Accuracy of Phonon Dispersion Calculations via Machine Learning Potentials

Recent advances of machine learning interatomic potentials (MLIPs) have improved both the accuracy and scalability of energy and force predictions in chemical systems for many practical applications, but accurate estimations of higher-order derivatives of the potential energy surface (PES) are needed to warrant replacing the expensive density functional perturbation theory (DFPT) step in thermal transport calculations with MLIPs. Here we present illustrations of some of the challenges in accurate predictions of phonon frequencies with MLIP while retaining its scalability to larger-scale systems. We also present some theoretical analysis of the nature of the challenges and review a few solution ideas and state-of-the-art models proposed in the recent literature.