ICML 2026 paper SDIR achieves record accuracy in precipitation nowcasting

Precipitation nowcasting has long faced a fundamental trade-off: regression models produce over-smoothed predictions that blur convective details and violate turbulence power laws, while diffusion models generate realistic but physically unanchored hallucinations. A new ICML 2026 paper from Yunlong Zhou and colleagues introduces Spectral-Decoupled Iterative Refinement (SDIR), a deterministic framework that reformulates nowcasting as progressive frequency-decoupled refinement. SDIR first extracts a stable low-frequency synoptic skeleton, then iteratively refines high-frequency textures under physical constraints, eliminating both blurring and hallucinations.

The framework features a dual-path architecture: the Synoptic Frequency-Guided Former (SFG-Former) with Scale-Adaptive Transformers captures global structure, while the Fourier Residual Refiner (FR-Refiner) uses Scale-Conditioned Fourier Neural Operators to model fine residuals. A key innovation is the Physically Consistent Power Spectral Density (PCPSD) loss with dynamic masking, which enforces a turbulence-consistent spectral distribution. Tested on three benchmarks, SDIR significantly outperforms state-of-the-art methods in spatial accuracy while achieving spectral fidelity competitive with diffusion-based approaches. This breakthrough promises reliable high-resolution operational nowcasting for weather-dependent industries and disaster response systems.