The authors present IrisFlow, a query-based flow-matching framework for the inverse design of multilayer optical coatings that supports open-vocabulary material selection. Unlike typical amortized neural inverse design methods that rely on fixed vocabularies and discretized continuous variables, IrisFlow supplies target spectra, wavelength grids, and candidate materials at query time.
- The model uses discrete flow matching for material sequences sampled from a candidate bank and continuous flow matching for layer thicknesses without discretization.
- A single 136M-parameter model designs stacks with 2 to 100 layers.
- On a 224-task benchmark, it reconstructs in-distribution targets faithfully and maintains accuracy on a held-out material bank without retraining.
- The system designs bands up to 1100 nm beyond its training envelope and outperforms an autoregressive baseline on that baseline's material library.
- Fabricated color-displaying coolers achieved a CIEDE2000 color error of 3.1-5.2 while retaining 93-95% solar near-infrared reflectance.
IrisFlow demonstrates that open-vocabulary design can be successfully carried through to fabricated coatings, offering a flexible alternative to closed-world inverse design approaches.