LCAONet: Message-passing with Physically Optimized Atomic Basis Functions

A Model capable of handling various elemental species and substances isessential for discovering new materials in the vast phase and compound space.Message-passing neural networks (MPNNs) are promising as such models, in whichvarious vector operations model the atomic interaction with its neighbors.However, conventional MPNNs tend to overlook the importance of physicochemicalinformation for each node atom, relying solely on the geometric features of thematerial graph. We propose the new three-body MPNN architecture with amessage-passing layer that utilizes optimized basis functions based on theelectronic structure of the node elemental species. This enables conveying themessage that includes physical information and better represents theinteraction for each elemental species. Inspired by the LCAO (linearcombination of atomic orbitals) method, a classical method for calculating theorbital interactions, the linear combination of atomic basis constructed basedon the wave function of hydrogen-like atoms is used for the presentmessage-passing. Our model achieved higher prediction accuracy with smallerparameters than the state-of-the-art models on the challenging crystallinedataset containing many elemental species, including heavy elements. Ablationstudies have also shown that the proposed method is effective in improvingaccuracy. Our implementation is available online.