the desired mechanical and RCF performance of martensitic steel.