Predicting mechanical properties of bainitic rail steels based on microstructure digitalization and machine learning

It is usually quite challenging to establish a quantitative chemical composition-microstructure-mechanical property relationship for bainitic steels. In this study, digital information was extracted based on EBSD data of bainitic rail steels, and two machine learning models were built to predict strength and impact energy based on chemical compositions and digitalized microstructural features. The models showed excellent prediction accuracy (R 2 > 84%) for yield strength, tensile strength and impact energy. All predicted values were within the error range of experimental measured values. Feature importance analysis suggested that C has a beneficial and detrimental effect on the strength and toughness, respectively; while a high block boundary density proved to have a positive effect on toughness, which agrees well with previous experimental observations. The obtained quantitative relationship between chemical composition, microstructure and mechanical properties can serve as a good guideline for the design of bainitic rail steels with an optimized combination of high strength and toughness.