Data generation and uncertainty quantification for universal graph deep learning interatomic potentials

One of the most exciting developments in recent years is the development of graph deep learning interatomic potentials such as the Materials 3-Body Graph Network (M3GNet), that have near-universal coverage across the entire periodic table. Such universal interatomic potentials (UIPs) have broad applications in the dynamic simulations and discovery of materials. However, current UIPs are still hampered by the lack of accurate potential energy surface (PES) data. In this talk, I will discuss these limitations and how advanced sampling approaches based M3GNet latent structural features can be used to generate high-quality PES data in an efficient manner. In addition, I will show that these approaches can also be used to quantify the uncertainty of a UIP for a structure. I will conclude with some observations on the relative performance of invariant versus equivariant UIP architectures, and a perspective on the challenges involved in UIPs with additional physics such as charges and magnetism.