Refining Clear-Sky AI Radiative Transfer Model Forward Predictions and Jacobian Accuracy for ATMS Using a ResNet and Physical Constraints

Radiative Transfer Models (RTMs) are pivotal in remote sensing and data assimilation, providing simulations of satellite-measured radiances and corresponding Jacobians—matrices that quantify how radiative quantities change in response to variations in atmospheric or surface parameters. These Jacobians lie at the heart of variational data assimilation and drive a wide range of inverse modeling applications. Although traditional RTMs yield high-fidelity results, they are computationally intensive, which is increasingly problematic given the high spatial, temporal, and spectral resolutions measured by modern sensors. Recent advances in artificial intelligence (AI) have led to AI-derived “fast” RTMs that substantially accelerate forward predictions, (e.g., brightness temperatures). However, accurately estimating and stabilizing AI-derived Jacobians remains a key challenge. In this paper, we present a new AI-CRTM framework based on a Residual Neural Network (ResNet), augmented by physical constraints using the reference Jacobians from the traditional Community Radiative Transfer Model (CRTM). By leveraging residual connections, the proposed approach mitigates vanishing gradients and captures complex, multi-scale dependencies in radiative transfer physics. Physical constraints are applied during training to align the learned partial derivatives with trusted RTM references, thus ensuring physically consistent Jacobians. The AI_CRTM was initially trained and tested for simulated measurements from the Advanced Technology Microwave Sounder (ATMS) in global clear-sky conditions. Forward-model accuracy is accessed through comparisons with both CRTM-simulated radiances and high-spectral-resolution line-by-line simulations from the MonoRTM model. The results demonstrate that incorporating these constraints, alongside ResNet’s residual structure, significantly improves both forward prediction accuracy and Jacobian fidelity. This hybrid strategy addresses ill-posed issues in AI Jacobian inference, broadens applicability to different atmospheric scenarios, and paves the way for rapid and robust data assimilation.