This paper presents a novel, physically informed machine learning (ML) framework for the accurate modeling and visualization of steel microstructure evolution during annealing. Utilizing a chained support vector regression (SVR) architecture with optimized hyperparameters, the model sequentially predicts key microstructural states, ensuring metallurgical consistency. The process begins by forecasting recrystallization fraction (RF) kinetics, which is critically constrained by the classical Johnson–Mehl–Avrami–Kolmogorov (JMAK) model. The resulting JMAK-corrected RF then serves as a fundamental input to subsequent SVR models, which forecast the average grain size (AGS) and, finally, essential image-based microstructural features (mean and standard deviation of pixel intensity). This chained approach inherently prioritizes physically sound outputs, avoiding the consistency issues of isolated predictions. A unique visualization methodology is introduced, which selects and maps the closest experimental inverse pole figure (IPF) maps to the predicted states. This robust, multi-stage framework establishes a powerful, data-driven tool for simulating complex material evolution, thus minimizing the need for extensive experimental operations in materials design and process optimization.
Shahab et al. (Thu,) studied this question.