Model-guided deep unfolding network for broadband reconstructive spectrometers with non-orthogonal aliasing

Reconstructive spectrometers based on wavelength-dependent spectral patterns are capable of delivering high resolution in compact devices. However, their practical deployment is fundamentally limited by the ill-posed inverse problem and severe non-orthogonal aliasing at large bandwidths. Here, we propose a model-guided deep unfolding network (MGDUN) that combines the forward physical model with learned priors. Experimental results demonstrate that the proposed method ensures high performance on various spectral profiles with non-orthogonal aliasing, even under severe noise. MGDUN also outperforms transmission matrix (TM) methods as well as deep learning (DL) baselines on the same dataset. MGDUN demonstrates accurate broadband reconstruction over a 20 nm window, achieving MAE/RMSE/SAM/APS of 0.0124/0.0213/3.43°/0.399 nm on random Lorentzian spectra, outperforming representative CNN, Transformer, DIP, and PnP-ADMM baselines. It further maintaining stable reconstruction under Gaussian, Poisson and drift perturbations. This moves beyond purely black-box approaches to continuous spectrum reconstruction under bandwidth constraints, and provides what is believed to be a new paradigm for DL-based algorithm design.