Efficient calculations of impurity diffusivity in metals by linearized multi-band embedded atom method potentials

Impurity diffusivity in metals has attracted much attention in materials science and engineering due to its significant relevance to the performance and integrity of materials and components. In computational simulations of impurity diffusivity, although molecular dynamics (MD) simulations using empirical potentials such as embedded atom method potentials (EAMs) have been widely used, the accuracy is often insufficient, especially when it comes to details such as isotope effects. On the other hand, while machine learning (ML) potentials trained on first-principles calculations can achieve high accuracy, their computational speed poses a challenge for statistically precise calculations of diffusion coefficients. In this study, we propose an extended version of EAM, called linearized multi-band EAM (LMB-EAM), which can be constructed by a force-matching method using first-principles calculation data as a training set with regularization, and validate its performance for impurity diffusivity in metals. For two test cases, H diffusion in bcc-W and O diffusion in liquid Na, we demonstrate that LMB-EAMs can be constructed with a small number of training data, outperform empirical potentials, and determine the diffusion coefficient including the isotope effect with reasonable accuracy, better than tested empirical potentials and slightly inferior to a tested ML potential. We also investigate which properties are important for accurate simulations of impurity diffusivity in solid metals, providing guidance for the construction of potential models.