Speaker
Jaesuk Park
(The University of Texas at Austin)
Description
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. Here we explore the combination of MLIPs with state-of-the-art ab initio theory of thermal transport, which requires accurate estimations of higher-order derivatives of the potential energy surface (PES) to reliably replace the expensive density functional perturbation theory (DFPT) step. We present examples of different variables in MLIP training workflow and suggest the extent to which they impact the overall model performance for phonon frequency predictions.
Primary authors
Jaesuk Park
(The University of Texas at Austin)
Prof.
Feliciano Giustino
(The University of Texas at Austin)
