DT-MIM: A network suitable for alloy microstructure evolution prediction and characterization

The evolution of microstructure is essential in determining the physical properties of materials. Accurate modeling and prediction of its dynamics are critical for unraveling the complex interrelationships between processing, structure, and performance. However, the inherent nonlinearity and non-stationarity of microstructural evolution, compounded by limited, variable, and uncontrollable experimental datasets, present major challenges to precise dynamic characterization. Here we address these challenges by focusing on critical deformation phenomena in Zr-4 alloys, including grain deformation, orientation transitions, and twin nucleation and growth. We introduce the Difference-aware Temporal Attention Memory In Memory (DT-MIM) network, an improved predictive recurrent neural network tailored to capture and represent the spatiotemporal dynamics of microstructural evolution. The model was trained and evaluated using data acquired with a custom scanning electron microscopy (SEM)-integrated system. The proposed DT-MIM model offers a novel framework for predicting microstructural evolution, employing a temporal attention mechanism on hidden-state differential information that effectively identifies highly dynamic and non-stationary transitions. Furthermore, the model demonstrates strong generalizability and can be extended to various alloy systems, providing a powerful new tool for material design and performance improvement.