Prediction of Hardenability Curves for Non-Boron Steels via a Combined Machine Learning Model

Abstract

Hardenability is one of the most basic criteria influencing the formulation of the heat treatment process and steel selection. Therefore, it is of great engineering value to calculate the hardenability curves rapidly and accurately without resorting to any laborious and costly experiments. However, generating a high-precision computational model for steels with different hardenability remains a challenge. In this study, a combined machine learning (CML) model including k-nearest neighbor and random forest is established to predict the hardenability curves of non-boron steels solely on the basis of chemical compositions: (i) random forest is first applied to classify steel into low- and high-hardenability steel; (ii) k-nearest neighbor and random forest models are then developed to predict the hardenability of low- and high-hardenability steel. Model validation is carried out by calculating and comparing the hardenability curves of five steels using different models. The results reveal that the CML model works well for its distinguished prediction performance with precise classification accuracy (100%), high correlation coefficient (≥0.981), and low mean absolute errors (≤3.6 HRC) and root-mean-square errors (≤3.9 HRC); it performs better than JMatPro and empirical formulas including the ideal critical diameter method and modified nonlinear equation. Therefore, this study demonstrates that the CML model combining material informatics and data-driven machine learning can rapidly and efficiently predict the hardenability curves of non-boron steel, with high prediction accuracy and a wide application range. It can guide process design and machine part selection, reducing the cost of trial and error and accelerating the development of new materials.