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

MM: Fachverband Metall- und Materialphysik

MM 17: Data-driven Materials Science: Big Data and Workflows II

MM 17.2: Vortrag

Dienstag, 10. März 2026, 14:15–14:30, SCH/A251

Broken neural scaling laws in machine learning for optical properties of metals — •Max Großmann, Marc Thieme, Malte Grunert, and Erich Runge — Institute of Physics and Institute of Micro- and Nanotechnologies, Technische Universität Ilmenau, 98693 Ilmenau, Germany

Neural scaling laws guide the development of machine-learning models and their training datasets. Here, we investigate them in the context of materials science, where data are inherently costly and scarce, using dielectric functions of metals as an example. We compute dielectric functions for 205,224 intermetallic compounds using high-throughput ab initio calculations and train two multi-objective graph neural networks, OptiMetal2B and OptiMetal3B—the latter incorporating three-body interactions—to predict the complex interband dielectric function and the Drude frequency. Systematic variations in the number of training data and model parameters reveal so called "broken" neural scaling laws. Data scaling follows a smoothly broken power law, with steeper slopes occurring above 20,000 materials. In contrast, parameter scaling follows a conventional power law but saturates at around ten million parameters. Including three-body interactions improves accuracy by about 12% but leaves scaling slopes essentially unchanged. These findings suggest that, in the context of spectroscopy, expanding high-quality datasets is a more effective way to improve machine-learning models than optimizing network architectures, increasing body order, or merely increasing network size.

Keywords: Neural Scaling Laws; Graph Neural Networks; Dielectric Functions; High-Throughput Calculations; Intermetallic Compounds