Extending the accuracy of embedded-atom method potentials for near-threshold collision dynamics using a multi-band formalism: A case study of bcc Fe

Accurate and computationally efficient interatomic potentials are essential for classical molecular dynamics simulations of radiation damage in nuclear materials. Although embedded-atom method (EAM) potentials are widely used because of their computational efficiency, most existing models fail to reproduce the directional dependence of threshold displacement energies (TDEs) observed experimentally and in first-principles molecular dynamics simulations of bcc Fe, a key model system for ferritic/martensitic steels. While machine-learning interatomic potentials (MLIPs) have substantially improved predictive accuracy, their high computational cost limits their broad applicability to large-scale cascade simulations and studies requiring extensive statistical sampling. In this work, we develop a multi-band EAM potential trained on density functional theory energies, forces, and stresses, aiming to bridge the accuracy–efficiency gap between conventional EAM potentials and MLIPs. The performance of the proposed model is evaluated through low-energy collision simulations near the TDE in bcc Fe. The resulting potential reproduces the directional dependence of TDEs with significantly improved accuracy compared with conventional EAM models. Detailed analyses reveal that kinetic energy dissipation during replacement-collision sequences and the spatial localization of defect structures play key roles in determining TDEs, particularly along the 〈111〉 direction. These mechanisms are captured more faithfully by the multi-band EAM model than by existing EAM potentials. These results demonstrate that the multi-band EAM potential provides an attractive balance between accuracy and computational efficiency, complementing conventional EAM potentials and computationally expensive MLIPs for radiation damage simulations.