Research map / Architectures and efficiency / Pruning and efficient tuning
U-Cast: A Surprisingly Simple and Efficient Frontier Probabilistic AI Weather Forecaster
Salva Rühling Cachay, Duncan Watson-Parris, Rose Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add deterministic MAE pre-training before CRPS fine-tuning (ablation) | U-Cast · t850 12h validation | CRPS: 0.225 -> 0.218 | Improved |
| add deterministic MAE pre-training before CRPS fine-tuning (ablation) | U-Cast · short-range surface and upper-air variables | CRPS: 3-5% lower CRPS | Improved |
| add deterministic MAE pre-training before CRPS fine-tuning (ablation) | U-Cast · stratospheric variables | CRPS: 5-15% lower CRPS | Improved |
| add deterministic MAE pre-training before CRPS fine-tuning (ablation) | U-Cast · 2t long range | CRPS: from-scratch occasionally recovers long-range scores | Mixed |
| use Muon instead of AdamW (ablation) | U-Cast · z500 1-3 day | CRPS: up to 15% lower CRPS | Improved |
| use MC Dropout instead of adaLN noise injection (ablation) | U-Cast · 1-7 day forecasts | CRPS: adaLN significantly degrades CRPS | Improved |
| use adaLN noise injection instead of MC Dropout (ablation) | U-Cast · spread-skill ratio | CRPS: better dispersion | Improved |
| vary dropout rate to 5% or 15% (ablation) | U-Cast · CRPS | CRPS: within +-1% CRPS | No clear effect |
| lower dropout rate to 5% (ablation) | U-Cast · spread-skill ratio | CRPS: SSR closest to 1 | Improved |
| increase training ensemble size from M=2 to M=4 (ablation) | U-Cast · CRPS | CRPS: marginal CRPS gains at doubled cost | No clear effect |
| add deep ensembling (K=4) | U-Cast · z500 1d | CRPS: 20.3 -> 19.6 | Improved |
| add deep ensembling (K=4) | U-Cast · z500 10d | CRPS: 256 -> 253 | Improved |
Closest papers by meaning
- Generalizing Weather Forecast to Fine-grained Temporal Scales via Physics-AI Hybrid Modeling (NeurIPS 2024 · similarity 0.86)
- Tyche: One Step Flow for Efficient Probabilistic Weather Forecasting (NeurIPS 2026 · similarity 0.86)
- Probabilistic Weather Forecasting with Hierarchical Graph Neural Networks (NeurIPS 2024 · similarity 0.85)
- EWMoE: An Effective Model for Global Weather Forecasting with Mixture-of-Experts (AAAI 2025 · similarity 0.84)
- Scaling transformer neural networks for skillful and reliable medium-range weather forecasting (NeurIPS 2024 · similarity 0.84)
- ForecastBench: A Dynamic Benchmark of AI Forecasting Capabilities (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
