Two-Stage Deformable-Convolutional Inverse Design of Nanophotonic Absorbers from Optical Spectra

A recent preprint on arXiv (2608.11860) presents a novel two-stage deformable-convolutional approach aimed at the inverse design of nanophotonic absorbers. This technique reconstructs geometries of metal-insulator-metal resonators from 80-dimensional absorption spectra, tackling the complexities of mapping spectra to geometries due to non-uniqueness and intricate features. The absorption spectrum is transformed into a 150x4x4 latent representation, which is then decoded into a 64x64 resonator mask. The training process merges supervised reconstruction with least-squares adversarial refinement, starting from the optimal supervised checkpoint. A three-run ablation study evaluates deformable convolution against standard convolution, involution, Dynamic Conv, and ODConv, achieving 20.79±0.31 dB PSNR and 0.8501±0.0082 SSIM, surpassing plain convolution by 2.16 dB and 0.0831. This research is significant for computational nanophotonics and inverse design, with implications for optical devices.