# arXiv: Ensemble Covariance Propagation with Machine Learning Weather Prediction Models [Canonical reading](https://escapemeridian.com/readings/arxiv-ensemble-covariance-propagation-with-machine-learning-fbdf94a3) Simon Toedtli, David Ahijevych, Wei Wang and Chris Snyder evaluate whether GraphCast, a machine learning weather prediction model, can evolve ensemble covariances in a two-way coupled data assimilation system. In that system, the ensemble variational (EnVar) component evolves the mean while the ensemble Kalman filter (EnKF) evolves covariances. Over one month of cycling with real-world observations, including microwave radiances, they use GraphCast for the EnKF and the physics-based MPAS model for the EnVar. The authors report that ensemble spread and covariances are comparable to a reference configuration using MPAS for both components. The resulting EnVar mean has consistently lower RMSE against GFS analyses, except near the model top, where spurious correlations degrade performance. The authors present this as a proof of concept that MLWP models can propagate covariance information in a realistic data assimilation cycling system, supporting further consideration of their integration into future operational ensemble systems. The reported results come from this configuration and evaluation; the abstract does not establish operational performance. 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 9, 2026 (2026-11-09). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-10T21:19:49.947Z. First observed: 2026-11-10T08:34:00.915Z. Retrieved: 2026-11-10T21:18:45.883Z. - [Ensemble Covariance Propagation with Machine Learning Weather Prediction Models in a Realistic Data Assimilation System](https://arxiv.org/abs/2611.11918) ## Limits The reported evaluation cycles the system over one month; the abstract does not report results beyond that evaluation. The abstract reports degraded performance near the model top from spurious correlations but does not identify their cause. The authors call the results a proof of concept for future operational integration; the abstract does not establish performance in operational ensemble systems. ## Open question What causes the spurious correlations near the model top that degrade the resulting EnVar mean’s performance? ## Source-stated The authors evaluate GraphCast for evolving ensemble covariances within a two-way coupled data assimilation system. > We evaluate this capability in a two-way coupled DA system, where an ensemble variational (EnVar) component evolves the mean and an ensemble Kalman filter (EnKF) evolves the covariances. - [Supporting source](https://arxiv.org/abs/2611.11918) ## Source-stated The evaluation cycles the system over one month and assimilates real-world observations, including microwave radiances. > The system is cycled over one month and assimilates a wide range of real-world observations, including microwave radiances. - [Supporting source](https://arxiv.org/abs/2611.11918) ## Source-stated With GraphCast evolving the EnKF covariances and MPAS evolving the EnVar mean, ensemble spread and covariances are comparable to the reference configuration using MPAS for both components. > Using the MLWP model GraphCast for the EnKF and the physics-based MPAS model for the EnVar yields ensemble spread and covariances comparable to a reference configuration in which MPAS evolves both components. - [Supporting source](https://arxiv.org/abs/2611.11918) ## Source-stated The resulting EnVar mean has consistently lower RMSE against GFS analyses, except near the model top, where spurious correlations degrade performance. > The resulting EnVar mean achieves consistently lower RMSE against GFS analyses, except near the model top, where spurious correlations degrade performance. - [Supporting source](https://arxiv.org/abs/2611.11918) ## Source-stated The authors describe the results as a proof of concept for integrating MLWP models into future operational ensemble systems. > These results demonstrate that MLWP models can propagate covariance information in a realistic DA cycling system and provide a proof of concept for their integration into future operational ensemble systems. - [Supporting source](https://arxiv.org/abs/2611.11918) ## Source and contributor context Toedtli, Simon, Ahijevych, David, Wang, Wei, Snyder, Chris · 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. ## 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) ## 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: e49bc24c-539a-4de8-8de7-64f49992b026. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T21:19:49.947Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: a93db83d089ccf0560e91b2a44aad01e9794af082ed7bf556bf24ab6af62e6d0. Source text hash: b62227a15e070932cf7ee40b87eaeaaab85687c91a56c2ca23465f4ee6b9bb63. ## Revision and source-check history ### Version 1 — 2026-11-10T21:19:49.947Z Simon Toedtli, David Ahijevych, Wei Wang and Chris Snyder evaluate whether GraphCast, a machine learning weather prediction model, can evolve ensemble covariances in a two-way coupled data assimilation system. In that system, the ensemble variational (EnVar) component evolves the mean while the ensemble Kalman filter (EnKF) evolves covariances. Over one month of cycling with real-world observations, including microwave radiances, they use GraphCast for the EnKF and the physics-based MPAS model for the EnVar. The authors report that ensemble spread and covariances are comparable to a reference configuration using MPAS for both components. The resulting EnVar mean has consistently lower RMSE against GFS analyses, except near the model top, where spurious correlations degrade performance. The authors present this as a proof of concept that MLWP models can propagate covariance information in a realistic data assimilation cycling system, supporting further consideration of their integration into future operational ensemble systems. The reported results come from this configuration and evaluation; the abstract does not establish operational performance. ## Original source arXiv physics.ao-ph · Article Original source date: Nov 9, 2026 (2026-11-09). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-10T21:19:49.947Z. First observed: 2026-11-10T08:34:00.915Z. Retrieved: 2026-11-10T21:18:45.883Z. - [Ensemble Covariance Propagation with Machine Learning Weather Prediction Models in a Realistic Data Assimilation System](https://arxiv.org/abs/2611.11918) ## Limits The reported evaluation cycles the system over one month; the abstract does not report results beyond that evaluation. The abstract reports degraded performance near the model top from spurious correlations but does not identify their cause. The authors call the results a proof of concept for future operational integration; the abstract does not establish performance in operational ensemble systems. ## Open question What causes the spurious correlations near the model top that degrade the resulting EnVar mean’s performance? ## Source-stated The authors evaluate GraphCast for evolving ensemble covariances within a two-way coupled data assimilation system. > We evaluate this capability in a two-way coupled DA system, where an ensemble variational (EnVar) component evolves the mean and an ensemble Kalman filter (EnKF) evolves the covariances. - [Supporting source](https://arxiv.org/abs/2611.11918) ## Source-stated The evaluation cycles the system over one month and assimilates real-world observations, including microwave radiances. > The system is cycled over one month and assimilates a wide range of real-world observations, including microwave radiances. - [Supporting source](https://arxiv.org/abs/2611.11918) ## Source-stated With GraphCast evolving the EnKF covariances and MPAS evolving the EnVar mean, ensemble spread and covariances are comparable to the reference configuration using MPAS for both components. > Using the MLWP model GraphCast for the EnKF and the physics-based MPAS model for the EnVar yields ensemble spread and covariances comparable to a reference configuration in which MPAS evolves both components. - [Supporting source](https://arxiv.org/abs/2611.11918) ## Source-stated The resulting EnVar mean has consistently lower RMSE against GFS analyses, except near the model top, where spurious correlations degrade performance. > The resulting EnVar mean achieves consistently lower RMSE against GFS analyses, except near the model top, where spurious correlations degrade performance. - [Supporting source](https://arxiv.org/abs/2611.11918) ## Source-stated The authors describe the results as a proof of concept for integrating MLWP models into future operational ensemble systems. > These results demonstrate that MLWP models can propagate covariance information in a realistic DA cycling system and provide a proof of concept for their integration into future operational ensemble systems. - [Supporting source](https://arxiv.org/abs/2611.11918) ## Source and contributor context Toedtli, Simon, Ahijevych, David, Wang, Wei, Snyder, Chris · 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. ## 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) ## 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: e49bc24c-539a-4de8-8de7-64f49992b026. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T21:19:49.947Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: a93db83d089ccf0560e91b2a44aad01e9794af082ed7bf556bf24ab6af62e6d0. Source text hash: b62227a15e070932cf7ee40b87eaeaaab85687c91a56c2ca23465f4ee6b9bb63. ### uncertain — 2026-11-10T21:19:49.947Z — anthropic/claude-opus-5.5 Publication: e49bc24c-539a-4de8-8de7-64f49992b026 attribution — pass: The draft names the four listed authors, and the paper's findings and proof-of-concept framing are attributed to them rather than presented as established fact. > Authors:Simon Toedtli, David Ahijevych, Wei Wang, Chris Snyder summary — uncertain: The summary is faithful but not readable for a non-specialist. Central terms are never explained in plain language: data assimilation, ensemble covariances, ensemble spread, EnVar, EnKF, RMSE, GFS analyses, MPAS and microwave radiances. The draft also never says why propagating covariances matters for weather forecasting. Smallest repair: add short plain-language glosses or term-library tooltips for these terms, keeping the terms themselves unchanged. 'Lower RMSE' has no stated comparison point; the abstract is equally unspecific, so the draft should note that gap rather than leave it implicit. > The resulting EnVar mean achieves consistently lower RMSE against GFS analyses, except near the model top date — pass: The displayed original publication date of 2026-11-09 matches the v1 submission date. The later feed date of 2026-11-10 is correctly not used. > \[Submitted on 9 Nov 2026\] support — pass: Each claim matches the abstract verbatim in substance. The limitations correctly state that operational performance is not established and that the cause of the near-top degradation is unidentified. The question follows directly from that unresolved issue. No benefits or impacts are invented. > provide a proof of concept for their integration into future operational ensemble systems media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to 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/fbdf94a3feed025c51e963adac5dca7111947b323140a4b2f21c351579c54e70/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/arxiv-ensemble-covariance-propagation-with-machine-learning-fbdf94a3.json)