Accelerating high-throughput phonon calculations via machine learning universal potentials

Phonons play a critical role in determining various material properties, but conventional methods for phonon calculations are computationally intensive, limiting their broad applicability. In this study, we present an approach to accelerate high-throughput harmonic phonon calculations using machine learning universal potentials (MLIPs) combined with an efficient training dataset generation strategy. Instead of computing phonon properties from a large number of supercells with small atomic displacements of a single atom, our approach uses a smaller subset of supercell structures where all atoms are randomly displaced by 0.01 to 0.05 UŮ, significantly reducing computational costs. We train a state-of-the-art MLIP based on multi-atomic cluster expansion (MACE), on a comprehensive dataset of 2738 materials with 77 elements, totaling 15,670 supercell structures, computed using high-fidelity density functional theory (DFT) calculations. The trained model is validated against phonon calculations for a held-out subset of 384 materials, achieving a mean absolute error (MAE) of 0.18 THz for vibrational frequencies from full phonon dispersions, 2.19 meV/atom for Helmholtz vibrational free energies at 300K, as well as a classification accuracy of 86.2% for dynamical stability of materials. A thermodynamic analysis of polymorphic stability in 126 systems demonstrates good agreement with DFT results at 300 K and 1000 K. In addition, the diverse and extensive high-quality DFT dataset curated in this study serves as a valuable resource for researchers to train and improve other machine learning interatomic potential models.