Research map / Time series / Causal and generative time series
Continuous Ensemble Weather Forecasting with Diffusion models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Continuous Ensemble Forecasting, a lead-time-conditioned diffusion model that samples temporally consistent ensemble weather trajectories in parallel, combined with autoregressive rollouts (ARCI) for hourly-resolution 10-day forecasts.
Problem
Diffusion-based probabilistic weather forecasting is computationally expensive and accumulates error over many autoregressive rollout steps, limiting temporal resolution and parallelism.
Approach
Condition a diffusion denoiser on lead time and sample each ensemble member's trajectory by reusing (fixed or Ornstein-Uhlenbeck-correlated) driving noise across lead times, so all lead times are generated in one parallel single-step pass; combine with 24h autoregressive steps and continuous interpolation (ARCI).
Key results
- ARCI-24/6h achieves z500 RMSE 765.6 and CRPS 355.2 at 10 days, beating Graph-EFM (817.1 / 373.6) and AR-6h (811.8 / 391.9) at 6h resolution and matching AR-24h (750.6 / 335.2).
- At 1h resolution, ARCI-24/1h stays as skillful as 24h forecasts while AR-1h degrades strongly; ARCI beats DYffusion in all variables and lead times for t850.
- Sampling a 10-day 6h-resolution member takes 8 s with parallelized ARCI versus 32 s for AR-6h.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ARCI-24/6h instead of AR-6h | Global weather forecasting · z500, 5 days | RMSE: 602.3 -> 560.9 | Helped |
| Use ARCI-24/6h instead of AR-6h | Global weather forecasting · z500, 5 days | CRPS: 287.8 -> 256.7 | Helped |
| Use ARCI-24/6h instead of AR-24h | Global weather forecasting · z500, 5 days | RMSE: 544.2 -> 560.9 | Hurt |
| Use ARCI-24/6h instead of AR-24h | Global weather forecasting · z500, 5 days | CRPS: 242.7 -> 256.7 | Hurt |
| Use ARCI-24/6h instead of AR-6h | Global weather forecasting · t850, 5 days | RMSE: 2.72 -> 2.6 | Helped |
| Use ARCI-24/6h instead of AR-6h | Global weather forecasting · t850, 5 days | CRPS: 1.34 -> 1.27 | Helped |
| Use ARCI-24/6h instead of AR-24h | Global weather forecasting · t850, 5 days | RMSE: 2.55 -> 2.6 | Hurt |
| Use ARCI-24/6h instead of AR-24h | Global weather forecasting · t850, 5 days | CRPS: 1.24 -> 1.27 | Hurt |
| Use ARCI-24/6h instead of AR-6h | Global weather forecasting · z500, 10 days | RMSE: 811.8 -> 765.6 | Helped |
| Use ARCI-24/6h instead of AR-6h | Global weather forecasting · z500, 10 days | CRPS: 391.9 -> 355.2 | Helped |
| Use ARCI-24/6h instead of AR-24h | Global weather forecasting · z500, 10 days | RMSE: 750.6 -> 765.6 | Hurt |
| Use ARCI-24/6h instead of AR-24h | Global weather forecasting · z500, 10 days | CRPS: 335.2 -> 355.2 | Hurt |
Surprising findings
- Continuous-only CI-6h performs well at short lead times but struggles at long horizons, with SSR of 0.6-0.7 at 10 days.
- All diffusion-based models show SSR below 1, indicating systematic underdispersion.
Limitations
- Only evaluated at 5.625 degree resolution; scaling to higher spatial resolution not shown.
- Correlation between initial and future states weakens with lead time, limiting continuous forecasting at long horizons.
- Sampling still requires many sequential network evaluations per diffusion step, slower than latent-variable models.
Experimental setup
Datasets: ERA5 (WeatherBench 5.625 deg)
Benchmarks: WeatherBench ERA5 5.625 deg 2018 test
Metrics: RMSE CRPS SSR temporal difference
Baselines: Deterministic U-Net AR-6h AR-24h CI-6h Graph-EFM DYffusion AR-1h
Builds on: GenCast EDM (elucidating diffusion models) DYffusion Graph-EFM probability flow ODE
Closest papers by meaning
- Tyche: One Step Flow for Efficient Probabilistic Weather Forecasting (NeurIPS 2026 · similarity 0.88)
- Elucidated Rolling Diffusion Models for Probabilistic Forecasting of Complex Dynamics (NeurIPS 2025 · similarity 0.88)
- Auto-Regressive Moving Diffusion Models for Time Series Forecasting (AAAI 2025 · similarity 0.87)
- ABC: Any-Subset Autoregression via Non-Markovian Diffusion Bridges in Continuous Time and Space (NeurIPS 2026 · similarity 0.87)
- Probabilistic Weather Forecasting with Hierarchical Graph Neural Networks (NeurIPS 2024 · similarity 0.86)
- On conditional diffusion models for PDE simulations (NeurIPS 2024 · similarity 0.86)
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.