SpecFormer: full-spectrum encoder–decoder inverse design of multilayer optical thin films

Inverse design of multilayer optical thin films aims to generate realizable structures from target spectra. Autoregressive generative models represented by OptoGPT have become an important paradigm; for multichannel targets spanning the visible, near-infrared, and thermal-infrared regions, explicitly organizing local spectral features, cross-window constraints, and global functional objectives can provide richer conditional representations for structure generation. Here, we propose SpecFormer, a full-spectrum encoder–decoder framework for multilayer thin-film inverse design. SpecFormer uses a spectrum encoder to organize the complete spectrum into token memory, summary memory, and a global embedding. A cross-attention structure decoder then generates variable-length material–thickness token sequences, followed by differentiable transfer-matrix refinement of the continuous thicknesses. On 3000 fixed test targets, SpecFormer achieves a mean Plain MAE of 0.0041 and a mean SSIM of 0.9993. Its Plain MAE is lower than the 0.0234 obtained by OptoGPT under the matched protocol; 73.60% and 92.67% of the samples reach MAE ≤ 0.005 and MAE ≤ 0.01, respectively, and the improvement remains stable under equal-channel, bandwidth-normalized, and physically weighted metrics. Linear probing, concept-direction editing, attention heatmaps, and activation patching show that the encoder converts target spectra into physical condition representations that are readable, directionally editable, and usable by the decoder. Fabrication-constraint and perturbation analyses quantify the effects of material-dependent minimum thicknesses, thickness errors, interface roughness, and optical-constant uncertainty. Under partial-channel inputs without retraining, all 3000 targets yield valid structures, with a mean Plain MAE of 0.0255 over the valid sampled points. SpecFormer is further applied to visible-transparency-tunable cooling windows, photosynthesis-selective cooling films, and band-edge thermal-management front coatings for crystalline-silicon photovoltaics, showing improvements or complementary advantages over task-related literature baselines. These results establish SpecFormer as a high-accuracy, interpretable, and physics-constrained framework for full-spectrum inverse design of multilayer optical thin films.