A domain-specific machine learning potential model for metallic materials spanning 53 elements

Alloys have been the cornerstone of human societal progress, from the Bronze Age to modern sustainable technologies. Yet, their atomic-scale behavior remains poorly understood, impeding their targeted optimization. Thus, reliable and efficient material design tools are urgently needed to accelerate alloy development. To address this demand, we develop a domain-specific machine learning potential (MLP) model spanning 53 metallic elements with balanced accuracy and efficiency. The model achieves DFT-level precision: energy mean absolute error (MAE) = 12 meV/atom, force MAE = 144 meV/Å, accurately predicts lattice parameters, elastic constants, and equation of states. We further validate its versatility through four alloy systems: (1) negative thermal expansion in Ti-Nb orthorhombic phases, (2) the Elinvar effect in Co₂₅Ni₂₅(TiZrHf)₅₀ intermetallic compound, (3) grain boundary segregation and high-temperature deformation in NbTaMoW multi-principal element alloy, and (4) precipitation pathway and θ′/Al interface segregation in Al-Cu-based alloy. This model provides a foundational tool for atomic-scale simulation, advancing materials research and accelerating alloy design.