A differentiable multi-physics learning model for robust single-image dehazing under complex atmospheric conditions

Restoring clear visibility from a single hazy image remains a challenging inverse problem when multiple atmospheric phenomena jointly affect image formation. In real environments, light propagation can be influenced not only by particle scattering but also by turbulence-induced blur and wavelength-dependent attenuation. Most existing dehazing approaches primarily address scattering, which limits their physical consistency and robustness in heavy haze or long-range scenes. In this work, we introduce a differentiable multi-physics learning framework that is designed to integrate scattering, turbulence, and spectral attenuation within a unified image formation model. The degradation process is expressed as a compact linear observation equation, where turbulence is represented through a spatial convolution operator and spectral effects are modeled via channel-dependent attenuation. Since the governing parameters cannot be directly inferred from a single observation, the formulated as an iterative refinement process guided by a learned approximation of the physical model. A lightweight neural estimator predicts the observation operators, while a learnable spectral module captures wavelength-dependent behavior. An iterative regularization network stabilizes the reconstruction process and prevents error propagation. Experiments conducted on the RESIDE-SOTS benchmark show competitive performance, achieving PSNR/SSIM values of 27.3/0.94 for indoor scenes and 30.2/0.94 for outdoor scenes with only a few optimization steps. Additional analyses confirm improved robustness under noise perturbations and severe illumination variations, highlighting the effectiveness of differentiable learning with physics-inspired constraints for real-world atmospheric image restoration.