Research map / Probabilistic and scientific modeling / Molecules and proteins
Probablistic Emulation of a Global Climate Model with Spherical DYffusion
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.
Spherical DYffusion combines the DYffusion dynamics-informed diffusion framework with SFNO backbones (plus time conditioning and MC dropout/stochastic depth) to produce probabilistic, stable 100-year global atmosphere emulations of FV3GFS.
Problem
Standard diffusion models are too slow for long autoregressive climate rollouts and DYffusion's UNet backbone mismatches spherical data; deterministic emulators like ACE cannot produce ensembles or capture the climate distribution.
Approach
Replace the UNet networks in DYffusion with spherical Fourier neural operators (SFNO) extended with time conditioning, and create inference stochasticity via MC dropout and stochastic depth, enabling ensemble climate emulation with under 3x the neural function evaluations of a deterministic SFNO.
Key results
- Average time-mean RMSE across 34 fields is about 49% above the noise floor, versus 110% for the next best method ACE.
- Ensemble-mean of 25 members reduces climate bias by 29.28% on average relative to single members.
- Runtime of 02:56 for a 10-year rollout versus 01:08 for ACE and 78:04 for physics-based FV3GFS (about 25x speedup), with 3(h-1) NFEs per step.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Spherical DYffusion instead of ACE | FV3GFS climate emulation · 10-year time-means, average across 34 predicted fields | RMSE relative to reference noise floor: 49.36% higher than the noise floor vs 110.47% | Helped |
| Average predictions across a 25-member ensemble | Spherical DYffusion · 10-year time-means, all variables | RMSE: 29.28% | Helped |
| Add MC dropout to ACE | ACE · 10-year time-mean climate biases, most fields | RMSE | Hurt |
| Replace SFNO backbones with the original UNet-based DYffusion architecture | DYffusion climate emulator · 10-year time-mean climate statistics | RMSE | Hurt |
| Add stochastic depth to the interpolator (ablation) | Spherical DYffusion interpolator · Interpolated timesteps 1 to 5 | spread-skill ratio: around 0.26 to 0.35 | Helped |
| Add stochastic depth at inference time (ablation) | ACE-STO · Climate emulation | climate bias | Hurt |
Surprising findings
- Stochastic ACE (ACE-STO via MC dropout) substantially underperforms its deterministic counterpart.
- Direct DYffusion with a UNet fails to match even the deterministic ACE baseline, even after ensemble averaging.
Limitations
- Only atmosphere emulated with fixed forcings and no greenhouse gas scenarios; no climate change generalization.
- Still roughly twice as slow as deterministic ACE at inference.
- Does not handle output-only variables such as precipitation and does not model initial-condition uncertainty.
Experimental setup
Datasets: FV3GFS 10-year initial-condition ensemble simulations (ACE dataset)
Benchmarks: FV3GFS 10-year validation simulation
Metrics: RMSE of 10-year time-mean bias MAE ensemble-mean RMSE spread-skill ratio CRPS ensemble variability
Baselines: ACE ACE-STO DYffusion FV3GFS 2x coarser (C48)
Builds on: DYffusion ACE Spherical Fourier Neural Operator Cold Diffusion Monte Carlo dropout
Closest papers by meaning
- DiffDA: a Diffusion model for weather-scale Data Assimilation (ICML 2024 · similarity 0.88)
- OmniCast: A Masked Latent Diffusion Model for Weather Forecasting Across Time Scales (NeurIPS 2025 · similarity 0.86)
- Tyche: One Step Flow for Efficient Probabilistic Weather Forecasting (NeurIPS 2026 · similarity 0.86)
- Causal Climate Emulation with Bayesian Filtering (NeurIPS 2025 · similarity 0.86)
- Validating Climate Models with Spherical Convolutional Wasserstein Distance (NeurIPS 2024 · similarity 0.85)
- Continuous Ensemble Weather Forecasting with Diffusion models (ICLR 2025 · similarity 0.85)
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.