Research map / Image generation / Fast diffusion sampling
DiffDA: a Diffusion model for weather-scale Data Assimilation
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 DiffDA, a denoising diffusion model that adapts the pretrained GraphCast forecaster as its backbone to assimilate sparse column observations into 0.25° global atmospheric states, with observations enforced at inference via soft-masked inpainting.
Problem
Data assimilation (producing accurate initial conditions from predicted states and sparse observations) is computationally expensive and tightly coupled to numerical forecast systems, and prior diffusion-based DA work has not reached ERA5-scale resolution.
Approach
Condition a diffusion model on the GraphCast forecast by reusing GraphCast's input/output shape (two time steps in, next step out), train it on predicted-state conditioning only, and inject sparse observations at inference time via a softbled mask and kriging-interpolated observations (RePaint-style resampling) rather than training with observations.
Key results
- With 4,000 observed columns (<0.4% of data) the single-step assimilated data beats the 48-hour forecast RMSE for z500, t850 and t2m.
- With 10,000 observed columns, 48-hour forecasts from assimilated inputs have errors below GraphCast 72-hour forecasts, implying lead-time loss under 24 hours.
- Inference on a single A100 produces an assimilated state in around 15 minutes at 0.25° resolution with 13 levels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase soft-mask Gaussian kernel standard deviation from 0.5 to larger values (ablation) | DiffDA with GraphCast backbone · Single-step assimilation and forecasts from assimilated data | RMSE: 94.6 -> 90.1 | Helped |
| Increase soft-mask Gaussian kernel standard deviation into the 2.0–3.0 range (ablation) | DiffDA with GraphCast backbone · Single-step assimilation and forecasts from assimilated data | RMSE: 200 -> 209 | Mixed |
| Use 40000 observed columns instead of 1000 | DiffDA with GraphCast backbone · Single-step assimilation | RMSE | Helped |
| Use DiffDA assimilation rather than interpolated observations | Autoregressive data assimilation with non-fixed observation locations · z500 and t850 | RMSE | Helped |
| Use DiffDA assimilation rather than interpolated observations | Autoregressive data assimilation with non-fixed observation locations · t2m | RMSE | Hurt |
| Use 48-hour forecast inputs for DiffDA assimilation rather than ERA5 inputs | DiffDA forecast evaluation · 10,000 observed columns; z500, t850, and t2m | RMSE | Helped |
Surprising findings
- Hard-mask conditioning performed poorly; a softbled mask with kriging-interpolated observations was needed for the observation conditioning to take effect.
Limitations
- Relies on simulated observations taken from ERA5 with no quality control of real observations.
- Supports only point observations at a single time step; does not handle satellite imagery or radar.
- Autoregressive cycles show error growth, especially for t2m, which exceeds interpolation errors after several iterations.
Experimental setup
Datasets: WeatherBench2 ERA5
Benchmarks: ERA5 (WeatherBench2)
Metrics: RMSE
Baselines: GraphCast 48h forecast interpolated observations ERA5 initial conditions
Builds on: DDPM RePaint GraphCast universal kriging
Closest papers by meaning
- FlowDAS: A Stochastic Interpolant-based Framework for Data Assimilation (NeurIPS 2025 · similarity 0.88)
- Probablistic Emulation of a Global Climate Model with Spherical DYffusion (NeurIPS 2024 · similarity 0.88)
- On conditional diffusion models for PDE simulations (NeurIPS 2024 · similarity 0.88)
- OmniCast: A Masked Latent Diffusion Model for Weather Forecasting Across Time Scales (NeurIPS 2025 · similarity 0.87)
- DAWP: A framework for global observation forecasting via Data Assimilation and Weather Prediction in satellite observation space (NeurIPS 2025 · similarity 0.86)
- Towards a Self-contained Data-driven Global Weather Forecasting Framework (ICML 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.