Combining physical process and deep learning models for reliable precipitation nowcasts

Deep learning has become increasingly popular in the field of precipitation nowcasting. However, single data-driven models often fail to adhere physical laws and can suffer from errors, such as blur, dissipation, and intensity inaccuracies. Moreover, the opaque nature of deep learning hinders our understanding of the precipitation process. In this study, we propose a deep learning model that combines more accurate physical laws and the multiscale characteristics of weather systems. Specifically, our model incorporates the advection-diffusion equation and Burger’s equation to better capture the growth, decay, and motion fields of mesoscale precipitation. By utilizing radar observations from the USA, our deep learning model produces precipitation nowcasts that are not only physically realistic but also provide consistent multiscale patterns for different types of extreme precipitation. Further evaluation reveals that incorporating physical priors enhances the ability of the deep learning model to understand the evolution of precipitation and generate reliable nowcasts. By bridging the gap between physical principles and data-driven learning, our model represents a significant advancement in extreme precipitation nowcasting.