Dresden 2026 – wissenschaftliches Programm

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

AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz

AKPIK 4: Focus: Deep Learning in Electromagnetics Research

AKPIK 4.2: Hauptvortrag

Dienstag, 10. März 2026, 14:30–15:00, BEY/0127

Machine--learning assisted design of metasurfaces — Lukas Mueller, Alexander Wolff, Janis Krieger, Steffen Klingel, Ralf Stemler, and •Marco Rahm — RPTU Kaiserslautern-Landau, Erwin-Schroedinger-Strasse, 67663 Kaiserslautern, Germany

We present several applications of machine-learning-assisted metasurface design. The first study focuses on maximizing the received signal power for two users positioned at different angles relative to a reconfigurable intelligent surface (RIS) operating at 27 GHz and 31 GHz. The RIS must function as a frequency-selective yet independently tunable beam steerer, making the optimization of the varactor bias voltages a challenging task. The optimized voltage matrices successfully steered beams at both frequencies over angles from 10° to 45°. In parallel, the work explores metasurface designs with independent control of reflection amplitude and phase using physics-informed machine learning. To significantly reduce training data requirements, a Temporal Coupled Mode (TCM) model was introduced to capture the dynamic tuning behavior using only four simulations instead of hundreds. Machine-learning models predict TCM parameters directly from the metasurface geometry, enabling fast optimization. Furthermore, transfer learning was applied to composite unit-cell design, achieving comparable accuracy with far fewer simulations than direct training.

Keywords: metasurfaces; machine learning; reconfigurable intelligent surfaces