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

Large scale atomistic simulations enable access to materials phenomena not easily probed experimentally or by quantum methods. Accurate, efficient interatomic potentials are essential yet challenging to develop for complex materials and phenomena. Machine learning potentials, in particular Deep Potential (DP), offer robust general-purpose models. This work introduces a methodology to specialise ML potentials for high-fidelity simulations where general models are insufficient. As a case study, a general DP is specialised for the mechanical response of titanium’s HCP and BCC phases while also capturing defect, thermodynamic, and structural properties. The resulting DP accurately reproduces structures, energies, elastic constants, and gamma-lines for Ti, as well as dislocation core structures, vacancy formation energies, phase transition temperatures, and thermal expansion, thereby enabling direct atomistic modelling of plasticity and fracture in Ti. The specialisation approach, DPspecX, is general and extensible to other systems and properties of interest.

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