# arXiv paper presents Sonny-ENS for probabilistic weather forecasting on one GPU [Canonical reading](https://escapemeridian.com/readings/arxiv-paper-presents-sonny-ens-for-probabilistic-weather-ff2e92a2) Minjong Cheon presents Sonny-ENS, an ensemble of three machine-learning networks trained in 2.7 A40 GPU-days. The paper reports lower continuous ranked probability score (CRPS) than the European Centre for Medium-Range Weather Forecasts ensemble on 96% of 570 targets for 2022, with an 8.3% lower mean CRPS at lead times of 1–10 days. It also reports outperforming GenCast on 69% of the targets. A 51-member, 15-day forecast took 69 seconds on one GPU. These results address two questions about machine-learning ensembles: whether their skill requires extensive training compute, and how their forecast spread reflects atmospheric flow. The paper reports that dropout injects uncertainty mainly in the tropics, while spread grows where the atmosphere is baroclinically unstable. The reported results are limited to the specified targets and year; the abstract does not establish performance in operational forecasting. What performance would Sonny-ENS achieve across other years and in operational forecasts? Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv physics.ao-ph · Article Original source date: Nov 1, 2026 (2026-11-01). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-07T09:30:53.954Z. First observed: 2026-11-03T08:51:17.830Z. Retrieved: 2026-11-03T19:53:14.185Z. - [Sonny-ENS: demystifying machine-learning ensemble weather forecasting](https://arxiv.org/abs/2611.01614) ## Limits The reported comparison is for 2022 and 570 targets; the abstract does not report performance across other years. The abstract does not establish how Sonny-ENS performs in operational forecasting. ## Open question What performance would Sonny-ENS achieve across other years and in operational forecasts? ## Source-stated Cheon presents Sonny-ENS as an ensemble of three networks, trained using 2.7 A40 GPU-days. > an ensemble of three 68.9-million-parameter networks with Monte Carlo dropout, trained on the continuous ranked probability score (CRPS) in 2.7 A40 GPU-days. - [Supporting source](https://arxiv.org/abs/2611.01614) ## Source-stated For 2022, the paper reports lower CRPS than the European Centre for Medium-Range Weather Forecasts ensemble on 96% of 570 targets, and 8.3% lower mean CRPS at lead times of 1–10 days. > On 2022, it has a lower CRPS than the European Centre for Medium-Range Weather Forecasts ensemble (ENS) on 96% of 570 targets and an 8.3% lower mean CRPS at lead times of 1-10 days. - [Supporting source](https://arxiv.org/abs/2611.01614) ## Source-stated The paper reports that Sonny-ENS outperforms the operational GenCast on 69% of the targets. > It outperforms the operational GenCast on 69% of the targets - [Supporting source](https://arxiv.org/abs/2611.01614) ## Source-stated The abstract reports that a 51-member, 15-day forecast takes 69 seconds on one GPU. > a 51-member, 15-day forecast takes 69 s on one GPU. - [Supporting source](https://arxiv.org/abs/2611.01614) ## Source-stated The paper reports that dropout injects uncertainty mainly in the tropics, while spread grows where the atmosphere is baroclinically unstable. > dropout injects uncertainty mainly in the tropics, whereas the spread grows where the atmosphere is baroclinically unstable - [Supporting source](https://arxiv.org/abs/2611.01614) ## Source and contributor context Cheon, Minjong · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Graphics processing unit A chip that performs many calculations at once; the main hardware used to train and run AI models. - [Definition reference](https://en.wikipedia.org/wiki/Graphics_processing_unit) ## Machine learning Methods that let computers learn patterns from examples instead of following rules written by hand. - [Definition reference](https://developers.google.com/machine-learning/glossary#machine-learning) ## AI compute The computing resources, meaning chips, data centers and electricity, used to train or run AI models. - [Definition reference](https://developers.google.com/machine-learning/glossary#compute) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 94f468df-9cef-4b83-86ce-f1734b484114. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-07T09:30:53.954Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 7a69bc701fa64dd92c464ac0917587ab8ebccf16fee14f4be304b52fe46abc64. Source text hash: aac97a659919d553b4ca96d974842417481cf05b137ce628b6e20b6ea4822782. ## Revision and source-check history ### Version 1 — 2026-11-07T09:30:53.954Z Minjong Cheon presents Sonny-ENS, an ensemble of three machine-learning networks trained in 2.7 A40 GPU-days. The paper reports lower continuous ranked probability score (CRPS) than the European Centre for Medium-Range Weather Forecasts ensemble on 96% of 570 targets for 2022, with an 8.3% lower mean CRPS at lead times of 1–10 days. It also reports outperforming GenCast on 69% of the targets. A 51-member, 15-day forecast took 69 seconds on one GPU. These results address two questions about machine-learning ensembles: whether their skill requires extensive training compute, and how their forecast spread reflects atmospheric flow. The paper reports that dropout injects uncertainty mainly in the tropics, while spread grows where the atmosphere is baroclinically unstable. The reported results are limited to the specified targets and year; the abstract does not establish performance in operational forecasting. What performance would Sonny-ENS achieve across other years and in operational forecasts? ## Original source arXiv physics.ao-ph · Article Original source date: Nov 1, 2026 (2026-11-01). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-07T09:30:53.954Z. First observed: 2026-11-03T08:51:17.830Z. Retrieved: 2026-11-03T19:53:14.185Z. - [Sonny-ENS: demystifying machine-learning ensemble weather forecasting](https://arxiv.org/abs/2611.01614) ## Limits The reported comparison is for 2022 and 570 targets; the abstract does not report performance across other years. The abstract does not establish how Sonny-ENS performs in operational forecasting. ## Open question What performance would Sonny-ENS achieve across other years and in operational forecasts? ## Source-stated Cheon presents Sonny-ENS as an ensemble of three networks, trained using 2.7 A40 GPU-days. > an ensemble of three 68.9-million-parameter networks with Monte Carlo dropout, trained on the continuous ranked probability score (CRPS) in 2.7 A40 GPU-days. - [Supporting source](https://arxiv.org/abs/2611.01614) ## Source-stated For 2022, the paper reports lower CRPS than the European Centre for Medium-Range Weather Forecasts ensemble on 96% of 570 targets, and 8.3% lower mean CRPS at lead times of 1–10 days. > On 2022, it has a lower CRPS than the European Centre for Medium-Range Weather Forecasts ensemble (ENS) on 96% of 570 targets and an 8.3% lower mean CRPS at lead times of 1-10 days. - [Supporting source](https://arxiv.org/abs/2611.01614) ## Source-stated The paper reports that Sonny-ENS outperforms the operational GenCast on 69% of the targets. > It outperforms the operational GenCast on 69% of the targets - [Supporting source](https://arxiv.org/abs/2611.01614) ## Source-stated The abstract reports that a 51-member, 15-day forecast takes 69 seconds on one GPU. > a 51-member, 15-day forecast takes 69 s on one GPU. - [Supporting source](https://arxiv.org/abs/2611.01614) ## Source-stated The paper reports that dropout injects uncertainty mainly in the tropics, while spread grows where the atmosphere is baroclinically unstable. > dropout injects uncertainty mainly in the tropics, whereas the spread grows where the atmosphere is baroclinically unstable - [Supporting source](https://arxiv.org/abs/2611.01614) ## Source and contributor context Cheon, Minjong · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Graphics processing unit A chip that performs many calculations at once; the main hardware used to train and run AI models. - [Definition reference](https://en.wikipedia.org/wiki/Graphics_processing_unit) ## Machine learning Methods that let computers learn patterns from examples instead of following rules written by hand. - [Definition reference](https://developers.google.com/machine-learning/glossary#machine-learning) ## AI compute The computing resources, meaning chips, data centers and electricity, used to train or run AI models. - [Definition reference](https://developers.google.com/machine-learning/glossary#compute) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 94f468df-9cef-4b83-86ce-f1734b484114. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-07T09:30:53.954Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 7a69bc701fa64dd92c464ac0917587ab8ebccf16fee14f4be304b52fe46abc64. Source text hash: aac97a659919d553b4ca96d974842417481cf05b137ce628b6e20b6ea4822782. ### uncertain — 2026-11-07T09:30:53.954Z — anthropic/claude-opus-5.5 Publication: 94f468df-9cef-4b83-86ce-f1734b484114 attribution — pass: The draft correctly attributes Sonny-ENS and its reported results to Minjong Cheon, the listed author of the arXiv paper. It frames the results as paper reports rather than established facts. > Authors:Minjong Cheon summary — uncertain: Central terms are left unexplained for a general reader: ensemble, CRPS (and that lower is better), GPU-days, Monte Carlo dropout, and baroclinically unstable. The draft also drops two pieces of context that carry the author's argument. First, the abstract contrasts 2.7 GPU-days with hundreds to thousands of accelerator-days. Second, the paper concludes that skillful probabilistic forecasts are within reach of a single GPU server. Smallest repair: add tooltips or a plain gloss for these terms, and restore the compute contrast and the closing thesis. The summary also drops the word 'operational' before GenCast. > whether their skill requires hundreds to thousands of accelerator-days of training date — pass: The originalPublished date of 2026-11-01 matches the v1 submission date shown on the arXiv page. The 2026-11-03 feed date is correctly not used. > \[Submitted on 1 Nov 2026\] support — pass: Every numeric claim matches the abstract: 2.7 A40 GPU-days, 96% of 570 targets, 8.3% lower mean CRPS, 69% versus GenCast, and 69 s for a 51-member, 15-day forecast. The dropout and spread finding also matches. The limitations and the question about other years and operational use are appropriately conditional and do not overstate the evidence. > It outperforms the operational GenCast on 69% of the targets, and a 51-member, 15-day forecast takes 69 s on one GPU. media — pass: The selected media is the verified direct link to the original arXiv abstract page, which is the most direct way to reach the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/ff2e92a2e2a3c330c2d5a81ab9b4ac2e7655f9fb7a416a7dccce1fea7b358afb/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/arxiv-paper-presents-sonny-ens-for-probabilistic-weather-ff2e92a2.json)