Fast All‐Sky Hyperspectral Infrared Radiative Transfer With Physically Recursive Cloud Treatment

Abstract Clouds strongly modify thermal infrared radiances through absorption, emission, and multiple scattering, yet rigorous line‐by‐line (LBL) gas absorption coupled with multi‐stream scattering remains too expensive for large ensembles of all‐sky hyperspectral simulations. We present a fast monochromatic radiative transfer (RT) model for the thermal infrared spectrum that retains the physical structure and vertical flexibility of the layered RT problem while embedding type‐wise trained radiative response models for the computationally dominant cloud‐layer scattering and emission calculations. The model combines accelerated LBL gas absorption, analytic thermal‐source integration, and an extended Adding‐Doubling framework that recursively couples clear and cloudy layers within a physically recursive framework. Cloud‐layer reflectance, transmittance, and emission responses are learned from DISORT solutions using single‐scattering optical properties, cloud optical thickness, spectral position, and angular coordinates as predictors, with radiance‐level constraints imposed during scattering training. Against LBL + DISORT references, the model reproduces clear‐sky brightness temperatures with an RMSE of 0.006 K and cloudy single‐layer, multilayer, mixed‐phase, and vertically extended cases with RMSEs below 0.05 K. A vertically distributed cirrus layer produces a maximum error of 0.55 K and RMSE of 0.28 K when compressed into one layer, whereas the proposed recursive treatment reduces the RMSE to 0.03 K. A full cloudy‐sky spectrum is computed approximately three orders of magnitude faster than the conventional LBL + DISORT pipeline. Application to FY‐4B/AGRI observations gives channel‐wise R 2 values of up to 0.96 after spectral convolution, indicating that the model can function as a physically constrained forward model for satellite radiance simulation.