Spatiotemporal prediction and characterization of microstructure evolution during in-situ heat treatment using a deep learning algorithm

Abstract

The austenite grain growth plays a significant role in improving the structures and properties of carbon steels. Modeling, characterizing, and predicting the temporal evolution of microstructures is crucial for understanding the relationship of processing-structure-property. Therefore, in this study, we designed a customized in-situ heating device that could be installed in the scanning electron microscope chamber to observe the temporal evolution process of the microstructures of 45# steel. We collected the in-situ experimental images during heating for downstream deep learning-based grain boundary extraction, spatiotemporal sequence characterization, and prediction. The proposed pipeline was fully verified in three experimental datasets under two distinct heating schemes in terms of (1) short-term and long-term prediction, (2) feeding different lengths of input sequence, and (3) transferring to other datasets. The proposed pipeline presented a high performance in quantitative evaluation metrics and qualitative comparisons between the predicted results and the ground truths in all three verification scenarios. These results are inspiring for researchers as they can preview the evolutional state in advance, thereby saving money, labor, and time with the assistance of deep learning-based predictive models. We believe that this pipeline opens opportunities for modeling, characterizing, and predicting evolutional state, which will aid in the analysis of processing-structure-property relationship, tailoring heat treatment, and improving material properties.