Dec 5, 2024
2:15pm - 2:30pm
Hynes, Level 2, Room 210
Chikashi Shinagawa1,So Takamoto1,Daiki Shintani1,Katsuhiko Nishimra1,Ju Li2
Preferred Networks, Inc.1,Massachusetts Institute of Technology2
Chikashi Shinagawa1,So Takamoto1,Daiki Shintani1,Katsuhiko Nishimra1,Ju Li2
Preferred Networks, Inc.1,Massachusetts Institute of Technology2
Accurate, universal, and fast interatomic potentials accelerate materials discovery. Neural network potentials (NNP) have initially achieved success in small molecule systems composed of few elements, and have since expanded to systems involving more elements and more diverse systems. Today, some NNPs cover most of the elements in the periodic table.<br/>Most of these NNPs are trained to reproduce potential energy surface (PES) of density functional theory (DFT) calculation with generalized gradient approximation (GGA) level exchange-correlation functional, such as Perdew–Burke–Ernzerhof (PBE) functional. Improvements in architectures and datasets have reduced errors between GGA-level DFT and NNPs, and the GGA error itself has become more pronounced.<br/>As a meta-GGA level correlation functional, which has more rich expressive power than GGA level functionals, r2SCAN exchange-correlation functional has recently been adopted in DFT calculations, showing improvements in properties like formation energies and densities compared to GGA. In the field of NNPs, some recent studies demonstrate simulations with higher accuracy to real-world properties by using NNPs pre-trained on GGA-PBE and re-trained with metaGGA or higher level data specific to the target materials and structures. However, for efficacy in materials discovery, it is essential to use a universal potential, which can accurately evaluate PES and other properties of a wide range of unknown structures.<br/>We have been developing PFP, which is an NNP aimed at materials discovery, characterized by its universality in reproducing diverse systems, including molecules, crystals, surfaces, and interfaces. We have continually improved both the universality and accuracy of PFP, which now supports most elements in the periodic table. Currently, PFP is trained with our inhouse GGA-PBE level dataset, and its accuracy to real-world properties is limited by the GGA level. Our goal is to refine PFP to reproduce r2SCAN level PES, achieving higher accuracy beyond the GGA level while maintaining its universality. In this presentation, we will show our progress towards r2SCAN level universal NNP and provide several examples of the improvements made possible through these advancements.