Regensburg 2025 – wissenschaftliches Programm

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

MM: Fachverband Metall- und Materialphysik

MM 18: SYMD contributed

MM 18.4: Vortrag

Mittwoch, 19. März 2025, 11:00–11:15, H23

Comparing linear and deep learning surrogate models of materials electronic structure — •Valdas Vitartas, Chen Qian, James Kermode, and Reinhard Maurer — University of Warwick, Coventry, UK

The self-consistent electronic Hamiltonian matrix from Density Functional Theory (DFT) gives access to the electronic band structure and the density of states of a material, albeit at a large computational cost. Over recent years, several surrogate models based on linear parametrization and deep learning have been proposed to efficiently learn the electronic Hamiltonian as a function of the configuration and composition of materials. In this work, we compare two such models, the ACEhamiltonians [npj Comput. Mater. 8, 158] and MACE-H. Both provide a representation of the Hamiltonian in atomic orbital basis in terms of an equivariant many-body expansion of local atomic environments. In the case of ACEhamiltonians, the model parametrization is linear; for MACE-H, the representation serves as input to a message-passing neural network. The models are trained on reduced, valence-only Hamiltonian matrices for bulk gold and silicon generated from all-electron DFT via an approximately eigenspectrum-conserving transformation. We discuss the inherent strengths and weaknesses of the models by illustrating their accuracy, performance, data efficiency, and their ability to predict electronic quantities of interest for out-of-distribution configurations.

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