Dresden 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
HL: Fachverband Halbleiterphysik
HL 23: Transport Properties
HL 23.5: Talk
Wednesday, March 11, 2026, 10:30–10:45, POT/0051
Deep-learning Hamiltonian Acceleration of Electrical Transport Predictions using the Non-Perturbative ab initio Kubo-Greenwood Method for Strongly Anharmonic Materials — •Juan Zhang1,2, Kisung Kang3, and Matthias Scheffler1 — 1The NOMAD Laboratory at the FHI of the Max Planck Society, Berlin — 2Department of Optical Science and Engineering, Fudan University, Shanghai — 3School of Materials Sciences and Engineering, Chonnam National University, Gwangju
Thermal insulators, e.g., needed for efficient thermoelectric materials, feature strong anharmonicity. As a consequence, a perturbative approach of electron-phonon interactions and even the phonon concept for describing vibrations may become invalid. The non-perturbative ab initio Kubo-Greenwood (aiKG) method provides an approach for evaluating electron and hole mobilities [1]. However, it requires substantial computational costs due to its high requirements with respect to statistical averages, large supercells, and extrapolation strategies to the zero-frequency limit. This work introduces an AI-assisted aiKG framework for the FHI-aims code, which incorporates the neural-network model, DeepH [2], trained to predict the Kohn-Sham Hamiltonian. Using the thermal insulator KI, we demonstrate the capabilities and predictive power of the approach, which substantially accelerates the calculation of large supercell electronic band structures, temperature-dependent spectral functions, and carrier mobilities with high accuracy.
[1] J. Quan et al. Phys. Rev. B 110, 235202 (2024).
[2] X. Gong, et al. Nat Commun 14, 2848 (2023).
Keywords: Deep-learning Hamiltonian; Electrical Transport; Kubo-Greenwood method