An Ultra-Fast Machine-Learning Potentials to Investigate the Phonon-Dislocation Interaction of Lead Selenide

Dislocations, common crystallographic defects from synthesis or thermal/mechanical processes, influence phonon dynamics in crystals, affecting materials phenomena like plastic flow, internal friction, and thermal resistance. To understand the phonon-dislocation interaction of lead selenide (PbSe), it is crucial to employ a multi-scale strategy where quantum simulations provide data for atomistic simulations. This study concentrates on creating an Ultra-Fast Force Fields (UF3) potential designed to connect quantum and atomistic simulations for PbSe. To derive the machine-learning potentials, we fit effective 2- and 3-body terms using B-spline basis sets with regularized linear regression. To generate structurally diverse training data, we run high pressure relaxations, ab initio MD calculations at various temperatures, and relaxations of defect systems. Further, to provide data that is both structurally and compositionally diverse, we sample the relaxation trajectories of structures produced from a genetic algorithm structure search. We also determine appropriate weightings of the training data to get efficient and accurate potentials, while avoiding overfitting. Our findings suggests that, in the case of lattice parameter, system energy, defect formation energies, and elastic constants calculations, the UF3 force fields for PbSe crystal predicts reasonable results compared to DFT with speeds that are thousands of times faster.