Machine learning guided methods in building chemical composition-hardenability model for wear-resistant steel

Abstract

Six machine learning methods, including artificial neural network, gradient boosting regression, random forest, etc. were used to conduct a comparative study of building chemical composition-hardenability model for wear resistant steel. The results indicated that artificial neural network method with 32 × 32 × 32 structure had the highest prediction accuracy among the six machine learning methods based on our study. Through Pearson’ s linear correlation heat map and the feature importance parameter in the gradient boosting regression method, the contributions of different alloying elements on hardenability could be predicted, which guided us to design the further chemical composition. Finally, a reverse microalloying design based on the target performance was carried out with the artificial neural network model and an end quenching experiment using actual steel was used to evaluate the performance of model. The predicted results, calculated results and experimental results were consistent. The combination of material data base and machine learning provided an efficient approach to design the chemical composition of steels.