To achieve high-precision radiative simulation as well as cloud and precipitation monitoring and forecasting, a physicallyconstrained, Artificial Intelligence (AI)augmented framework for satellitebased radiation, cloud, and precipitation analysis was constructed, namely the DaYu Radiation-Cloud-Precipitation Analysis System (DaYu-RCPAS). It consists of four core components: the DaYu Radiative Transfer Model (DaYu-RTM), the DaYu CLoud Analysis System (DaYu‐CLAS), the DaYu PRecipitation Analysis System (DaYu‐PRAS), and the DaYu Nowcast System (DaYu-Nowcast).
As the physical forward operator of DaYu-RCPAS, DaYu-RTM is designed for efficient simulation of satellite imager radiances in cloudy atmospheres, and it incorporates three major physical modules: gaseous absorption parameterizations using alternate mapping correlated k-distribution methods, radiative transfer solvers based on the discrete ordinate adding approximation, and neural network parameterizations of cloud particle optical properties. The DaYu-CLAS framework is composed of single-layer cloud retrieval, double-layer cloud retrieval, and cloud forecasting models. Specifically, it integrates several single-layer cloud retrieval models, including Cloud-ResUNet, Cloud-SmaAtUNet, and CloudDiff. For double-layer cloud retrieval, it adopts Overlap-CloudDiff and a hybrid algorithm combining DaYu-RTM simulation with deep learning. Additionally, it also incorporates the Cloud-FNO model for cloud forecasting. It is worth noting that based on the Cloud-SmaAtNet model in the DaYu-CLAS, we have constructed a Global Cloud physical property Products (DaYu-GCP) covering 2000–2022, with a temporal resolution of 3 hours and a spatial resolution of 0.07°. To complete the DaYu system, DaYu‑PRAS focuses on precipitation analysis, incorporating TPWDiff‑CB for all‑weather precipitable water retrieval and RePPIC‑Net for real‑time precipitation phase and intensity monitoring and forecasting. By further incorporating the DaYu-Nowcast model, RePPIC‑Net enables nowcasting of rain‑snow transitions up to three hours ahead, addressing a long‑standing gap in winter hazard warnings. Besides, to address the limitations of existing AI forecast models in representing mesoscale and convective-scale weather systems, DaYu-Nowcast adopts a hybrid deterministic-probabilistic forecasting framework and was developed to capture the spatiotemporal evolution of cloud systems. Within this framework, the DaYu-AHI model achieved high-accuracy short-term forecasts of brightness temperature cloud images from thermal infrared channels, with lead times up to 6 hours, a temporal resolution of 0.5 hours and a spatial resolution of 0.05°. 
Overall, DaYuRCPAS provides a unified, physicallyinformed framework that bridges radiative transfer simulation, cloud property retrieval, and precipitation nowcasting, facilitating the study of cloudradiation interactions and the diagnosis of extreme weather. In the future, DaYuRCPAS will continue to optimize the algorithms and expand the application scope of satellite data.
 

机器学习