Sign in

Salva Rühling Cachay

@salvarc7.bsky.social
918 followers 230 following 17 posts

ML PhD student at UC San Diego. Into AI for Science, especially climate & weather. salvarc.github.io

PostsRepliesMedia
Salva Rühling Cachay @salvarc7.bsky.social · 04/06/2026
Together with @dwatsonparris.bsky.social and @yuqirose.bsky.social!
010
Salva Rühling Cachay @salvarc7.bsky.social · 04/06/2026
*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...
arxiv.org
U-Cast: A Surprisingly Simple and Efficient Frontier Probabilistic AI Weather Forecaster
AI-based weather forecasting now rivals traditional physics-based ensembles, but state-of-the-art (SOTA) models rely on specialized architectures and massive computational budgets, creating a high bar...
120
Salva Rühling Cachay @salvarc7.bsky.social · 04/06/2026
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.
110
Salva Rühling Cachay @salvarc7.bsky.social · 04/06/2026
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.
110
Salva Rühling Cachay @salvarc7.bsky.social · 04/06/2026
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.
120
Salva Rühling Cachay @salvarc7.bsky.social · 04/06/2026
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.
120
Salva Rühling Cachay @salvarc7.bsky.social · 04/06/2026
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
110
Salva Rühling Cachay @salvarc7.bsky.social · 04/06/2026
🌎⚡ 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. 🧵
The Efficiency-Accuracy Pareto Frontier. We visualize
forecast skill (y-axis, % improvement over IFS ENS), inference la-
tency (x-axis), and training cost (bubble size). Our model (top-left)
achieves state-of-the-art performance while requiring an order of
magnitude less compute for training and/or inference compared
to leading baselines.
1113
Salva Rühling Cachay @salvarc7.bsky.social · 27/11/2025
A huge thank you to my brilliant collaborators @nvidia -Miika Aittala, @karstenkreis.bsky.social, Noah Brenowitz, Arash Vahdat & Morteza Mardani-and @yuqirose.bsky.social @ucsandiego.bsky.social 👇 See you next week in San Diego! Paper: arxiv.org/abs/2506.20024
000
Salva Rühling Cachay @salvarc7.bsky.social · 27/11/2025
𝗧𝗵𝗲 𝗥𝗲𝘀𝘂𝗹𝘁𝘀: 3) 𝗣𝗵𝘆𝘀𝗶𝗰𝗮𝗹 𝗥𝗲𝗮𝗹𝗶𝘀𝗺: ERDM matches the power spectra of operational physics-based models (IFS ENS), solving the "blurriness" problem common in AI weather models.
120
Salva Rühling Cachay @salvarc7.bsky.social · 27/11/2025
𝗧𝗵𝗲 𝗥𝗲𝘀𝘂𝗹𝘁𝘀: 1) Up to 🚀 𝟱𝟬% 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁 in probabilistic CRPS skill on Navier-Stokes dynamics, with strong calibratio; 2) Up to 🌍 10% improvement on ERA5 global weather forecasting (1.5° resolution) over autoregressive EDM baselines;
110
Salva Rühling Cachay @salvarc7.bsky.social · 27/11/2025
𝗢𝘂𝗿 𝗰𝗼𝗻𝘁𝗿𝗶𝗯𝘂𝘁𝗶𝗼𝗻: We unify rolling diffusion with the high-fidelity design of EDM by adapting EDMs core components—noise schedule, loss weighting, sampler—and supplementing it with a hybrid 3D backbone and a simple but effective initialization strategy.
100
Salva Rühling Cachay @salvarc7.bsky.social · 27/11/2025
The Gap: Existing methods struggle to balance fidelity and efficiency. ❌ Autoregressive models ignore temporal dependencies & may accumulate error ❌ Full "video" diffusion is computational- and data-inefficient We chose a third path: Rolling Diffusion—but gave it an upgrade..
100
Salva Rühling Cachay @salvarc7.bsky.social · 27/11/2025
🌍 Modeling chaos isn't just about predicting the next step—it's about modeling how uncertainty grows over time.🌪️ I’m thrilled to share Elucidated Rolling Diffusion Models (ERDM), accepted to #NeurIPS2025! We unify rolling diffusion with EDM for forecasting complex systems🧵👇
180
Reposted by Salva Rühling Cachay
David Rolnick @drolnick.bsky.social · 10/01/2025
Internship in our group at Mila in reinforcement learning + graphs for reducing energy use in buildings. More info and submit an application by Jan 13 here: forms.gle/TCChXnvSAHqz... Questions? Email donna.vakalis@mila.quebec with [intern!] in the subject line.
1144
Salva Rühling Cachay @salvarc7.bsky.social · 12/12/2024
Come talk to us tomorrow at Poster session 3: Thursday 11am-2pm at East Hall A-C #3905! (Or ping me if you'd like to chat outside of the poster session!)
0100
Reposted by Salva Rühling Cachay
Stephan Hoyer @handle.invalid · 21/11/2024
The new ACE2 climate emulator from Oliver Watt-Meyer et al has very compelling results, with results that look comparable to NeuralGCM. Congrats to the AI2 team! arxiv.org/abs/2411.112...
arxiv.org
ACE2: Accurately learning subseasonal to decadal atmospheric variability and forced responses
Existing machine learning models of weather variability are not formulated to enable assessment of their response to varying external boundary conditions such as sea surface temperature and greenhouse...
04511
Salva Rühling Cachay @salvarc7.bsky.social · 17/11/2024
Thanks!
030
Salva Rühling Cachay @salvarc7.bsky.social · 17/11/2024
🙌🏽🙋🏽‍♂️
110
Reposted by Salva Rühling Cachay
Jerry Lin 🇺🇦 @jlin404.com · 16/11/2024
I made a starter pack for those working in or adjacent to Machine Learning for Earth System Modeling! Apologies if I forgot anyone, and feel free to suggest people to add :) go.bsky.app/C5DQNCe
137122