WMO AI Webinars

Achieving skilful, long-lead precipitation nowcasting remains challenging, particularly when relying on a single observation source. Here we present NowAlpha, an operational radar-only precipitation nowcasting system that extends skilful prediction to 410 minutes. NowAlpha formulates nowcasting as latent-space diffusion video generation: sequences of radar reflectivity are encoded by a continuous visual tokenizer, and a diffusion model generates future latent trajectories that are decoded back to physically plausible reflectivity evolutions. We adopt NVIDIA’s Cosmos tokenizer with pretrained weights, and find that reusing the pretrained tokenizer improves forecast quality compared with training the tokenizer from scratch, indicating that large-scale, general-purpose visual tokenization transfers effectively to the weather domain. In midlatitude regimes, NowAlpha reduces spurious westward tendencies and more faithfully reproduces organized wintertime coastal convective bands and cyclonic precipitation structures. Finally, NowAlpha is validated in operations through a research-to-operations cycle, incorporating iterative feedback from professional forecasters to improve reliability for decision support.