# Wilder, Swaminathan, Storkey and Kuhlbrodt evaluate machine-learning prediction of subgrid eddy kinetic energy [Canonical reading](https://escapemeridian.com/readings/wilder-swaminathan-storkey-and-kuhlbrodt-evaluate-machine-02682d62) In their paper, Thomas Wilder, Ranjini Swaminathan, David Storkey and Till Kuhlbrodt evaluate whether a convolutional neural network can predict subgrid eddy kinetic energy, a quantity used in eddy closures for climate models. These closures approximate how eddies affect the larger-scale flow when a model cannot resolve those smaller processes. The authors compare models trained across regions and test how performance changes when regions or input features are held out. They report that models generalize to unseen regions when training data have similar feature and target distributions; models can fail when those distributions shift or filtering scales differ. Holding out features improves some generalizable models, but effects vary by region: some feature spaces help in some regions and worsen performance in others. The authors also report that predicted power spectral density generally diverges at small scales, although some feature spaces reduce that divergence. Their best model, trained on all regions, achieves time-mean R2 of approximately 0.95 and anomaly R2 of approximately 0.8–0.9. The work lays initial groundwork for a machine-learned scale-aware GEOMETRIC eddy closure for the UK's Earth System Model; the abstract reports evaluation results, not performance in a climate-model deployment. 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 4, 2026 (2026-11-04). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-10T04:08:43.095Z. First observed: 2026-11-05T05:00:21.003Z. Retrieved: 2026-11-10T04:07:56.913Z. - [Evaluating the prediction of subgrid eddy kinetic energy from a machine learning model](https://arxiv.org/abs/2611.06539) ## Limits The evidence is the paper's abstract; it gives no further detail here about the experiments or their results. The abstract describes offline evaluation and initial groundwork; it does not report performance in a climate-model deployment. ## Open question What feature and target distributions are similar enough for these models to generalize to an unseen region? ## Source-stated The authors use an offline learning approach with a convolutional neural network to predict subgrid eddy kinetic energy, as initial groundwork for a machine-learned scale-aware GEOMETRIC eddy closure for the UK's Earth System Model. > In this work we employ an offline learning approach with a convolutional neural network to predict subgrid eddy kinetic energy ($KE\_e$), laying the initial groundwork for a machine-learned scale-aware GEOMETRIC eddy closure for the UK's Earth System Model. - [Supporting source](https://arxiv.org/abs/2611.06539) ## Source-stated The authors report that models generalize to unseen regions when training data share similar feature and target distributions with those regions; distributional shift or different filtering scales can hinder generalization. > ML parameterization models generalize to unseen regions when their training data share similar feature and target distributions with those regions. Models that fail to generalize incur distributional shift, or employ different filtering scales - [Supporting source](https://arxiv.org/abs/2611.06539) ## Source-stated When features are held out, performance improves among generalizable models, but effects vary across regions and feature spaces. > When holding out features, there is improvement among the generalizable ML models, with some feature spaces leading to improvements in some regions but degrading performance in others. - [Supporting source](https://arxiv.org/abs/2611.06539) ## Source-stated The authors report that power spectral density generally diverges at small scales, though some feature spaces minimize the divergence. > Generally, power spectral density of $KE\_e$ tend to diverge at small scales, though some feature spaces are able to minimize this divergence. - [Supporting source](https://arxiv.org/abs/2611.06539) ## Source-stated The best-performing model trained on all regions achieves time-mean R2 of approximately 0.95 and anomaly R2 of approximately 0.8–0.9. > Overall, the best performing model trained on all regions achieves time-mean and anomaly R2 scores of R2$\\approx 0.95$ and R2$\\approx 0.8 - 0.9$, respectively. - [Supporting source](https://arxiv.org/abs/2611.06539) ## Source and contributor context Wilder, Thomas, Swaminathan, Ranjini, Storkey, David, Kuhlbrodt, Till · 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) ## Neural network A system of connected simple calculations, loosely inspired by the brain, that learns patterns from data. - [Definition reference](https://developers.google.com/machine-learning/glossary#neural-network) ## 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) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 83606d55-8860-4584-8a7a-efa197270ae9. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T04:08:43.095Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 3f4db7cd59b9e44ef4e052f29e28e43dca913e25809b07a6389e2aba351c66a7. Source text hash: 2d4f985814f4e663de1a1955b97b989954423bbabd1f6eedbf30623384466c85. ## Revision and source-check history ### Version 1 — 2026-11-10T04:08:43.095Z In their paper, Thomas Wilder, Ranjini Swaminathan, David Storkey and Till Kuhlbrodt evaluate whether a convolutional neural network can predict subgrid eddy kinetic energy, a quantity used in eddy closures for climate models. These closures approximate how eddies affect the larger-scale flow when a model cannot resolve those smaller processes. The authors compare models trained across regions and test how performance changes when regions or input features are held out. They report that models generalize to unseen regions when training data have similar feature and target distributions; models can fail when those distributions shift or filtering scales differ. Holding out features improves some generalizable models, but effects vary by region: some feature spaces help in some regions and worsen performance in others. The authors also report that predicted power spectral density generally diverges at small scales, although some feature spaces reduce that divergence. Their best model, trained on all regions, achieves time-mean R2 of approximately 0.95 and anomaly R2 of approximately 0.8–0.9. The work lays initial groundwork for a machine-learned scale-aware GEOMETRIC eddy closure for the UK's Earth System Model; the abstract reports evaluation results, not performance in a climate-model deployment. ## Original source arXiv physics.ao-ph · Article Original source date: Nov 4, 2026 (2026-11-04). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-10T04:08:43.095Z. First observed: 2026-11-05T05:00:21.003Z. Retrieved: 2026-11-10T04:07:56.913Z. - [Evaluating the prediction of subgrid eddy kinetic energy from a machine learning model](https://arxiv.org/abs/2611.06539) ## Limits The evidence is the paper's abstract; it gives no further detail here about the experiments or their results. The abstract describes offline evaluation and initial groundwork; it does not report performance in a climate-model deployment. ## Open question What feature and target distributions are similar enough for these models to generalize to an unseen region? ## Source-stated The authors use an offline learning approach with a convolutional neural network to predict subgrid eddy kinetic energy, as initial groundwork for a machine-learned scale-aware GEOMETRIC eddy closure for the UK's Earth System Model. > In this work we employ an offline learning approach with a convolutional neural network to predict subgrid eddy kinetic energy ($KE\_e$), laying the initial groundwork for a machine-learned scale-aware GEOMETRIC eddy closure for the UK's Earth System Model. - [Supporting source](https://arxiv.org/abs/2611.06539) ## Source-stated The authors report that models generalize to unseen regions when training data share similar feature and target distributions with those regions; distributional shift or different filtering scales can hinder generalization. > ML parameterization models generalize to unseen regions when their training data share similar feature and target distributions with those regions. Models that fail to generalize incur distributional shift, or employ different filtering scales - [Supporting source](https://arxiv.org/abs/2611.06539) ## Source-stated When features are held out, performance improves among generalizable models, but effects vary across regions and feature spaces. > When holding out features, there is improvement among the generalizable ML models, with some feature spaces leading to improvements in some regions but degrading performance in others. - [Supporting source](https://arxiv.org/abs/2611.06539) ## Source-stated The authors report that power spectral density generally diverges at small scales, though some feature spaces minimize the divergence. > Generally, power spectral density of $KE\_e$ tend to diverge at small scales, though some feature spaces are able to minimize this divergence. - [Supporting source](https://arxiv.org/abs/2611.06539) ## Source-stated The best-performing model trained on all regions achieves time-mean R2 of approximately 0.95 and anomaly R2 of approximately 0.8–0.9. > Overall, the best performing model trained on all regions achieves time-mean and anomaly R2 scores of R2$\\approx 0.95$ and R2$\\approx 0.8 - 0.9$, respectively. - [Supporting source](https://arxiv.org/abs/2611.06539) ## Source and contributor context Wilder, Thomas, Swaminathan, Ranjini, Storkey, David, Kuhlbrodt, Till · 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) ## Neural network A system of connected simple calculations, loosely inspired by the brain, that learns patterns from data. - [Definition reference](https://developers.google.com/machine-learning/glossary#neural-network) ## 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) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 83606d55-8860-4584-8a7a-efa197270ae9. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T04:08:43.095Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 3f4db7cd59b9e44ef4e052f29e28e43dca913e25809b07a6389e2aba351c66a7. Source text hash: 2d4f985814f4e663de1a1955b97b989954423bbabd1f6eedbf30623384466c85. ### uncertain — 2026-11-10T04:08:43.095Z — anthropic/claude-opus-5.5 Publication: 83606d55-8860-4584-8a7a-efa197270ae9 attribution — pass: The draft correctly attributes the paper and its findings to Thomas Wilder, Ranjini Swaminathan, David Storkey and Till Kuhlbrodt. Reported results stay framed as the authors' claims. > Authors:Thomas Wilder, Ranjini Swaminathan, David Storkey, Till Kuhlbrodt summary — uncertain: The draft omits the paper's stated central question: whether global parameterizations are needed or regional closures are enough. Saying the authors compare models trained across regions does not convey this. Repair: add that question. The draft also omits the stated motivation that ML closures could help reduce model biases and uncertainty. The headline metric R2 and power spectral density are not explained, so non-specialists cannot judge what roughly 0.95 or the small-scale divergence means. Suggest term-library tooltips for these terms. > We address the question of whether global parameterizations are needed or if regional closures are sufficient by evaluating a number of neural network models through region and feature ablation experiments. date — pass: The originalPublished date of 2026-11-04 matches the v1 submission date on arXiv. The feed date of 2026-11-05 is later and was correctly not used. > \[Submitted on 4 Nov 2026\] support — pass: Each claim closely tracks the abstract: generalization conditions, feature-ablation effects, small-scale spectral divergence and R2 scores. The draft correctly distinguishes offline evaluation from deployment in a climate model. The question follows from the evidence without assuming an answer. > ML parameterization models generalize to unseen regions when their training data share similar feature and target distributions with those regions. 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/02682d62d34b21aef5d99cd3badc723ab02f622e0385b1e79f02e788c1d2d55a/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/wilder-swaminathan-storkey-and-kuhlbrodt-evaluate-machine-02682d62.json)