Abstract Numerical weather prediction is the cornerstone of modern weather forecasting, yet its operational implementation demands vast computational resources. While artificial intelligence (AI)-based forecasting models offer a computationally efficient alternative, these purely data-driven approaches often sacrifice physical consistency. Here, we bridge physics-based and AI-based models through a novel, efficient hybrid framework that integrates a low-resolution atmospheric dynamical core with a neural operator in the multigrid architecture. This framework achieves performance comparable to that of state-of-the-art medium-range global weather forecasting models, while incurring much lower training costs, and simultaneously enhances the physical consistency that black-box models often lack. Furthermore, our framework provides substantial flexibility in the choice of dynamical cores, since the training process of the neural network does not require gradient propagation through the dynamical core, which ensures scalability to a wide range of operational forecasting systems.
HU et al. (Fri,) studied this question.