Machine-learning-aided analysis of relationship between crystal defects and macroscopic mechanical properties of TWIP steel.

Establishing efficient methods to obtain quantitative data on crystal defect evolution is vital for understanding material properties. Dynamic Transmission Electron Microscopy (TEM) captures crystal defects in materials undergoing plastic deformation, generating vast datasets with high temporal and spatial resolution. However, manual analysis of these images is labor-intensive, and automated, unbiased analysis remains a challenge. In this study, we developed a U-net-based machine learning approach to analyze TEM videos of crystal defect evolution in Twinning-Induced Plasticity (TWIP) steels with different grain sizes. The method overcame challenges like field-of-view translation and nonuniform defect motion. This approach quantitatively measured defect evolution as a function of time and strain with the same temporal resolution as the original videos, detecting even minor changes with high accuracy. We use this technique to quantitatively reveal the switch of the dominant plastic deformation mechanism with grain size and the relaxation of elastic strain due to the rapid increase in stacking faults. Our results validate the use of U-net models for efficient semantic segmentation of TEM videos, enabling accurate quantitative analysis. This work advances TEM video analysis and provides new insights into the deformation mechanisms of materials.