Martensite Start Temperature Prediction through a Deep Learning Strategy Using Both Microstructure Images and Composition Data

Abstract

In recent decades, various previous research has established empirical formulae or thermodynamic models for martensite start temperature (Ms) prediction. However, most of this research has mainly considered the effect of composition and ignored complex microstructural factors, such as morphology, that significantly affect Ms. The main limitation is that most microstructures cannot be digitized into numerical data. In order to solve this problem, a convolutional neural network model that can use both composition information and microstructure images as input was established for Ms prediction in a medium-Mn steel system in this research. Firstly, the database was established through experimenting. Then, the model was built and trained with the database. Finally, the performance of the model was systematically evaluated based on comparison with other, traditional AI models. It was proven that the new model provided in this research is more rational and accurate because it considers both composition and microstructural factors. In addition, because of the use of microstructure images for data augmentation, the deep learning had a low risk of overfitting. When the deep-learning strategy is used to deal with data that contains both numerical and image data types, obtaining the value matrix that contains interaction information of both numerical and image data through data preprocessing is probably a better approach than direct linking of the numerical data vector to the fully connected layer.