Title: Machine Learning of Mechanical Properties of Steels

Abstract: The mechanical properties are essential for structural materials. The analyzed 360 data on four mechanical properties of steels, viz. fatigue strength, tensile strength, fracture strength, and hardness, are selected from the NIMS database, including carbon steels, and low-alloy steels. Five machine learning algorithms were applied on the 360 data to predict the mechanical properties and random forest regression illustrates the best performance. The feature selection was conducted by random forest and symbolic regressions, leading to the four most important features of tempering temperature, and alloying elements of carbon, chromium, and molybdenum to the mechanical properties of steels. Besides, mathematic expressions were generated via symbolic regression, and the expressions explicitly predict how each of the four mechanical properties varies quantitatively with the four most important features. The present work demonstrates the great potential of symbolic regression in the discovery of novel advanced materials.