CHROMA: Conversational Human-Readable Optical Multilayer Assembly for Natural Language-Driven Inverse Design of Structural Coloration

Mingxuan Li, Jungtaek Kim, Oliver Hinder, Paul Leu

Keywords: Structural Color, Inverse Design, LLM, Multilayer Thin Films

Abstract: We present CHROMA, a human-in-the-loop system that turns natural language-based requests into manufacturable multilayer thin-film designs for structural coloration. CHROMA combines a frozen large language model encoder with a compact, trainable Transformer decoder over a discrete material--thickness vocabulary, and couples decoding with a differentiable transfer-matrix-method verifier. Users write prompts like ``teal reflection at $60^\circ$, six layers, include ZnO, avoid Al next to Mn.'' Then, CHROMA proposes a stack, simulates the optical response, and enforces hard constraints. We describe dataset construction from physics-generated spectra paired with templated natural-language paraphrases, report scaling trends over encoder size and decoder depth, and demonstrate angle-resolved analyses of selected stacks. CHROMA's conversational interface provides exceptional flexibility for defining design targets.

Submission Track: Paper Track (Short Paper)

Submission Category: AI-Guided Design

Institution Location: Pittsburgh, US

Submission Number: 81