Prediction of martensitic transformation start temperature of steel using thermodynamic model, empirical formulas, and machine learning models

Abstract

Three methods are used to predict the martensitic transformation start temperature (Ms ) of steel. Based on the database containing 832 compositions and corresponding Ms data, prediction models are built, modified, and trained. Firstly, Ms was re-calculated by establishing a thermodynamic model to link the martensitic transformation driving force (Gibbs free energy difference of martensite and austenite) with resistance (elastic strain energy, plastic strain energy, interface energy, and shearing energy). Secondly, the existing Ms data is cleaned and re-predicted using traditional empirical formulas within different composition application ranges. Thirdly, four different algorithms in machine learning including random forest, k nearest neighbor, linear regression, and decision tree are trained to predict 832 new Ms values. By comparing the Ms results re-predicted by the mentioned three methods with the original Ms values, the accuracy is evaluated to identify the optimal prediction model.