Prediction of Ac<sub>3</sub> and Martensite Start Temperatures by a Data-driven Model Selection Approach

Abstract

Four different information criteria, which are widely used for model selection problems, are applied to reveal the explanatory variables for phase transformation temperatures of steels, austenitise temperature (Ac3) and martensite-start temperature (Ms). Using existing datasets for CCT diagram for various steels, the predictive equations for these critical temperatures are derived. A number of empirical equations have been proposed to enable efficient prediction of the the Ac3 and Ms temperatures of steels. However, the key parameters in those equations are usually chosen based on researchers’ trials and errors. In this study, the performance of the information criteria is evaluated first using a simulated dataset mimicking the characteristics of those for the Ac3 and the Ms temperatures. Then the criteria are applied to the experimental data obtained from two different sources. The key parameters are chosen for the Ac3 and Ms temperatures and the derived equations are found to be in better agreement with experimental data than the previous empirical equations. Thus, it was clarified that the methods can be applied to automatically discover the hidden mechanism from complex multi-dimensional datasets of steels’ chemical composition.