| --- |
| license: mit |
| pipeline_tag: other |
| --- |
| |
| # PRISM: Position-encoded Regressive Inverse Spectral Model |
|
|
| PRISM is a unified decoder-only autoregressive transformer designed for **inverse thin-film optical design**. Given a target optical spectrum, it generates a multilayer thin-film stack (specifying materials and thicknesses) whose physical response matches the target. |
|
|
| - **Paper:** [PRISM: Position-encoded Regressive Inverse Spectral Model for Multilayer Thin-Film Design](https://huggingface.co/papers/2605.26502) |
| - **Project Page:** [PRISM Playground](https://www.prism-playground.com/) |
| - **Code:** [GitHub Repository](https://github.com/wang-henry4/prism) |
|
|
| ## Architecture Key Innovations |
|
|
| PRISM introduces three primary architectural ideas that distinguish it from prior sequence-to-sequence approaches: |
|
|
| 1. **Spectrum prefix conditioning**: The target spectrum is projected into a single learned token and prepended to the decoder sequence, utilizing causal self-attention for target injection. |
| 2. **Cumulative-depth RoPE**: Instead of standard sequential token indices, Rotary Position Embeddings (RoPE) use the running cumulative physical depth (in nm) of the film stack. This provides the attention mechanism with a physically meaningful distance metric related to optical path length. |
| 3. **Dual output heads**: A shared transformer backbone feeds two specialized heads: |
| - **Material Head**: Predicts discrete material selection. |
| - **Thickness Head**: A multi-layer MLP that treats thickness as a continuous regression target (nm), predicting a thickness for *every* material in the vocabulary at each position to enable joint beam search. |
|
|
| ## Getting Started |
|
|
| ### Installation |
|
|
| ```bash |
| git clone https://github.com/wang-henry4/prism |
| cd prism |
| pip install -e . |
| ``` |
|
|
| ### Evaluation |
|
|
| You can evaluate a checkpoint using the provided script. It decodes structures, re-simulates them via the Transfer Matrix Method (TMM), and compares them against target spectra. |
|
|
| ```bash |
| python evaluate.py \ |
| --checkpoint saved_models/inverse/inverse_v1/best.pt \ |
| --val_path ./data/val/part_000.arrow \ |
| --nk_dir ./nk --n_samples 1000 \ |
| --beam_width 5 --length_penalty 0.3 \ |
| --plot_dir ./plots/inverse_eval |
| ``` |
|
|
| ## Citation |
|
|
| ```bibtex |
| @article{wang2026prism, |
| title={PRISM: Position-encoded Regressive Inverse Spectral Model for Multilayer Thin-Film Design}, |
| author={Wang, Runtian and Xue, Renhao and Chen, Baige and Wu, Hao}, |
| journal={arXiv preprint arXiv:2605.26502}, |
| year={2026} |
| } |
| ``` |
