Metasurface colour router inverse-designed with a neural-network-accelerated genetic algorithm, fabricated and measured
A U-Net trained on 140,000 finite-difference time-domain simulations stood in for the solver inside a genetic algorithm, which evaluated candidate layouts for a single-layer silicon nitride colour router over a 2 µm Bayer cell; the fabricated device measured 88.8% total transmission against 95.4% predicted, though with per-colour routing efficiencies and crosstalk worse than simulated.
- Model
- U-Net surrogate (about 4.9 million parameters) inside a genetic algorithm
- Field
- Physics
- Date
- 2026-09-08
- Human collaborators
- Hyoseok Park, Samuel Kim, Duk-Yong Choi, Yeonsang Park, Myungjae Lee
What was found
Camera sensors filter colour by absorption, discarding most incoming light; a colour router instead steers red, green and blue light onto their sub-pixels. The design here is a 16 by 16 grid of 600 nm silicon nitride pillars at 125 nm pitch. The surrogate predicts full transmitted field maps about times faster than the solver, which let the search run 4,000 generations against the roughly ten thousand structures a solver-driven genetic algorithm can afford. Because it predicts fields rather than a single score, the same trained network can be reused for other photodiode aperture layouts without new training data. In a colour-chart reconstruction using the measured spectra, the router gave brighter images than a literature colour-filter-array response at some cost in spatial resolution.
Novelty check
OpenAlex, 2026-11-12: search "metasurface color router deep learning genetic algorithm", 15 records. Earlier colour routers were already designed by genetic algorithm (Zou et al. 2022, cited by the paper, 84% measured energy-utilisation efficiency), by freeform optimisation (Science Advances 2024) and with deep learning (Nanoscale 2024; a deep-learning metalens RGB-NIR router, Nanomaterials 2024). What is new is the field-map surrogate inside the genetic search and its reuse across aperture layouts, not deep-learning design of colour routers as such.
Caveats and known objections
Peer-reviewed in Reports on Progress in Physics, open access, by the developers at Chungnam National University with Seoul National University and the Australian National University; no independent replication. The paper's claim to the best performance among compared routers rests on simulated values. Measured peak routing efficiencies were 56.2, 61.2 and 36.5% for red, green and blue, and measured crosstalk was about 48 to 51%, higher than simulated, so the experimental advance over the 2022 genetic-algorithm design is not clear-cut. The 3.6-fold transmission figure in press coverage compares the measured router with a literature colour-filter-array response in a reconstruction, not with a fabricated sensor. The surrogate code is public; trained weights, training data and measured spectra are available on request. Autonomy is search-scaffold: the objective, design grid and genetic algorithm were human-built, and the network is a fast simulator inside that loop.
No independent check is recorded on this entry yet.
Reading the primary source closely enough to say whether it supports the claim counts as a check, and you are credited on the entry.
Or on GitHub: submit a check challenge the grade send a correction or send a pull request
Flag this for triage
Signals order the review queue and nothing else. They are never published, and they never move a grade: that takes a citation.
Entry history (1 event)
- AddedEntered the registry graded Peer reviewed and Search scaffold.
Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.
Graded peer reviewed for verification and search scaffold for autonomy. What these mean.
Cite this entry
whataifound.org. (2026). Metasurface colour router inverse-designed with a neural-network-accelerated genetic algorithm, fabricated and measured. whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2026-09-08-metasurface-color-router-unet-ga
BibTeX
@misc{whataifound-chungnamnationaluniversity-2026-ga,
title = {Metasurface colour router inverse-designed with a neural-network-accelerated genetic algorithm, fabricated and measured},
author = {{whataifound.org}},
year = {2026},
howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
note = {Result by Chungnam National University. Verification: Peer reviewed. Autonomy: Search scaffold.},
url = {https://whataifound.org/finding/2026-09-08-metasurface-color-router-unet-ga}
}