EAC-Net: Real-space charge density via equivariant atomic contributions

Abstract: Charge density is a fundamental quantity in quantum simulations, yet its accurate computation remains a major bottleneck. We present the Equivariant Atomic Contribution Network (EAC-Net), a deep learning framework for efficient and accurate charge density prediction. By introducing an atom-grid coupling mechanism, EAC-Net integrates the strengths of grid-based and basis-function-based models, achieving simultaneous improvements in accuracy and efficiency. We evaluated EAC-Net on a wide variety of systems, including amorphous solids, molecular liquids, surface structures, and metallic alloys, and found that it consistently achieves high accuracy with prediction errors typically below 1%. We further develop EAC-mp by training on Material Project's CHGCAR datasets, which achieves state-of-the-art accuracy comparable to existing large charge density models while providing atomic-decomposed charge densities. The model demonstrates strong zero-shot prediction capabilities across diverse material systems. Moreover, EAC-Net generalizes well beyond the training distribution, supporting downstream applications such as non-self-consistent band structure calculations under structural perturbations. By bridging local chemical environments and global charge distributions, EAC-Net provides a scalable and general framework for accelerating electronic structure prediction, with potential applications in high-throughput materials screening and machine-learning-driven simulation workflows.