IrisFlow | Zhiyi Li

IrisFlow is a generative model for the inverse design of multilayer optical coatings. You hand it a design request: the spectrum you want, the wavelength band you care about, the materials you can actually deposit, given only as their optical curves, and how many layers you allow. It generates a pool of candidate stacks from pure noise, denoising the discrete material choices and the continuous thicknesses together. Every candidate is re-simulated exactly, so you select by whatever rule matters to you and take the recipe to the deposition chamber.

Try the model How it works Paper

Why IrisFlow

Try the real model

The other panels on this page are illustrations. This one runs IrisFlow itself: your request goes to a Hugging Face Space, which borrows a GPU for a few seconds, generates several designs and re-simulates each one with the transfer-matrix method (normal incidence, from air onto a fused-silica substrate; layer 1 sits on the substrate). The materials that were never seen in training are given to the model only as their n,k curves. The model was trained on 380–1400 nm, but it takes the wavelengths as input: you can choose one or several bands and go out to 2500 nm, as in the paper’s wavelength-extension tests, using the materials whose measured n,k data reach that far. Hugging Face gives each anonymous visitor only a few GPU runs; after that, or when no GPU is free, the request runs on a CPU, which takes a little longer. If nobody has used the Space for a while it sleeps, and the first request takes about a minute to wake it.

How it works

One trained model, and what it gives a coating designer.

Paper

IrisFlow is described in the preprint below. I conceived and designed the study, developed the framework, implemented the code, ran the simulations and benchmarks, and wrote the manuscript; my co-authors fabricated the samples and supervised the project.

Flexible Inverse Design of Optical Multilayer Coatings Using Discrete-Continuous Flow Matching

Zhiyi Li, Yuheng Jin, Yidan Huang, Nan Chen, Hongyan Fu, and Yikun Bu

arXiv preprint arXiv:2607.08392, 2026

First author · preprint under review

Abs arXiv Bib HTML Website

Optical multilayer coatings are used in imaging optics, lasers, telecommunications, solar cells, and radiative cooling, where the spectral response required by each application is achieved by a well-designed layer structure. However, existing design methods have limitations. Local methods refine a starting design efficiently but depend on its quality, and global searches require up to millions of transfer-matrix evaluations per target. Existing deep-learning inverse-design models are fast, but they are trained for a fixed set of materials and a fixed wavelength band, and they either optimize the thicknesses for a fixed material sequence or choose them from a discrete grid. In this paper, we present IrisFlow, a generative model that addresses these limitations by taking each candidate material as its optical constants and the target spectrum together with its wavelengths as input. It generates the material and the continuous thickness of every layer jointly, by coupled discrete and continuous flow matching in a single network. We demonstrate its performance through design with materials outside the training library and on bands beyond the training range, and through comparisons with classical optimizers and an existing deep-learning model. We also demonstrate its usability in a practical design workflow by designing four colored solar-reflective coatings with the calibrated optical constants of an existing deposition system. Combined with a local refinement of its ten best designs, it reduces the number of transfer-matrix evaluations about 500-fold relative to needle optimization at comparable accuracy. The code, trained model, and training and test data are publicly available.
@article{li2026irisflow, title = {Flexible Inverse Design of Optical Multilayer Coatings Using Discrete-Continuous Flow Matching}, author = {Li, Zhiyi and Jin, Yuheng and Huang, Yidan and Chen, Nan and Fu, Hongyan and Bu, Yikun}, journal = {arXiv preprint arXiv:2607.08392}, year = {2026}, note = {First author · preprint under review}, }