Machine Learning Approaches to Predicting and Understanding the Mechanical Properties of Steels

The mechanical properties are essential for structural materials. A dataset consists of 102 carbon steel and 258 low-alloy steel was built from the NIMS steel fatigue dataset. And five machine learning algorithms were applied on the dataset to predict the mechanical properties, including fatigue strength, tensile strength, fracture strength, and hardness of steels. Random forest shows a tremendous predictive power of those properties of steels. The prominent features of those properties were selected based on random forest, and symbolic regressions, the tempering temperature and the alloying elements of carbon, chromium, and molybdenum were shown to play essential roles on the mechanical properties. Besides, the mathematic expressions were generated via symbolic regression, and the expression can directly predict the mechanical properties with heat treatment conditions and compositions. The symbolic regression algorithm shows excellent potential in feature selection and expression discovery in materials science