WIND: Weather Inverse Diffusion for Zero-Shot Atmospheric Modeling

Deep learning has revolutionized weather and climate modeling, yet the current landscape remains fragmented: highly specialized models are typically trained individually for distinct tasks. To unify this landscape, we introduce WIND, a single pre-trained foundation model capable of replacing specialized baselines across a vast array of tasks. Crucially, in contrast to previous atmospheric foundation models, we achieve this without any task-specific fine-tuning. To learn a robust, task-agnostic prior of the atmosphere, we pre-train WIND with a self-supervised video reconstruction objective, utilizing an unconditional video diffusion model to iteratively reconstruct atmospheric dynamics from a noisy state. At inference, we frame diverse domain-specific problems strictly as inverse problems and solve them via posterior sampling. This unified approach allows us to tackle highly relevant weather and climate problems, including probabilistic forecasting, spatial and temporal downscaling, sparse reconstruction and enforcing conservation laws purely with our pre-trained model. We further demonstrate the model’s capacity to generate physically consistent counterfactual storylines of extreme weather events under global warming scenarios. By combining generative video modeling with inverse problem solving, WIND offers a computationally efficient paradigm shift in AI-based atmospheric modeling.

💡 Research Summary

The paper introduces WIND (Weather Inverse Diffusion), a single foundation model for atmospheric science that eliminates the need for task‑specific fine‑tuning. The authors treat atmospheric data as short video sequences and pre‑train an unconditional diffusion model using a “diffusion‑forcing” self‑supervised objective: each frame in a sequence receives an independently sampled noise level, and the network learns to reconstruct the clean sequence from any mixture of noisy and clean frames. Importantly, the model does not receive the noise levels as input, forcing it to infer uncertainty directly from the corrupted data. This training scheme enables the model to accept clean context frames from previous windows without distribution shift, allowing arbitrarily long roll‑outs and stable generation.

At inference time, every downstream task (probabilistic forecasting, spatial and temporal down‑scaling, sparse reconstruction, enforcement of physical conservation laws, and generation of counterfactual climate storylines) is cast as an inverse problem of the form Y = A(X) + η, where A is a task‑specific forward operator and η is Gaussian observation noise. The posterior score ∇ log p(Z|Y) is decomposed into a prior score supplied by the pretrained diffusion model and a likelihood score estimated via Moment Matching Posterior Sampling (MMPS). MMPS provides a principled way to incorporate the observation constraints while preserving the multivariate structure of the diffusion prior, avoiding the pitfalls of point‑estimate posterior samplers.

The authors evaluate WIND on ERA5 reanalysis data at 1.5° resolution (70 variables, sequence length = 5, 6‑hour stride). For probabilistic forecasting, they generate 10‑member ensembles for 14‑day lead times from 24 initial conditions in 2021. Using Continuous Ranked Probability Score (CRPS) and Spread‑Skill Ratio (SSR), WIND consistently outperforms a strong autoregressive diffusion baseline (AR‑UVIT) after the first few days and converges to well‑calibrated ensembles (SSR ≈ 1) within two weeks. A 20‑year unconstrained rollout demonstrates that AR‑UVIT develops non‑physical spikes across all variables, whereas WIND maintains physical consistency across the full spectral range.

In spatial down‑scaling (4× resolution increase), WIND is compared against a Fourier Neural Operator and a UVIT‑based model on a full year of 2021 data. WIND achieves higher PSNR and SSIM, indicating superior recovery of fine‑scale structures that are critical for impact modeling. The paper also shows that enforcing a dry‑air‑mass conservation constraint does not degrade forecast skill, confirming that physical guidance can be added without sacrificing performance. Finally, counterfactual experiments illustrate how WIND can generate physically plausible climate scenarios under a +2 °C warming, altering mean precipitation and temperature fields while respecting dynamical consistency.

Limitations include the use of relatively coarse 1.5° data (the model has not been tested at operational 0.25° resolution) and the absence of explicit noise‑level conditioning, which may limit reconstruction quality under extreme corruption. Future work could explore multi‑scale architectures, high‑resolution fine‑tuning, and integration of more complex physical constraints such as energy conservation or convection‑precipitation coupling.

Overall, WIND demonstrates that a single, self‑supervised diffusion prior can serve as a universal atmospheric model. By framing downstream tasks as inverse problems and solving them with posterior sampling, the approach delivers a versatile, zero‑shot solution that rivals or exceeds specialized models across a broad spectrum of weather and climate applications, heralding a paradigm shift toward foundation‑model‑centric atmospheric science.