Dresden 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 9: Topical Session: Physics-driven Artificial Intelligence for Materials II
MM 9.4: Talk
Monday, March 9, 2026, 16:45–17:00, SCH/A251
Making equivariant graph neural network prediction of electronic structure properties fast and accurate — •Chen Qian1, Valdas Vitartas1, James Kermode1, and Reinhard J. Maurer1,2 — 1University of Warwick, UK — 2University of Vienna, AT
Machine learning predictions of band structures and equivariant electronic properties, such as real-space density functional theory (DFT) operator matrices and response properties, have the potential to accelerate electronic structure prediction while avoiding expensive ab initio calculations. However, most current models struggle to strike a balance between prediction accuracy and inference speed. Following our previous work on the equivariant graph neural network MACE-H [arXiv:2508.15108], we assess the model’s performance on DFT operator matrices and, subsequently, on a property based on electron-phonon response, the electronic friction tensor. We compare its applications across various datasets. Furthermore, we analyze several existing algorithm- and hardware-based acceleration methods for the computationally intensive Clebsch-Gordan tensor product in terms of accuracy and computational efficiency, and discuss their respective suitable application scenarios. To this end, we present the MACE-H2 framework, which features an O(3) equivariant graph neural network with many-body expansion and suitable acceleration approaches and provides separate routines for DFT operator matrices and electron-phonon response prediction. The model achieves high accuracy and inference speed and is suitable for high-throughput band-structure calculations and material discovery.
Keywords: Machine Learning; Deep Hamiltonian; Electronic Structure; Electronic friction; Density Functional Theory