Dresden 2026 – wissenschaftliches Programm

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

MM: Fachverband Metall- und Materialphysik

MM 9: Topical Session: Physics-driven Artificial Intelligence for Materials II

MM 9.5: Vortrag

Montag, 9. März 2026, 17:15–17:30, SCH/A251

Predicting the Thermal Properties of Semiconductor Defects with Equivariant Neural Networks — •Jonas A. Oldenstaedt, Manuel Grumet, Xiangzhou Zhu, Patrick Rinke, and David A. Egger — Physics Department, TUM School of Natural Sciences, Technical University of Munich, 85748 Garching, Germany

Predicting temperature-dependent properties of defects in semiconductors remains computationally challenging, even with density functional theory (DFT), due to the large supercells and long simulation times required for the calculations. In our recent work [1], we developed an active-learning workflow to accelerate defect calculations by combining two equivariant graph neural networks, trained using DFT calculations: MACE for predicting energies and forces needed in molecular dynamics, and DeepH-E3 for predicting electronic Hamiltonians needed for computing electronic properties across many thermally-excited configurations. We discuss the performance of our approach for predicting structural and electronic properties of intrinsic defects in the prototypical semiconductor GaAs, and demonstrate calculation accuracy comparable to DFT at much reduced computational cost. Furthermore, we discuss extensions of our approach to predict the thermal behavior of defects in more complex semiconductors such as halide perovskites.

[1] X Zhu, P. Rinke and D. A. Egger, arXiv:2511.18398 (2025).

Keywords: defects; semiconductors; machine learning; equivariant neural networks