A Hybrid Data-Driven and Physics-Based Framework for Atmospheric Radiative Transfer Modeling

Accurate and efficient radiative transfer computations are critical for numerical weather forecast and climate prediction, but current radiative transfer models are computationally expensive. We developed a neural network (NN) scheme that emulates a physics-based radiative transfer scheme, achieving 4–5 times faster computation with high accuracy for well-trained, targeted scenarios. However, NN models struggle with extrapolation to out-of-sample weather or climate conditions. To address this, we propose HRadNN, a hybrid framework that uses a NN scheme by default and switches to a physics-based scheme whenever energy conservation is not maintained. Online simulations demonstrate HRadNN's accuracy in present-day climate and stability in extreme warming/cooling scenarios (up to 9 K change of the sea surface temperature) even without training specific to these scenarios. This indicates that machine learning can accelerate radiative transfer computation while maintaining accuracy and efficiency, provided it is quality-controlled by physical laws such as energy conservation.