Specialising neural network potentials for accurate properties and application to the mechanical response of titanium

Accurate interatomic potentials are crucial for large-scale atomistic simulations of materials. This paper presents a methodology for specializing machine learning potentials, specifically the Deep Potential (DP) method, to achieve high fidelity simulations of complex phenomena. The authors specialize a general-purpose DP to model the mechanical response of titanium's HCP and BCC allotropes. The resulting potential accurately predicts various properties, including structures, energies, elastic constants, γ-lines, dislocation core structures, vacancy formation energies, phase transition temperatures, and thermal expansion, enabling direct atomistic modeling of titanium's plastic and fracture behavior. The specialization approach, DPspecX, is general and applicable to other materials and properties.

Tongqi Wen, Rui Wang, Lingyu Zhu, Linfeng Zhang, Han Wang, David J. Srolovitz, Zhaoxuan Wu