Predicting the martensite start temperature of steels via a combination of deep learning and multi-scale data mining

Abstract

The martensite start ($M_s$) temperature, plays a significant role in guiding the alloy design and heat treatment process for steels. However, due to the complexity of martensite transformation, it remains challenging to establish more generalized models. In this study, the database with three scales was first built by multi-scale data mining, without relying on thermodynamic software. Then a convolutional neural network (CNN) model, as well as four traditional machine learning (ML) models, were trained to predict $M_s$ using multi-scale database. The CNN model exhibits the smallest error, and the five models all perform better than those trained solely by alloy composition, demonstrating the benefits of feature diversity. The benchmarking test indicates that the CNN model has higher accuracy, compared to the empirical equations, JMatPro software, and thermodynamic model. Besides, the simplified CNN models trained with remaining features after each stage of the ‘three-stage feature screening’ all show an error of only about 1 K from the original CNN model, illustrating the effectiveness of the current feature screening strategy and good robustness of the CNN model. Moreover, the CNN model can be utilized to predict the $M_s$ of the alloys with unknown compositional combinations, then new insights about the impacts of alloy elements on austenite stability and alloy design can be revealed. The integration of multi-scale data mining into a deep learning framework represented by CNN, offers a recipe for predicting certain attributes that are involved in complicated relationships with alloy composition.