Regensburg 2025 – wissenschaftliches Programm

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

HL: Fachverband Halbleiterphysik

HL 3: Focus Session: Machine Learning of semiconductor properties and spectra

HL 3.5: Hauptvortrag

Montag, 17. März 2025, 11:45–12:15, H17

OptiMate: Artificial intelligence for optical spectra — •Malte Grunert and Max Großmann — Theoretical Physics I, Institute of Physics, Technische Universität Ilmenau, 98693 Ilmenau, Germany

Machine Learning (ML) is currently transforming computational materials science - many material properties can now be predicted to ab initio accuracy almost instantly using modern ML techniques. Until recently, optical properties such as absorption in the UV/VIS region were excluded from the ever-growing list of machine-learned properties, due to lack of high-quality training data for electronic excitations. To address this missed opportunity, we present OptiMate [1], a graph attention neural network trained on the largest high-quality, high-throughput dataset of optical properties available to date – a dataset that we have independently generated. OptiMate is capable of predicting the complex optical properties of a wide class of materials up to the XUV range with quantitative accuracy. In addition, OptiMate learns physical properties of spectra like continuity and causality directly from high-quality data, without such properties being enforced by constraints in the model architecture or with penalties during training. We detail the workings of OptiMate, show that it (and probably other complex models) learns a surprisingly meaningful representation of the material space, and preview current developments such as transfer learning to higher levels of theory.

[1] M. Grunert et al., Phys. Rev. Mater. 8, L122201 (2024)

Keywords: Machine Learning; Graph Neural Networks; Dielectric Function; Optical Materials; High-Throughput