Neural reparameterization for nonlocal metasurface topology optimization
APL Photonics · 01.10.2025 · 10 (10)
DOI: 10.1063/5.0283122 ↗Kurz gefasst
Fachartikel in der Zeitschrift APL Photonics, online erschienen am 13.10.2025. Verfasst von Lincoln Clark und 2 weiteren Autor:innen. Die Arbeit wurde bisher 5-mal zitiert (Crossref) und nennt 58 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Australian Research Council.
Abstract (Originalsprache)
Metasurfaces are emerging as a paradigm for performing all-optical, energy efficient analog computing and image processing in an ultra-compact form-factor. While various operations have been demonstrated with simple metasurface geometries, broadening the range of available processing outputs will require non-trivial and non-intuitive designs. Topology optimization (TO) is a powerful inverse design technique capable of designing such metasurfaces with significantly enhanced optical properties. The two most important design considerations are the quality of the optical performance and the minimum feature size in the resulting device. Larger features can reduce the mismatch between design and realized device performance, increasing fabrication speed and extending the scalability of the fabrication. Metasurface TO is typically performed on a pixel grid to determine the material distribution at each point. Here, we compare two pixel-based methods to neural-network based parameterizations, where the parameters of a neural-network are trained to output a metasurface design. We produce 150 different computational metasurface design tasks to compare the performance of each parameterization with regard to the performance and feature sizes. In comparison to the best pixel-based approach, a hybrid neural-network and pixel method produced designs with a median minimum feature size 2.8 times larger with only a 1.4 times increase in RMS error. Compared to a pixel method capable of producing similar critical dimensions, the optical performance of the hybrid approach is improved by 39%. Finally, as a proof of concept, we use this neural-network based parameterization to design an angular bandpass filter with relatively large critical dimensions and non-trivial performance.
Förderung
- Australian Research Council
Zitieren
APA
Clark, L., Roberts, A., & Wesemann, L. (2025). Neural reparameterization for nonlocal metasurface topology optimization. APL Photonics, 10(10). https://doi.org/10.1063/5.0283122
Vancouver
Clark L, Roberts A, Wesemann L. Neural reparameterization for nonlocal metasurface topology optimization. APL Photonics. 2025;10(10). doi:10.1063/5.0283122
BibTeX
@article{Clark2025,
author = {Clark, Lincoln and Roberts, Ann and Wesemann, Lukas},
title = {Neural reparameterization for nonlocal metasurface topology optimization},
journal = {APL Photonics},
year = {2025},
volume = {10},
number = {10},
doi = {10.1063/5.0283122}
}
EUVIMED