A Generic and Extensible Prediction Method for Martensitic Transformation Temperature Combining Thermodynamic Guided Data Mining and Deep Learning

Chenchong Wang*，Kaiyu Zhu，Yong Li，Chunguang Shen，Xiaolu Wei，Wei Xu

Northeastern University, Shenyang, 110819, China

ABSTRACT: Martensite transformation temperature is one of the critical microstructure information for various steels related with metastable austenite. However，the complex mechanism of martensite transformation greatly limits performance of different models. So, to build a generic model for martensite transformation temperature prediction with high accuracy and high extensibility, deep data mining guided by thermodynamic and deep learning strategy were combined in this research. The data mining guided by thermodynamic was used to improve the quality of the data set and enrich the information of the data set, and adopting deep learning to improve the model's prediction accuracy for multi-dimensional database. By deep data mining, a hierarchical database with three levels of information was established. It included composition as the first level, driving force (DF) and stacking fault energy (SFE) as the second level, and friction work (W_f) information from machine learning analysis as the third level. Then, convolutional neural network (CNN) model was used for the final prediction. The mean and maximum of evaluation results of all 100 DDM-CNN models were taken as the evaluation indices. The prediction results are shown in Fig. 1. It could be concluded that by digging the deep mechanism information and integrating the advantages of thermodynamic, traditional machine learning and deep learning models, the current model showed greatly improved extensibility both within and beyond the composition range of the original database. This research provides an inspiring guidance that, with the help of deep data mining from physical mechanism, deep learning could partially get rid of the limitation of data amount and become even more powerful than traditional machine learning for solving the complex problems with small sample database.

Figure 1. The mean absolute error of different model

Keywords: Martensite transformation; Deep learning; data mining