Almeida, Rodrigo, Otero, Noelia, Arndt, Jost, Baur, Simon, Samek, Wojciech, Ma, Jackie · Aug 31, 2026
arXiv authors separate uncertainty sources in AI weather forecasts
Markdown · JSON · Original source
The authors describe making Aardvark Weather probabilistic by adding separate sources of uncertainty to its observation encoder and processor. End-to-end systems forecast directly from raw Earth observations, replacing the numerical weather prediction pipeline, including data assimilation, but the abstract says they issue no uncertainty. Learned, input-dependent noise represents uncertainty inherited from observations; Monte Carlo dropout represents uncertainty in learned dynamics. A nested ensemble attributes forecast spread to these sources using a law-of-total-variance decomposition, cross-checked by withholding observation streams. The authors report that probabilistic improves the mean forecast by 11% on average across variables and lead times. They also report that the ensemble is close to calibrated against ERA5 through the medium range, is at least as accurate as the deterministic model at weather stations, and beats it in relative economic value at every lead . It remains considerably behind the operational ECMWF IFS ensemble. The authors describe component-attributed uncertainty as a way to make end-to-end forecasts more transparent and a step toward observation-driven digital twins of the atmosphere.
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
physics.ao-ph · Article
Source details
Original source date: Aug 31, 2026 (2026-08-31).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-12T09:17:23.535Z. First observed: 2026-11-10T08:33:47.356Z. Retrieved: 2026-11-12T09:04:03.387Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The supplied evidence is the abstract; it does not provide the underlying evaluation details or quantify the reported gap from the operational ECMWF IFS ensemble. The abstract describes a step toward observation-driven digital twins of the atmosphere; it does not report a deployed digital twin or demonstrate operational use.
Open question
Across which variables and lead times, and by how much, does the ensemble remain behind the operational ECMWF IFS ensemble?
Source-stated
The authors describe end-to-end weather forecasting systems as producing forecasts directly from raw Earth observations, replacing the numerical weather prediction pipeline, including data assimilation.
produce skillful global gridded and station forecasts directly from raw Earth observations, replacing the numerical weather prediction pipeline, including data assimilation
Source-stated
The authors add separate stochastic mechanisms to the observation encoder and processor to represent observation-related and learned-dynamics uncertainty.
learned, input-dependent noise at the observation encoder, aimed at the uncertainty inherited from the observing system, and Monte Carlo dropout in the processor
Source-stated
The authors report probabilistic finetuning improves the mean forecast by 11% on average across variables and lead times.
Probabilistic finetuning significantly improves the mean forecast, by 11% on average across variables and lead times.
Source-stated
The authors report the ensemble is close to calibrated against ERA5 through the medium range and remains considerably behind the operational ECMWF IFS ensemble.
The ensemble is close to calibrated against ERA5 through the medium range, is at least as accurate as the deterministic model at weather stations while beating it in relative economic value at every lead time, and remains considerably behind the operational ECMWF IFS ensemble.
Source-stated
The authors present component-attributed uncertainty as a step toward more transparent end-to-end forecasts and observation-driven digital twins of the atmosphere.
Component-attributed uncertainty makes end-to-end forecasts more transparent, a step toward observation-driven digital twins of the atmosphere.
Source and contributor context
Almeida, Rodrigo, Otero, Noelia, Arndt, Jost, Baur, Simon, Samek, Wojciech, Ma, Jackie · 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.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: acfdcff6-4ae7-4e30-88d5-efa6ec69b123.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-12T09:17:23.535Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 63e750ac71b4476c85ad847495f089ff6b5413eb14e2900c446c9f883ea0afa4. Source text hash: 1b73586aad73c6f6045f24e5b3c8b33da45237d58aac8f329fc3ca084658ae31.
Revision and source-check history
Version 1 · 2026-11-12T09:17:23.535Z
arXiv authors separate uncertainty sources in AI weather forecasts
The authors describe making Aardvark Weather probabilistic by adding separate sources of uncertainty to its observation encoder and processor. End-to-end systems forecast directly from raw Earth observations, replacing the numerical weather prediction pipeline, including data assimilation, but the abstract says they issue no uncertainty. Learned, input-dependent noise represents uncertainty inherited from observations; Monte Carlo dropout represents uncertainty in learned dynamics. A nested ensemble attributes forecast spread to these sources using a law-of-total-variance decomposition, cross-checked by withholding observation streams. The authors report that probabilistic improves the mean forecast by 11% on average across variables and lead times. They also report that the ensemble is close to calibrated against ERA5 through the medium range, is at least as accurate as the deterministic model at weather stations, and beats it in relative economic value at every lead . It remains considerably behind the operational ECMWF IFS ensemble. The authors describe component-attributed uncertainty as a way to make end-to-end forecasts more transparent and a step toward observation-driven digital twins of the atmosphere.
Original source
physics.ao-ph · Article
Source details
Original source date: Aug 31, 2026 (2026-08-31).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-12T09:17:23.535Z. First observed: 2026-11-10T08:33:47.356Z. Retrieved: 2026-11-12T09:04:03.387Z.
Limits
The supplied evidence is the abstract; it does not provide the underlying evaluation details or quantify the reported gap from the operational ECMWF IFS ensemble. The abstract describes a step toward observation-driven digital twins of the atmosphere; it does not report a deployed digital twin or demonstrate operational use.
Open question
Across which variables and lead times, and by how much, does the ensemble remain behind the operational ECMWF IFS ensemble?
Source-stated
The authors describe end-to-end weather forecasting systems as producing forecasts directly from raw Earth observations, replacing the numerical weather prediction pipeline, including data assimilation.
produce skillful global gridded and station forecasts directly from raw Earth observations, replacing the numerical weather prediction pipeline, including data assimilation
Source-stated
The authors add separate stochastic mechanisms to the observation encoder and processor to represent observation-related and learned-dynamics uncertainty.
learned, input-dependent noise at the observation encoder, aimed at the uncertainty inherited from the observing system, and Monte Carlo dropout in the processor
Source-stated
The authors report probabilistic finetuning improves the mean forecast by 11% on average across variables and lead times.
Probabilistic finetuning significantly improves the mean forecast, by 11% on average across variables and lead times.
Source-stated
The authors report the ensemble is close to calibrated against ERA5 through the medium range and remains considerably behind the operational ECMWF IFS ensemble.
The ensemble is close to calibrated against ERA5 through the medium range, is at least as accurate as the deterministic model at weather stations while beating it in relative economic value at every lead time, and remains considerably behind the operational ECMWF IFS ensemble.
Source-stated
The authors present component-attributed uncertainty as a step toward more transparent end-to-end forecasts and observation-driven digital twins of the atmosphere.
Component-attributed uncertainty makes end-to-end forecasts more transparent, a step toward observation-driven digital twins of the atmosphere.
Source and contributor context
Almeida, Rodrigo, Otero, Noelia, Arndt, Jost, Baur, Simon, Samek, Wojciech, Ma, Jackie · 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.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: acfdcff6-4ae7-4e30-88d5-efa6ec69b123.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-12T09:17:23.535Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 63e750ac71b4476c85ad847495f089ff6b5413eb14e2900c446c9f883ea0afa4. Source text hash: 1b73586aad73c6f6045f24e5b3c8b33da45237d58aac8f329fc3ca084658ae31.
uncertain · 2026-11-12T09:17:23.535Z · anthropic/claude-opus-5.5
Publication: acfdcff6-4ae7-4e30-88d5-efa6ec69b123
attribution — pass: The reading attributes the method and results to the paper's authors throughout, using phrases like 'the authors report' and 'the abstract says.' It does not present the source's claims as established fact. The headline's 'arXiv authors' is generic but accurate. Naming the team, for example 'Almeida et al.', would be a minor improvement.
Authors:Rodrigo Almeida, Noelia Otero, Jost Arndt, Simon Baur, Wojciech Samek, Jackie Ma
summary — uncertain: The reading is faithful, but a non-specialist may not grasp the core stakes. It omits the abstract's motivation that these systems replace the traditional pipeline 'at a fraction of its cost.' It also never says plainly that the change turns a single forecast into a range of likely outcomes. Central terms are left unexplained: ECMWF IFS ensemble (the operational benchmark), ERA5, Monte Carlo dropout, law-of-total-variance decomposition, relative economic value and data assimilation. Smallest repair: restore the cost clause, add one plain sentence on the single-forecast-to-range change, and add tooltips for ECMWF IFS, ERA5 and Monte Carlo dropout.
replacing the numerical weather prediction pipeline, including data assimilation, at a fraction of its cost. These systems are deterministic and issue no uncertainty.
date — pass: The originalPublished date of 2026-08-31 matches the v1 submission date. It is not the 9 Nov 2026 revision date or the 10 Nov feed date.
[Submitted on 31 Aug 2026 (v1), last revised 9 Nov 2026 (this version, v2)]
support — pass: Each claim matches the abstract verbatim, including the 11% gain, the near-calibration against ERA5, the station-accuracy and economic-value results, and the gap from ECMWF IFS. The limitations correctly note that the gap is not quantified and that no deployed digital twin is reported. The question follows from evidence the abstract leaves unquantified and is not leading. No benefits or costs are invented.
remains considerably behind the operational ECMWF IFS ensemble
media — pass: The selected media is the verified arXiv abstract page for this paper, which is the direct original source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Further training of an existing model to adapt it to a particular task or use.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivFurther training of an existing model to adapt it to a particular task or use.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv