Hybrid physics-neural network computational spectrometer based on a random diffractive structure

Miniaturized spectrometers often trade spectral resolution and bandwidth against footprint and reconstruction fidelity. We report a structure-algorithm co-designed computational spectrometer that combines a random micropinhole diffractive encoder with hybrid physics-neural reconstruction. The encoder generates wavelength-sensitive diffraction patterns with low inter-wavelength correlation. The reconstruction first obtains a nonnegativity-constrained, smoothness-regularized Tikhonov estimate and then fuses it with radial speckle features using a neural network. Experiments over the 500-700 nm band achieve a reconstructed full width at half maximum (FWHM) of 1.46 nm for a single spectral peak and resolve doublets separated by 1.80 nm. The hybrid approach improves accuracy and noise robustness over physics-only and data-driven baselines, offering a compact route to broadband, high-resolution computational spectroscopy.