TL;DR
WeatherNext 2 generates 100+ weather scenarios in under 60 seconds on single TPU, outperforming previous model on 99.9% of variables.
Key Points
- Generates hundreds of weather scenarios in <1 minute using single TPU vs. hours on traditional supercomputers
- Achieves 99.9% improvement over WeatherNext Gen across all variables and lead times (0-15 days)
- Implements novel Functional Generative Network (FGN) architecture injecting noise directly into model for physically realistic predictions
- Available now in Earth Engine, BigQuery; Vertex AI early access program launched for custom inference
Why It Matters
This represents a significant engineering breakthrough in probabilistic weather modeling—moving from deterministic to ensemble forecasting at scale. For developers and researchers, the sub-minute inference on commodity hardware (single TPU) vs. supercomputers fundamentally changes accessibility and deployment economics for operational weather prediction systems. The FGN approach of learning marginals to predict joints has implications beyond meteorology for any complex systems requiring coherent multivariate predictions.
Source: blog.google