Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 2: Invited Talk: Richard Hennig
MM 2.1: Hauptvortrag
Montag, 9. März 2026, 09:30–10:00, SCH/A251
Deep-Learning and Generative AI for the Discovery of Electron-Phonon Superconductors — •Richard Hennig — University of Florida, Gainesville, Florida, USA
The search for new superconductors with higher critical temperatures, Tc, and critical fields, Hc, is limited by the cost of electron-phonon calculations and the vastness of compositional and structural space. To overcome both obstacles, we develop an integrated deep-learning workflow for conventional, electron-phonon-mediated superconductors.
First, we introduce BEE-NET, an ensemble of equivariant graph neural networks trained to predict the Eliashberg spectral function α2F(ω) and Tc directly from crystal structures, optionally augmented by the phonon density of states. Unlike traditional approaches that learn Tc directly, predicting α2F(ω) treats superconductors and non-superconductors on equal footing and, together with explicit phonon-spectrum information, leverages electron-phonon physics to improve predictions for rare superconducting materials. Embedded in a multi-stage screening pipeline that combines elemental substitution strategies with machine-learned interatomic potentials, BEE-NET scans over 1.3× 106 candidate structures and down-selects to 741 dynamically and thermodynamically stable compounds with DFT-confirmed Tc>5 K, including two experimentally realized new superconductors.
Finally, I will briefly show how guided diffusion models for crystal structures and our Open Materials Generation (OMatG) framework based on stochastic interpolants extend this physics-informed, data-driven approach to the generative design of new superconductors.
Keywords: Deep learning; Superconductivity; Materials discovery; Generative AI; Experimental validation