Abstract

Although DFT has long been the workhorse method of computational chemistry, materials science, and physics, all beasts of burden eventually need a break. Machine learning interatomic potentials (MLIPs) have recently begun to bridge the divide between chemistry-dependent classical interatomic potentials and DFT . Often covering nearly all the periodic table, the accuracy of an MLIP depends heavily on its training data. Extant MLIP -ready datasets, such as the Materials Project, tend to be noisy and favor basins on the potential energy surface. More recent efforts to purpose-build datasets for training MLI Ps have favored maximalist approaches, possibly leading to dataset duplication. In my talk, I’ll give a broad overview of the accuracy and computational considerations entering the construction of MLIP datasets, including relevant background on DFT . I’ll then discuss two recent dataset generation efforts I’ve contributed to, MatPES and MP-ALOE, which have sought to maximize dataset information density and accuracy. I’ll finish with an outlook to where these efforts lead.