Conditional Generative Adversarial Network for Liquid Crystal Metasurface Spectrometers

Reconstructive spectrometers based on tunable metasurface optics and computational algorithms have emerged as a promising solution for compact spectral analysis over a broad wavelength range. However, the spectral reconstruction process is inherently ill-posed since a single sensor response can correspond to multiple possible spectra. Discriminative deep learning models cannot solve this problem since they merely learn deterministic mappings, leading to unstable solutions in the presence of measurement noise and spectral ambiguities. We proposed a conditional generative adversarial network model to address this challenge. Based on the experimental response matrix of a fabricated liquid crystal metasurface device, we demonstrate that the proposed model can resolve Gaussian peaks with 0.5-nm full width at half maximum across a 130-nm operational bandwidth and is robust against noise up to 20-dB signal-to-noise ratio in simulation.