Machine learning-accelerated calculation of diffusion coefficients in crystalline structures from molecular dynamics

Diffusion is one of the most fundamental physical processes governing materials phenomena. However, the calculation of diffusion coefficients is a representative example of atomistic simulations that require a substantial computational cost despite their conceptual simplicity. This poses a technical bottleneck in the investigation of diffusion kinetics based on atomistic simulations, particularly for studies based on ab initio and machine learning interatomic potentials, which require meaningfully larger computational cost compared to those based on traditional interatomic potentials. In the present work, we developed a novel machine learning model that predicts reliable diffusion coefficients from raw mean squared displacement data obtained over simulation times shorter than 200 ps, achieving accuracy within one order of magnitude. We demonstrate that the model generalizes well across diverse systems and diffusion scenarios beyond those in the training set. The present approach offers a substantial advancement in computational efficiency for diffusion studies, enabling the exploration of diffusion kinetics in a wide range of materials with substantially reduced computational cost and time.