High-quality full stokes polarimetric spectroscopy reconstruction using a model-compensated physics-informed neural network for channeled spectropolarimetry

Channeled spectropolarimetry (CSP) enables the simultaneous acquisition of full Stokes parameters spectra in a single-shot, providing a robust solution for dynamic and complex optical measurements. However, accurate spectral reconstruction is often hindered by systematic errors and the limitations of a simplified CSP physical model, resulting in reduced reliability of its applications. To address these challenges, this paper introduces a model-compensated physics-informed neural network (MC-PINN), which integrates an improved physical model with deep learning to enhance reconstruction performance. The MC-PINN framework incorporates a prediction network to approximate data and physical laws, together with a compensation network to correct discrepancies between the observed data and the physical model. This approach significantly reduces errors caused by model inaccuracies, enabling high-quality reconstruction even with limited and noisy data. Simulations and experiments confirm that MC-PINN outperforms traditional methods, achieving superior accuracy and robustness in reconstructing high-frequency spectral features and complex polarization states. This work demonstrates the potential of MC-PINN to improve data reliability, thereby expanding the applicability of CSP to diverse and challenging scenarios.