Abstract Accurate and efficient fast calculation of microwave atmospheric transmittance is the key to ensuring the accuracy of microwave radiative transfer simulations, and also an indispensable critical link in satellite microwave remote sensing and data assimilation. Traditional fast parameterization methods, such as Optical Depth in Pressure Space (ODPS), rely on empirical regression with handcrafted predictors, which may fail to capture complex nonlinear relationships and vertical dependencies inherent in atmospheric absorption. Driven by advances in deep learning (DL), a physically constrained DL‐based transmittance model is proposed for the Microwave Humidity Sounder‐2 aboard FengYun‐3F (FY‐3F MWHS‐II) satellite. Three architectures—Multilayer Perceptron (MLP), Bidirectional LSTM (BiLSTM), and Multi‐head attention‐based Transformer—are developed and evaluated against transmittance generated by Monochromatic Radiative Transfer Model (MonoRTM), incorporating realistic spectral response functions. The Transformer model achieves the lowest root mean square error (RMSE) of 0.005 among DL models in water vapor channels, attributed to its self‐attention mechanism effectively capturing vertical dependencies in atmospheric features. Integrated into the Advanced Radiative Transfer Modeling System (ARMS), the DL parameterization achieves brightness temperature simulation accuracy comparable to ODPS, outperforming ODPS in window and temperature channels, though slightly underperforming in water vapor channels. However, raw DL Jacobian matrices exhibit non‐physical oscillations and biases. To enhance physical consistency, a fine‐tuning strategy is applied, penalizing deviations from reference Jacobians. After fine‐tuning, the model shows significantly improved physical consistency in the Jacobian matrices and adjoint model, and has been further verified in ARMS. The results validate the considerable potential of DL for parameterizing atmospheric gas absorption.
Huang et al. (Fri,) studied this question.