Physics-Informed Attention Mechanism and Generalization Capability of Deep Learning-Based Grain Growth Evolution Prediction

This study introduces a boundary-masked attention mechanism for grain growth prediction, enhancing the Out-Of-Distribution (OOD) generalization of machine learning models previously trained on idealized synthetic data. The proposed model significantly improved performance in various test cases, notably achieving a Structural Similarity Index Measure (SSIM) increase from 0.6221 to 0.7609 and reducing mean grain size error from 8.75% to 3.57% for bimodal grain size distributions. These findings underscore the potential of physics-informed architectures to enhance model robustness and accuracy in real-world applications, allowing for effective generalization without the need for retraining.