LCAONet: Message-passing with physically optimized atomic basis functions

Abstract

A Model capable of handling various elemental species and substances is essential for discovering new materials in the vast phase and compound space. Message-passing neural networks (MPNNs) are promising as such models, in which various vector operations model the atomic interaction with its neighbors. However, conventional MPNNs tend to overlook the importance of physicochemical information for each node atom, relying solely on the geometric features of the material graph. We propose the new three-body MPNN architecture with a message-passing layer that utilizes optimized basis functions based on the electronic structure of the node elemental species. This enables conveying the message that includes physical information and better represents the interaction for each elemental species. Inspired by the LCAO (linear combination of atomic orbitals) method, a classical method for calculating the orbital interactions, the linear combination of atomic basis constructed based on the wave function of hydrogen-like atoms is used for the present message-passing. Our model achieved higher prediction accuracy with smaller parameters than the state-of-the-art models on the challenging crystalline dataset containing many elemental species, including heavy elements. Ablation studies have also shown that the proposed method is effective in improving accuracy. Our implementation is available online.