Regensburg 2025 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

MM: Fachverband Metall- und Materialphysik

MM 18: SYMD contributed

MM 18.5: Vortrag

Mittwoch, 19. März 2025, 11:15–11:30, H23

MACE-H: Equivariant Hamiltonian prediction with many-body expansion message passing — •Chen Qian, Valdas Vitartas, James Kermode, and Reinhard J. Maurer — University of Warwick, UK

The machine learning prediction of Kohn-Sham Density Functional Theory (DFT) Hamiltonians has the potential to accelerate the prediction of electronic properties, such as electronic band structures and electron-phonon coupling, while avoiding computationally expensive self-consistent field iterations. We introduce the MACE-H graph neural network, which combines the MACE body-order expansion message passing scheme with node-degree expansion blocks to efficiently generate messages that incorporate all relevant SO(3) irreducible representations. This model achieves high accuracy and high computational efficiency in capturing the local chemical environment. We demonstrate the model performance using several open materials benchmark datasets for 2D materials, achieving sub-meV prediction errors on matrix elements. Moreover, we discuss how the many-body expansion achieves higher data efficiency and examine its effect on out-of-distribution prediction for nanostructures featuring long-range interactions. To assess prediction outputs, we analyze the correlation between errors and hermiticity. The high computational efficiency and accuracy make the model a good candidate for electronic structure prediction in large-scale systems and high-throughput material screening.

Keywords: Machine Learning; Hamiltonian matrix; Electronic Structure; Density Functional Theory