ML PhD student at UC San Diego. Into AI for Science, especially climate & weather.
https://salvarc.github.io/
Salva Rühling Cachay
Loading...
The result: U-Cast matches GenCast and outperforms IFS ENS on WeatherBench 2 at 1.5°, with >10× less training compute than leading CRPS models and >10× faster inference than diffusion.
Salva Rühling Cachay
Is that complexity actually necessary? Surprisingly, no.
U-Cast is a fairly standard U-Net made probabilistic with Monte Carlo Dropout + a two-stage curriculum: pre-train deterministically on MAE, then briefly fine-tune on CRPS.
The point isn't that specialized architectures are obsolete (we're clear about U-Cast's limitations). It's that a strong, efficient model + recipe lets far more people actually train and fine-tune these models, not just consume them.
Together with @dwatsonparris.bsky.social and
@yuqirose.bsky.social!
A single forward pass per member keeps sampling fast. Decoupling the learning of dynamics from the learning of uncertainty keeps training cheap: the probabilistic stage is only ~15% of the budget.
*Fine-tuning takes ~1 day (1.2 H200-GPU-days) on a pre-trained deterministic backbone; add ~7 H200-days to train that backbone from scratch (~8 total).
📄 arxiv.org/abs/2604.09041
💻 github.com/Rose-STL-Lab...
Salva Rühling Cachay
Today's frontier ensemble forecasters are caught in a cost trap:
• Diffusion (GenCast): slow to sample, dozens of forward passes per forecast
• CRPS-trained (AIFS-CRPS, FGN): slow to train, a full ensemble every gradient step
• Most use bespoke nets (graph, spherical,...) that compound complexity
Salva Rühling Cachay
🌎⚡ A frontier (1.5°, 15-day) ensemble weather forecast in ~3 seconds, from a probabilistic model you can train in ~1 day on a single H200 GPU?*
Meet U-Cast, our new #ICML2026 paper. 🧵
