EAC-Net: Predicting Real-Space Charge Density via Equivariant Atomic Contributions.

Area of Science:

• Computational chemistry
• Materials science
• Quantum mechanics

Background:

• Charge density is fundamental to density functional theory (DFT).
• Deep learning approaches for charge density prediction aim to accelerate electronic-structure calculations.
• Current methods face trade-offs between physical priors and representational flexibility or expressiveness and efficiency.

Purpose of the Study:

• Introduce a novel deep learning framework, the Equivariant Atomic Contribution Network (EAC-Net).
• Bridge existing paradigms in charge density prediction.
• Develop an accurate, efficient, and physically grounded method for predicting charge density.

Main Methods:

• Decompose total charge density into symmetry-consistent, atom-centered contributions.
• Couple contributions to real space, avoiding direct grid or basis prediction.
• Utilize an equivariant network architecture.

Main Results:

• Achieve high accuracy with errors typically below 1% across the periodic table.
• Demonstrate strong generalization to diverse chemical environments.
• Produce physically meaningful atomic charges consistent with chemical intuition.

Conclusions:

• EAC-Net provides an accurate and efficient framework for charge density prediction.
• The method integrates physical priors for improved interpretability and consistency.
• This approach advances deep learning applications in electronic-structure calculations.