Toward High-Fidelity and Temporal-Consistency Generation for Radar Echo Perception via Differential Diffusion Model

Precise and timely radar echo extrapolation is challenging in short-term meteorological nowcasting, given the inherent uncertainty and nonlinearity of weather systems. The diffusion and denoising processes in diffusion models correspond with the intrinsic mechanisms of ‘disorder enhancement’ (entropy increase) and ‘ordered structure formation’ (entropy decrease) in atmospheric motion. Nevertheless, we find that directly applying diffusion models fails to capture the key dynamics. We argue that radar echo evolution is governed by two core processes: advection-dominated bulk displacement and local growth/decay-dominated intensity changes. We thus propose a Physics-Guided Decoupled Diffusion Model (PG-DDM) which injects physical priors through a dual-path design: an Advection Path that derives the instantaneous spatial state from a single frame, and a Growth/Decay Path that captures local processes via inter-frame differences. Our method leverages these complementary physical cues to decouple the learning of motion and change, producing high-fidelity extrapolation sequences. Code is available at https://github.com/azier33/DDM.