Guided analysis of fracture toughness and hydrogen-induced embrittlement crack growth rate in quenched-and-tempered steels using machine learning

Abstract

This study focuses on developing a machine learning (ML) model, specifically a Bayesian-optimized deep neural network, leveraging numerical simulation data for the prediction of fracture toughness (KIC) and crack growth rate (CGR) in Quenched-and-Tempered Steel (AISI 4140 alloy) under hydrogen embrittlement conditions. The proposed model demonstrated superior accuracy with Root Mean Squared Error (RMSE) values of 0.052 for KIC and 0.084 for CGR, surpassing conventional ML models including random forest (0.094 for KIC, 0.139 for CGR), gradient boosting (0.184 for KIC, 0.196 for CGR), support vector regression (0.142 for KIC, 0.110 for CGR), and decision tree regression (0.119 for KIC, 0.133 for CGR). The findings also revealed that input parameters such as temperature, hydrogen concentration, and hydrostatic stress carry higher weight factors compared to other parameters in predicting KIC. Similarly, the prediction of high CGR values, ranging from 5.9 × 10−5 to 9.5 × 10−5 m/s, was associated with the importance of hydrostatic stress, strain rate, and initial crack length in the prediction model. This underscores the model's ability to capture the intricate dependencies of output objectives on input features. Furthermore, a comprehensive case study, informed by the ML model, highlights the potential for tuning specific processing input parameters to manage hydrogen embrittlement effectively, contributing to a deeper understanding of its dynamics.