深度学习电子结构计算简介

陈语诺; 段文晖; 徐勇

doi:10.7693/wl20260802

物理

深度学习电子结构计算简介

, ,

陈语诺, 段文晖, 徐勇. 2026: 深度学习电子结构计算简介, 物理, 55(8): 526-532. doi: 10.7693/wl20260802

引用本文:

陈语诺, 段文晖, 徐勇. 2026: 深度学习电子结构计算简介, 物理, 55(8): 526-532. doi: 10.7693/wl20260802

CHEN Yu-Nuo, DUAN Wen-Hui, XU Yong. 2026: An introduction to deep learning electronic structure calculations, Physics, 55(8): 526-532. doi: 10.7693/wl20260802

Citation:

CHEN Yu-Nuo, DUAN Wen-Hui, XU Yong. 2026: An introduction to deep learning electronic structure calculations, Physics, 55(8): 526-532. doi: 10.7693/wl20260802

深度学习电子结构计算简介

清华大学物理系北京 100084

通讯作者: 徐勇,email:yongxu@tsinghua.edu.cn

基金项目:
教育部学科突破计划先导项目(批准号：JYB2025XDXM408)、国家重点研发计划(批准号：2024YFA1409100)、国家自然科学基金(批准号：52388201；12334003；12421004；12361141826)、上海科技大学“AI4S Initiative”(批准号：SHTAI4S202504)资助项目，北京市量子人工智能重点实验室支持项目

An introduction to deep learning electronic structure calculations

Department of Physics, Tsinghua University, Beijing 100084, China

Corresponding author: XU Yong, yongxu@tsinghua.edu.cn

摘要: 第一性原理电子结构计算可精准预测材料的物理和化学性质，已成为物理学、材料科学和化学的重要研究手段。然而，传统第一性原理方法计算成本较高，难以满足大尺度、高通量的材料计算需求。深度学习技术的发展为电子结构计算开辟了新路径。文章将介绍深度学习电子结构计算的发展背景和基本思想，以深度学习哈密顿量方法为例，重点讨论如何基于重要物理先验设计出可解释性的神经网络模型，实现电子结构的精准预测。此外，还将介绍电子结构的通用材料模型、无监督学习等领域前沿的最新进展。相关研究表明，人工智能正在推动第一性原理计算从传统计算范式向数据与物理共同驱动的新范式演进，为未来的科学计算和材料发现开辟了新途径。
- 深度学习 /
- 电子结构计算 /
- 第一性原理方法 /
- 材料发现
Abstract: First-principles electronic structure calculations can accurately predict the physical and chemical properties of materials, and have become important research tools in physics, materials science, and chemistry. However, conventional first-principles methods are computationally demanding, which limits their applications in large-scale and high-throughput materials research. Recent advances in deep learning have created new opportunities for electronic structure calculations. This article introduces the background and fundamental concepts of deep learning applied to electronic structure analysis. Taking the deep learning Hamiltonian approach as a representative example, we focus on how interpretable neural-network models can be designed by incorporating important physical priors to achieve accurate predictions of electronic structures. We also review recent progress in emerging directions such as universal materials models for electronic structures and unsupervised learning. These developments suggest that artificial intelligence is driving first-principles calculations from conventional computational approaches toward a new paradigm jointly driven by data and physics, opening new avenues for future scientific computing and materials discovery.
- deep learning /
- electronic structure calculations /
- first-principles methods /
- materials discovery .

LeCun Y，Bengio Y，Hinton G. Nature，2015，521：436

Li H，Wang Z，Zou N et al. Nat. Comput. Sci.，2022，2：367

Tang Z，Chen H，Li Y et al. Nat. Comput. Sci.，2025，1：14.

Martin R M. Electronic Structure：Basic Theory and Practical Methods. Cambridge University Press，2020

Wang Y，Li Y，Tang Z et al. Sci. Bull.，2024，69：30

Li H，Tang Z，Gong X et al. Nat. Comput. Sci.，2023，3：321

Gong X，Li H，Zou N et al. Nat. Commun.，2023，14：2848

Li Y，Tang Z，Chen Z et al. Phys. Rev. Lett.，2024，133：076401

Li H，Tang Z，Fu J et al. Phys. Rev. Lett.，2024，132：096401

Ihrig A C，Wieferink J，Zhang I Y et al. New Journal of Physics， 2015，17：093020

Tang Z，Li H，Lin P et al. Nat. Commun.，2024，15：8815

计量

文章访问数: 138
HTML全文浏览数: 138
PDF下载数: 5
施引文献: 0

深度学习电子结构计算简介

通讯作者: 徐勇,email:yongxu@tsinghua.edu.cn

清华大学物理系北京 100084

收稿日期: 2026-07-10

基金项目:

关键词:

摘要: 第一性原理电子结构计算可精准预测材料的物理和化学性质，已成为物理学、材料科学和化学的重要研究手段。然而，传统第一性原理方法计算成本较高，难以满足大尺度、高通量的材料计算需求。深度学习技术的发展为电子结构计算开辟了新路径。文章将介绍深度学习电子结构计算的发展背景和基本思想，以深度学习哈密顿量方法为例，重点讨论如何基于重要物理先验设计出可解释性的神经网络模型，实现电子结构的精准预测。此外，还将介绍电子结构的通用材料模型、无监督学习等领域前沿的最新进展。相关研究表明，人工智能正在推动第一性原理计算从传统计算范式向数据与物理共同驱动的新范式演进，为未来的科学计算和材料发现开辟了新途径。

English Abstract

An introduction to deep learning electronic structure calculations

Corresponding author: XU Yong, yongxu@tsinghua.edu.cn

清华大学物理系北京 100084

Received Date: 2026-07-10

Keywords:

Abstract: First-principles electronic structure calculations can accurately predict the physical and chemical properties of materials, and have become important research tools in physics, materials science, and chemistry. However, conventional first-principles methods are computationally demanding, which limits their applications in large-scale and high-throughput materials research. Recent advances in deep learning have created new opportunities for electronic structure calculations. This article introduces the background and fundamental concepts of deep learning applied to electronic structure analysis. Taking the deep learning Hamiltonian approach as a representative example, we focus on how interpretable neural-network models can be designed by incorporating important physical priors to achieve accurate predictions of electronic structures. We also review recent progress in emerging directions such as universal materials models for electronic structures and unsupervised learning. These developments suggest that artificial intelligence is driving first-principles calculations from conventional computational approaches toward a new paradigm jointly driven by data and physics, opening new avenues for future scientific computing and materials discovery.

全文HTML

参考文献 (11)

下载: 全尺寸图片幻灯片

返回文章

用微信扫码二维码

分享至好友和朋友圈

深度学习电子结构计算简介

通讯作者: 徐勇,email:yongxu@tsinghua.edu.cn

An introduction to deep learning electronic structure calculations

Corresponding author: XU Yong, yongxu@tsinghua.edu.cn

计量

出版历程

深度学习电子结构计算简介

通讯作者: 徐勇,email:yongxu@tsinghua.edu.cn

English Abstract

An introduction to deep learning electronic structure calculations

Corresponding author: XU Yong, yongxu@tsinghua.edu.cn

全文HTML

目录