Deep Charge: A Deep Learning Model of Electron Density from One-Shot Density Functional Theory Calculation

Taoyuze Lv,Zhicheng Zhong,Yuhang Liang,Feng Li,Jun Huang,Rongkun Zheng

2023-09-26

Abstract:Electron charge density is a fundamental physical quantity, determining various properties of matter. In this study, we have proposed a deep-learning model for accurate charge density prediction. Our model naturally preserves physical symmetries and can be effectively trained from one-shot density functional theory calculation toward high accuracy. It captures detailed atomic environment information, ensuring accurate predictions of charge density across bulk, surface, molecules, and amorphous structures. This implementation exhibits excellent scalability and provides efficient analyses of material properties in large-scale condensed matter systems.

Materials Science,Computational Physics

What problem does this paper attempt to address?

This paper presents a deep learning model called Deep Charge for predicting electronic density from a density functional theory (DFT) calculation. Electronic density is the basis of material properties, especially in the field of materials science and condensed matter physics. Traditional DFT methods are computationally expensive, limiting their application in large-scale systems.

The Deep Charge model utilizes local atomic environment information and naturally preserves physical symmetry, enabling high-precision training with only one DFT calculation. It can accurately predict the electronic density of various structures such as crystals, surfaces, molecules, and amorphous structures. It has good scalability and is suitable for efficient analysis of large condensed matter systems.

The paper compares the Deep Charge model with other machine learning models for predicting electronic density, including methods based on basis functions and grid points. It points out that the Deep Charge model, by adopting the embedding framework of Deep Potential, can more effectively handle complex and large-scale systems with only a small amount of training data. The results show that Deep Charge performs excellently in the lattice, surface, and complex structures of materials such as silicon and aluminum, with small errors, indicating its ability to accurately capture environmental changes.

The paper concludes by discussing the performance and scalability of the model and proposing possible future directions for improvement, such as optimizing truncation functions and loss functions, as well as adopting concurrent learning strategies to improve efficiency. The Deep Charge model provides a new approach for rapid prediction of charge density and contributes to the research of large-scale condensed matter systems with first-principles accuracy.