Transformer-based structural identification in molecular dynamics simulations

Large-scale molecular dynamics (MD) simulations generate enormous amounts of atomic data, increasing the need for robust and efficient structural identification. Conventional methods generally classify individual atoms based on their local atomic environments and can be sensitive to thermal fluctuations. In this study, we develop a Transformer-based framework that extends structural identification from atom-wise classification to phase identification at arbitrary positions in material space. The proposed method employs the self-attention mechanism to evaluate surrounding atomic environments and their correlations. Using face-centered cubic (FCC) and liquid configurations of Ni, the model incorporating atomic environments up to the fifth-nearest-neighbor range maintained high classification accuracy for both high-temperature FCC and undercooled liquid configurations, demonstrating robustness against thermal fluctuations. Application to a solid–liquid coexistence system showed that the method captured the overall interface and provided a smoother spatial representation than conventional atom-centered identification. The proposed framework enables discrete atomic configurations to be mapped onto a spatial distribution of phase states, providing a non-atom-centric approach to structural analysis and a potential bridge between atomistic and continuum descriptions.