Deep learning tight-binding (DeePTB) approach for large-scale electronic simulations

发布时间2025-01-03文章来源 上海科技大学作者责任编辑系统管理员

报告人简介:
顾强强,中国科学技术大学人工智能与数据科学学院预聘副教授、博士生导师。2021年于北京大学物理学院量子材料科学中心获得理学博士学位,2021-2024年在北京大学数学科学学院大数据科学研究中心从事博士后研究工作,并同时在北京科学智能研究院 (AISI) 担任兼职研究员,于2024年11月加入中国科学技术大学工作。主要从事机器学习与计算材料模拟相关方向的研究工作,以一作/共一/通讯作者身份发表在Nat. Commun.、Phys. Rev. Lett、Sci. Bull.、Adv. Mater.、Nati. Sci. Rev、PNAS.等多篇学术论文,主持开发 DeePTB、DFTIO、TBSOC 多个开源软件。

讲座摘要:
Density functional theory (DFT) is a powerful tool for simulating the electronic properties of materials, but its high computational cost limits its application to large-scale systems. This report introduces the innovative deep learning-based models in the DeePTB package: DeePTB-SK[1] and DeePTB-E3[2] and their applications on non-equilibrium Green’s function (NEGF) quantum transport simulations[3]. DeePTB-SK is a deep learning-enhanced Slater-Koster (SK) tight-binding (TB) approach that efficiently predicts local environment-dependent SK parameters, thereby predicting TB Hamiltonians. This enables efficient electronic structure simulations of systems containing millions of atoms. DeePTB-E3 aims to predict multiple quantum operator matrices in the LCAO-DFT framework with state-of-the-art accuracy while dramatically improving computational efficiency. Its core innovation lies in designing a strictly localized equivariant message-passing model (SLEM), constructing local equivariant representations for quantum operators including Hamiltonian, density matrix, and overlap matrix. These works represent significant progress in combining deep learning with traditional quantum mechanics methods, paving the way for large-scale electronic structure calculations and material simulations.

邀请人:张雪峰,李刚