Machine learning local potentials for accelerated electron-phonon interactions calculations within the projector augmented-wave framework

Abstract

Electron-phonon interactions govern carrier dynamics, transport, and many optical and quantum phenomena in solids, but finite-displacement calculations within the projector augmented-wave (PAW) framework require up to 6N self-consistent supercell calculations for an N-atom system, making them costly for large, low-symmetry, and disordered materials. We introduce MLLocP, a machine learning strategy that learns the self-consistent local potential on real-space grids and supplies its displacement derivatives to the evaluation of all-electron PAW electron-phonon matrix elements. Analysis of the PAW decomposition identifies the local-potential contribution as a natural target for machine learning, while the remaining PAW quantities are retained within the established VASP/PHELEL framework. Using targeted grid-point sampling and feature-diversity selection for efficient model training, we validate MLLocP for elemental Cu, polar BAs, and chemically disordered Cu3Au. The learned potentials and their resulting displacement derivatives closely reproduce direct density functional theory calculations, yielding transport coefficients within approximately 3.5% for BAs mobility, 8% for Cu conductivity, and 11% for Cu3Au conductivity. For the 32-atom Cu3Au special quasirandom structure, using 40 sampled configurations reduces the number of self-consistent calculations by about fivefold relative to the 192 displaced structures required directly; using 20 or 10 configurations increases the acceleration to approximately 10- and 20-fold, with conductivity errors of 14.3% and 18.6%, respectively. MLLocP thus provides a scalable route to PAW electron-phonon calculations in complex materials while preserving the underlying all-electron formalism.