A Universal Foundation Model for Electronic Structure Prediction

The convergence of big data and artificial intelligence is forging a new research paradigm in computational materials science. In this talk, I will mainly present our development of a universal foundation model for electronic structure prediction (see https://github.com/QuantumLab-ZY/HamGNN for all the source codes), which marks a significant advancement in this direction. Our work encompasses three core components:

First, we developed HamGNN, a Hamiltonian-centric graph neural network that is E(3)-equivariant. By rigorously preserving the inherent symmetries of physical systems, HamGNN accurately learns the electronic Hamiltonian matrices for both molecules and solids. It has demonstrated exceptional transferability and generalization capabilities across diverse systems, including allotropes of carbon and silicon, SiO₂ isomers, and BixSey compounds.

Second, building upon this foundation, we have trained for the first time a universal Hamiltonian model on a vast dataset of first-principles calculations for nearly all crystal structures in the Materials Project. This model can predict electronic structures for materials across the entire periodic table, successfully tackling highly complex systems such as multi-element compounds, solid-state electrolytes, moiré heterostructures, and metal-organic frameworks (MOFs).

Third, to precisely capture relativistic effects crucial for quantum materials, we introduced a novel strategy that decomposes the spin-orbit coupling (SOC) Hamiltonian into a spin-independent term and an SOC correction term. This approach enables the accurate modeling of SOC effects in complex and diverse systems, paving the way for the accelerated discovery of novel quantum materials.

Finally, these powerful models have been integrated into our proprietary software package, PASP (Property Analysis and Simulation Package for materials). Leveraging these developments, we have established the world's first AI-driven, large-scale electronic structure database and an open online prediction platform (https://sci-ai.cn), aiming to democratize and accelerate materials discovery and design.