Predicting multiple fatigue properties of twinning-induced plasticity steels by black-box and white-box machine learning

Abstract

Predicting multiple fatigue properties of metals under a wide range of conditions is still a challenge, as massive high-dimension inputs and multiple outputs are involved. We systematically conduct fatigue experiments on TWIP steel under various conditions, including different preloading methods, temperatures, strain amplitudes and mean strains. Using experimental data, we propose both black-box and white-box machine learning models to predict the fatigue performance of TWIP steel. The black-box model employs dimensionality reduction, clustering and regression techniques to achieve simultaneous predictions for fatigue life and maximum stress amplitude. The predicted fatigue lives are 100% within 3✕ error band and 88.31% within 2✕ error band. The predicted maximum stress amplitudes are all within 1.51✕ error band. The white-box model utilizes symbolic regression and matching analysis to automatically discover several predictive formulas for fatigue life and maximum stress amplitude, without any predefined equations. The three optimal fatigue life prediction formulas yield 100% predicted values within 3✕ error band and 98% within 2✕ error band. The two optimal maximum stress amplitude prediction formulas yield predicted values all within 1.09✕ error band. Based on the results, we discuss the applicability of our models and propose suggestions for future developments in machine learning fatigue performance predictions.