Exploring time-series transformers for spatio-temporal prediction of microstructural evolution of polycrystalline grain

Abstract

This study introduces a novel microstructure prediction model (polycrystalline grain transformers: PGTs) for metal solidification processes, leveraging the capabilities of transformer networks. The model’s primary purpose is to predict the dynamic evolution of polycrystalline microstructures within additive manufacturing conditions. To facilitate the training of the PGTs model, we generated a comprehensive dataset through a two-dimensional multi-phase field numerical simulation model. The inherent strength of transformer networks lie in its multi-head attention mechanism, which enables the PGTs model to autonomously identify crucial feature information within the input sequence, particularly related to structural evolution. Our research outcomes underscore the robustness of the PGTs model, as it closely aligns with predictions obtained from conventional phase-field calculations. Importantly, this model exhibits a remarkable improvement in computational efficiency, surpassing traditional phase-field methods by several orders of magnitude. This breakthrough opens up new avenues for research, offering an efficient and cost-effective strategy for delving deep into the intricate relationship between solidification processes and material mechanical properties. Notably, we mark this study as the pioneering application of transformer networks in numerical simulations of the metal solidification process, showcasing its immense potential across various domains, including material design.