A Fast Polarized Radiative Transfer Model for Multi‐Layer Cloudy Atmospheres

Polarized radiative transfer models (PRTMs) are essential in atmospheric remote sensing and particle microphysics researches, but their applications have been limited due to their high computational cost and complexity. This study proposes an efficient plane‐parallel PRTM by a combination of machine learning (ML) techniques and traditional physical radiative transfer models. The atmosphere is decomposed into layers with independent optical properties, each have type‐specific polarized scattering parameterizations. To be more specific, a physics‐informed neural network (PINN) model that directly incorporates physical constraints into the ML process is developed to capture complex polarized interactions of cloudy layers. By independent training for atmospheric layers of different types, we can effectively enhance the generalization capability across realistic atmospheric profiles. The interconnection of the polarized states across multiple layers is achieved using the adding‐doubling method, which is optimized through a discretization scheme for efficient computation. The PINN model demonstrates remarkable physical consistency when simulating the radiative properties of both ice and water clouds. The mean absolute relative errors (MAREs) are 0.7% for ice clouds and 1.0% for water clouds, with the latter showing errors in regions characterized by strong scattering peaks. For multi‐layer cloud cases, the model produces an MARE of 0.9%. Compared to rigorous PRTMs, our model improves the computational efficiency by 3 orders of magnitude. The proposed PRTM can be readily adapted to account for distinct particle types, especially aerosols, by modifying the optical parameterization in the training data set, showing promise for more extensive applications in atmospheric research.