We make an end-to-end AI weather model (Aardvark Weather) probabilistic with learned noise in the observation encoder and MC dropout in the processor, so a nested ensemble can split forecast uncertainty into observation-driven (aleatoric) and model-driven (epistemic) parts. The fine-tuned ensemble improves the deterministic mean by 4.2%, is calibrated through the medium range, and wins in CRPS at every lead time. Code and weights are released.</p>\n","updatedAt":"2026-09-01T10:03:20.476Z","author":{"_id":"67a5cc696122d1f9189f7893","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/ZjdKJ7k8RzqDbUzZQC7tS.jpeg","fullname":"Rodrigo Almeida","name":"rodrigoalmeida1994","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.9416351914405823},"editors":["rodrigoalmeida1994"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/ZjdKJ7k8RzqDbUzZQC7tS.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.30795","authors":[{"_id":"6a96895ecd6ebc484732ef6c","user":{"_id":"67a5cc696122d1f9189f7893","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/ZjdKJ7k8RzqDbUzZQC7tS.jpeg","isPro":false,"fullname":"Rodrigo Almeida","user":"rodrigoalmeida1994","type":"user","name":"rodrigoalmeida1994"},"name":"Rodrigo Almeida","status":"claimed_verified","statusLastChangedAt":"2026-09-01T08:29:18.165Z","hidden":false},{"_id":"6a96895ecd6ebc484732ef6d","name":"Noelia Otero","hidden":false},{"_id":"6a96895ecd6ebc484732ef6e","name":"Jost Arndt","hidden":false},{"_id":"6a96895ecd6ebc484732ef6f","name":"Simon Baur","hidden":false},{"_id":"6a96895ecd6ebc484732ef70","name":"Wojciech Samek","hidden":false},{"_id":"6a96895ecd6ebc484732ef71","name":"Jackie Ma","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/67a5cc696122d1f9189f7893/NOV-xYztzd9a-B0IkPpu8.png","https://cdn-uploads.huggingface.co/production/uploads/67a5cc696122d1f9189f7893/HzjgK5sgB2T6Lvq_5I9n5.png"],"publishedAt":"2026-08-31T00:00:00.000Z","submittedOnDailyAt":"2026-09-01T00:00:00.000Z","title":"Uncertainty-Aware End-to-End AI Weather Forecasting: Disentangling Observation and Model Contributions","submittedOnDailyBy":{"_id":"67a5cc696122d1f9189f7893","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/ZjdKJ7k8RzqDbUzZQC7tS.jpeg","isPro":false,"fullname":"Rodrigo Almeida","user":"rodrigoalmeida1994","type":"user","name":"rodrigoalmeida1994"},"summary":"End-to-end weather forecasting systems produce skillful global gridded and station forecasts directly from raw Earth observations, replacing the numerical weather prediction pipeline, including data assimilation, at a fraction of its cost. These systems are deterministic and issue no uncertainty. Here we render the Aardvark Weather model probabilistic by attaching one stochastic mechanism to each component: learned, input-dependent noise at the observation encoder, capturing aleatoric uncertainty inherited from the observing system, and Monte Carlo dropout in the processor, capturing epistemic uncertainty in the learned dynamics. The resulting nested ensemble attributes forecast spread to the two sources through a law-of-total-variance decomposition, cross-checked by withholding observation streams. Probabilistic finetuning significantly improves the mean forecast, by 4.2% on average across variables and lead times. The ensemble is calibrated against ERA5 through the medium range (spread-skill ratio 0.98), keeps station RMSE within 2.4% of the deterministic model while beating it in CRPS at every lead time, and trails the operational ECMWF ensemble. The encoder branch behaves as observation-driven uncertainty. Component-attributed uncertainty makes end-to-end forecasts more transparent, a step toward observation-driven digital twins of the atmosphere.","upvotes":0,"discussionId":"6a96895ecd6ebc484732ef72","projectPage":"https://gitlab.hhi.fraunhofer.de/ai-aml/uqe2e","ai_summary":"The Aardvark Weather model is made probabilistic by adding learned observation noise and Monte Carlo dropout to its encoder and processor, yielding a calibrated nested ensemble that improves mean forecasts and separates aleatoric and epistemic uncertainty.","ai_keywords":["Monte Carlo dropout","nested ensemble","law-of-total-variance decomposition","aleatoric uncertainty","epistemic uncertainty","probabilistic finetuning","observation encoder"],"ai_summary_model":"thinkingmachines/Inkling-Small","organization":{"_id":"6a6b35a681d2ba06fb45e26b","name":"FraunhoferHHI","fullname":"Fraunhofer-Institut für Nachrichtentechnik, Heinrich-Hertz-Institut, HHI","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/668d33dbaf57d6e4b5658ca9/PqXQqD5jUZLmLHfpUUW6W.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[],"acceptLanguages":["en"],"organization":{"_id":"6a6b35a681d2ba06fb45e26b","name":"FraunhoferHHI","fullname":"Fraunhofer-Institut für Nachrichtentechnik, Heinrich-Hertz-Institut, HHI","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/668d33dbaf57d6e4b5658ca9/PqXQqD5jUZLmLHfpUUW6W.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.30795.md","query":{}}">
Uncertainty-Aware End-to-End AI Weather Forecasting: Disentangling Observation and Model Contributions
Abstract
The Aardvark Weather model is made probabilistic by adding learned observation noise and Monte Carlo dropout to its encoder and processor, yielding a calibrated nested ensemble that improves mean forecasts and separates aleatoric and epistemic uncertainty.
End-to-end weather forecasting systems produce skillful global gridded and station forecasts directly from raw Earth observations, replacing the numerical weather prediction pipeline, including data assimilation, at a fraction of its cost. These systems are deterministic and issue no uncertainty. Here we render the Aardvark Weather model probabilistic by attaching one stochastic mechanism to each component: learned, input-dependent noise at the observation encoder, capturing aleatoric uncertainty inherited from the observing system, and Monte Carlo dropout in the processor, capturing epistemic uncertainty in the learned dynamics. The resulting nested ensemble attributes forecast spread to the two sources through a law-of-total-variance decomposition, cross-checked by withholding observation streams. Probabilistic finetuning significantly improves the mean forecast, by 4.2% on average across variables and lead times. The ensemble is calibrated against ERA5 through the medium range (spread-skill ratio 0.98), keeps station RMSE within 2.4% of the deterministic model while beating it in CRPS at every lead time, and trails the operational ECMWF ensemble. The encoder branch behaves as observation-driven uncertainty. Component-attributed uncertainty makes end-to-end forecasts more transparent, a step toward observation-driven digital twins of the atmosphere.
Community
We make an end-to-end AI weather model (Aardvark Weather) probabilistic with learned noise in the observation encoder and MC dropout in the processor, so a nested ensemble can split forecast uncertainty into observation-driven (aleatoric) and model-driven (epistemic) parts. The fine-tuned ensemble improves the deterministic mean by 4.2%, is calibrated through the medium range, and wins in CRPS at every lead time. Code and weights are released.
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2608.30795 in a model README.md to link it from this page.
Cite arxiv.org/abs/2608.30795 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.