Research map / Image generation / Text-to-image diffusion
Align-DA: Align Score-based Atmospheric Data Assimilation with Multiple Preferences
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.
Align-DA formulates atmospheric data assimilation as a latent score-based generative process and aligns the background-conditioned prior with DPO using assimilation-accuracy, 48-hour forecast, and geostrophic-balance rewards.
Problem
Atmospheric data assimilation is ill-posed due to sparse observations, and traditional background-error priors need empirical tuning and lack explicit physical or forecast-skill constraints.
Approach
Model the background-conditioned prior p(z|z_b) with a latent score-based diffusion model, then align it with Diffusion-DPO using multi-reward preference pairs (assimilation accuracy, 48h forecast skill, geostrophic balance) as a soft constraint, so observation guidance (ObsFree, Repaint, DPS) is applied on top of the aligned prior.
Key results
- Align-Repaint-M reduces analysis MSE by 4.12% relative to Repaint at 1% observation density (0.0623 to 0.0598).
- Align-ObsFree-M reduces z500 WRMSE by 7.94% relative to ObsFree (104.55 to 96.86) at 1% observations.
- Single physical-reward alignment (-P) yields Geo-Score gains (e.g., 93.85 to 95.00) but slightly degrades overall assimilation accuracy and forecast skill.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-reward alignment | ObsFree · ERA5 1% observations (analysis) | MSE: 0.0643 -> 0.0611; (-5.44%) | Helped |
| add multi-reward alignment | Repaint · ERA5 1% observations (analysis) | MSE: 0.0623 -> 0.0598; (-4.12%) | Helped |
| add multi-reward alignment | DPS · ERA5 1% observations (analysis) | z500 WRMSE: 93.4654 -> 88.4235; (-5.70%) | Helped |
| add multi-reward alignment | ObsFree · ERA5 1% observations (48-hour forecast) | MSE: 0.1146 -> 0.1080; (-6.12%) | Helped |
| add multi-reward alignment | Repaint · ERA5 1% observations (48-hour forecast) | MSE: 0.1123 -> 0.1068; (-5.12%) | Helped |
| add multi-reward alignment | DPS · ERA5 1% observations (48-hour forecast) | z500 WRMSE: 193.4167 -> 189.6935; (-1.96%) | Helped |
| add single physical reward alignment (ablation) | ObsFree · ERA5 1% observations (analysis) | MSE: 0.0643 -> 0.0645; (+0.24%) | Hurt |
| add single physical reward alignment (ablation) | Repaint · ERA5 1% observations (analysis) | MSE: 0.0623 -> 0.0624; (+0.20%) | Hurt |
| add single physical reward alignment (ablation) | DPS · ERA5 1% observations (analysis) | z500 WRMSE: 93.4654 -> 92.4301; (-1.12%) | Helped |
| add single physical reward alignment (ablation) | DPS · ERA5 1% observations (48-hour forecast) | MSE: 0.0969 -> 0.0972; (+0.29%) | Hurt |
| add multi-reward alignment (ablation) | Repaint · ERA5 10% observations (analysis) | z500 WRMSE: -6.42% | Helped |
| add multi-reward alignment (ablation) | DPS · ERA5 10% observations (analysis) | z500 WRMSE: -7.71% | Helped |
Surprising findings
- Physical-reward-only alignment (-P) slightly worsens overall MSE and forecast skill, while multi-reward alignment improves them.
- Alignment benefits shrink as observation guidance becomes stronger; the observation-free baseline gains most and DPS gains least.
Limitations
- Physical reward is limited to geostrophic balance at 500 hPa over a few variables, which may not generalize to full operational DA.
- The alignment is offline and depends on the quality and diversity of the fixed preference dataset.
- Single physical-reward alignment degrades overall assimilation accuracy and forecast skill.
Experimental setup
Datasets: ERA5 GDAS prepbufr
Benchmarks: ERA5 2019 6-hourly assimilation cycles 48-hour forecast skill on ERA5
Metrics: MSE MAE WRMSE
Baselines: ObsFree Repaint DPS Align-DPS-P Align-Repaint-P
Builds on: Diffusion-DPO DPS (diffusion posterior sampling) Repaint DiffDA Bradley-Terry preference model
Closest papers by meaning
- DiffDA: a Diffusion model for weather-scale Data Assimilation (ICML 2024 · similarity 0.86)
- FlowDAS: A Stochastic Interpolant-based Framework for Data Assimilation (NeurIPS 2025 · similarity 0.84)
- DAWP: A framework for global observation forecasting via Data Assimilation and Weather Prediction in satellite observation space (NeurIPS 2025 · similarity 0.84)
- Latent-EnSF: A Latent Ensemble Score Filter for High-Dimensional Data Assimilation with Sparse Observation Data (ICLR 2025 · similarity 0.84)
- Align-RAG: Alignment Is All You Need for TSFM In-Context Learning (NeurIPS 2026 · similarity 0.84)
- Scaling Storm-Resolving Atmospheric AI Simulation to the Entire Planet (NeurIPS 2026 · similarity 0.83)
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.