Prediction and characterization of microstructure evolution based on deep learning method and in-situ scanning electron microscope.

Microstructure significantly affects materials' physical properties. Predicting and characterizing temporal microstructural evolution is valuable and helpful for understanding the processing-structure-property relationship but is rarely conducted on experimental data for its scarcity, unevenness, and uncontrollability. As such, a self-designed in-situ tensile system in conjunction with a scanning electron microscope was adopted to observe the grain evolution during the tensile process. We then used a deep learning-based model to capture grain growth behavior from the experimental data and characterize grain boundary and orientation evolution. We validated the framework's effectiveness by comparing the predictions and ground truths from quantitative and qualitative perspectives, using data from (1) a tensile experimental dataset and (2) a phase-field simulation dataset. Based on the two datasets, the model's predicted results showed good agreement with ground truths in the short term, and local differences emerged in the long term. This pipeline opened an opportunity for the characterization of microstructure evolution and could be easily extended to other scenarios, such as dendrite growth and martensite transformation. [Display omitted] • It's the first study using a deep learning method to predict and characterize temporal grain evolution on experimental data. • We accurately predict future grain evolution without any subjective assumptions and hand-crafted features. • We can provide valuable reference and guidance for future experiments based on the model's performance. • This pipeline opens a new way of shortening experimental time and characterizing temporal microstructure evolution. [ABSTRACT FROM AUTHOR]