Mainz 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
Q: Fachverband Quantenoptik und Photonik
Q 8: Nanophotonics and Integrated Photonics II
Q 8.6: Talk
Monday, March 2, 2026, 18:30–18:45, P 3
Deep Learning Based Inverse Design of Nanophotonic Devices — •David Lemli1,2,3, Marco Butz1,2,3, Marlon Becker4, Benjamin Risse4, and Carsten Schuck1,2,3 — 1Department for Quantum Technology, University of Münster, Heisenbergstr. 11, 48149 Münster, Germany — 2Center for Soft Nanoscience, Busso-Peus-Str. 10, 48149 Münster, Germany — 3Center for Nanotechnology, Heisenbergstr. 11, 48149 Münster, Germany — 4Department for Geoinformatics, University of Münster, Heisenbergstr. 2, 48149 Münster, Germany
Photonic integrated circuits constitute a key platform for all areas of quantum technology, driving the need for nanophotonic components that achieve high optical performance while adhering to fabrication constraints such as minimum feature sizes.
Topology optimization provides a powerful framework for designing highly efficient, compact, and multifunctional photonic devices. Here, we present Memory Metropolis (MeMe), a deep-learning enhanced discrete topology optimization algorithm that employs deep template networks, a novel neural network architecture for generating proposal distributions in simulated annealing. By promoting the clustering of individual pixels, MeMe produces device geometries compatible with state-of-the-art lithographic processes. We experimentally validate the performance of optimized devices on the emerging tantala-on-insulator platform. Fabrication compatibility naturally emerges from MeMe's optimization process, representing a key algorithmic innovation in discrete inverse design.
Keywords: Inverse Design; Machine Learning; Nanophotonic Devices