Pith. sign in

REVIEW 2 major objections 5 minor 37 references

This paper claims that machine learning weather models can be tested for physical consistency by rotating or reversing the planet's coordinates, and that GraphCast and NeuralGCM fail these tests while a physics-based general circulation mod

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 09:33 UTC pith:WPONKNA6

load-bearing objection Solid, novel testing framework; the Reverse Longitude result for GraphCast is weakened by inconsistent inputs, so the headline claim overreaches. the 2 major comments →

arxiv 2607.20716 v1 pith:WPONKNA6 submitted 2026-07-22 physics.ao-ph

Spatial Generalization Tests for Machine Learning-based Weather Models to Assess Physical Consistency

Maren H\"over , Milan Kl\"ower , Christian Schroeder de Witt , Hannah M. Christensen This is my paper
classification physics.ao-ph MSC 86A1068T07
keywords spatial generalizationmachine learning weather predictionphysical consistencycoordinate invarianceGraphCastNeuralGCMgeneralization errordaily cycle shortcuts
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper asks whether machine learning weather models have learned the spatially invariant physics of the atmosphere, or instead rely on shortcuts tied to present-day locations and times. It introduces three coordinate transformations—rotate the planet in longitude, reverse latitude, reverse longitude—applied consistently to initial and boundary conditions, then compares forecasts on the transformed planet with baseline forecasts after undoing the transformation. A physics-based general circulation model passes with only rounding errors, while GraphCast and NeuralGCM fail. The analysis shows that GraphCast learns the daily cycle partly from date-time and coordinate inputs rather than from solar radiation, and that NeuralGCM's encoder-decoder imprints 'ghost continents' where continents appear in the training data. The authors argue that passing such spatial generalization tests is necessary before ML weather and climate models can be trusted under changing boundary conditions such as melting ice and land-use change.

Core claim

The central claim is that spatial generalization can be tested cleanly by exploiting the invariance of the atmosphere's governing equations under three involutive coordinate transformations. Once the required sign changes are applied to winds, forcing, and where possible rotation and time, a physically consistent weather model should produce forecasts that, after reverting the transformation, match the baseline. The paper reports that GraphCast violates this: its generalization error is comparable to or larger than its forecast error, and it maps the daily cycle partly from date-time and coordinate information rather than from incident solar radiation. NeuralGCM's encoder-decoder also fails

What carries the argument

The machinery is a set of three spatial involutions—half-turn in longitude, reflection in latitude, and reflection in longitude—applied to every initial and boundary field, together with the sign changes they induce in wind components, Coriolis or planetary rotation, and, where the model allows, time and solar forcing. The test metric is the generalization error: the difference between a baseline forecast and the forecast on the transformed planet after reversing the transformation. A model passes if this error is negligible relative to its own forecast error; a physics-based general circulation model demonstrates passability with errors at rounding level.

Load-bearing premise

The test cases are treated as physically self-consistent, so that a model failure indicates a learned spatial non-invariance rather than an inconsistent input set; this is violated for GraphCast's Reverse Longitude case, where time cannot be made local or backward and the Coriolis parameter is learned and cannot be changed, and it remains unproven for NeuralGCM, where only the encoder-decoder is tested without full time integration.

What would settle it

Run the Reverse Longitude test on an ML weather model that accepts local, coordinate-dependent time and allows the planetary rotation or Coriolis parameter to be negated. If the generalization error drops to rounding-error level, the paper's attribution of failure to coordinate- and time-dependent shortcuts is supported; if errors persist, the failure may instead stem from the test's internal inconsistency. Alternatively, run NeuralGCM's full model with a manual pressure-to-sigma conversion; if ghost continents disappear during time integration, the claim that the encoder-decoder itself breaks

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Forecast-skill metrics alone are insufficient to certify physical consistency in ML weather models; spatial generalization tests can expose nonphysical shortcuts.
  • ML weather models that fail these tests cannot be used for scientific experiments such as idealized solar forcing, land-use change, or melting sea ice.
  • A model that genuinely generalizes spatially should be more robust to shifting climate conditions and may require less frequent retraining.
  • The proposed test protocol can be extended beyond the three base cases to seven combined transformations and other rotation angles depending on grid symmetry.
  • If the daily-cycle finding is correct, removing date-time inputs and relying on physical radiative forcing could eliminate one identified source of nonphysical mapping.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If temporal generalization depends on spatial generalization, then models failing these tests are unlikely to be reliable climate projection tools, regardless of their weather forecast skill.
  • The same protocol could be turned into a training signal by augmenting training data with spatially transformed fields, potentially forcing architectures to learn invariant physics.
  • The identified dependence on coordinates and time suggests a concrete architectural change—replace coordinate/time features with physically meaningful boundary and forcing fields—that could be tested by retraining and re-running the tests.
  • The test framework could be extended to probabilistic and ensemble forecast models, checking whether their internal distributions remain invariant under the same transformations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper proposes three spatial generalization test cases for ML-based weather models: Rotate Longitude, Reverse Latitude, and Reverse Longitude. The tests transform all input and boundary fields and appropriate wind components (and, where possible, planetary rotation and time information) so that a spatially invariant model should reproduce the baseline forecast after the transformation is reverted. The authors apply the tests to GraphCast small and to NeuralGCM's encoder-decoder, with SpeedyWeather as a physics-based control. They report that GraphCast fails all three tests, that NeuralGCM's encoder-decoder produces 'ghost continents,' and that SpeedyWeather reproduces the baseline to rounding-error precision. They further diagnose GraphCast's daily-cycle dependence on date-time and coordinate information. The central claim is that current ML weather models rely on location/time shortcuts and lack the spatial invariance required for climate-change generalization.

Significance. The proposed tests are simple, parameter-free, and directly address an important gap: forecast skill metrics do not measure physical consistency. The paper ships reproducible code and uses a 365-member ensemble, and the SpeedyWeather control plus the Appendix A invariance derivation are useful reference points. If the central claim holds, the work provides a practical diagnostic that could guide future architecture design. However, the strongest claims currently outrun the evidence: the Reverse Longitude test for GraphCast is admitted to be physically inconsistent (time and Coriolis are not adapted), and the NeuralGCM evaluation covers only the encoder-decoder, not the full forecast model. These issues do not invalidate the overall approach, but they require substantial revision before the stated conclusions are supported.

major comments (2)
  1. [Sec. 3.2, Sec. 5.1 (Table 2), Fig. 3] The Reverse Longitude test for GraphCast is not physically self-consistent. The authors state that time would need to run backwards or be formulated as local time, and that the Coriolis parameter is learned and cannot be changed. Therefore the model receives a reversed planet with an unreversed time input and no Ω→−Ω, so the reported 1.5–3× forecast-error generalization error may be dominated by this known input inconsistency rather than by learned spatial non-invariance. This undermines the statement in Sec. 3.1 that 'GraphCast fails each generalization test' and the abstract's claim that all boundary conditions and forcings were adapted accordingly. Please either drop Reverse Longitude from the GraphCast failure claims, provide a controlled experiment with consistent time/Coriolis handling, or explicitly re-label it as a physical-consistency test rather than a spatial-invariance test.
  2. [Sec. 3.3] The NeuralGCM evaluation is limited to the encoder-decoder module without any time integration. The assertion that 'errors will only be further amplified' is not demonstrated and is not logically guaranteed: the physics-based dynamical core could damp or correct initial-condition errors introduced by the encoder-decoder. Hence the abstract's claim that 'GraphCast and NeuralGCM fail these tests' overstates the NeuralGCM evidence. Please either run the full forecast loop (e.g., with a manual sigma-coordinate conversion and an explicit statement of what is excluded) or restrict the claim to 'NeuralGCM's encoder-decoder component fails to generalize spatially.'
minor comments (5)
  1. [Abstract and Sec. 1] The statement 'Physics-based general circulation models simulate a rotated/reversed planet with only rounding errors' is supported only by SpeedyWeather and by a 2D incompressible Navier-Stokes proof. Full GCMs include moist processes, radiation, and land-surface schemes not covered by the proof. Please qualify the claim as applying to the tested model or add a limitation statement.
  2. [Sec. 3.3] The phrase 'errors will only be further amplified' is an unsupported assertion. Consider replacing it with a statement of intent to test the full model in future work.
  3. [Sec. 5.5, Eq. (1)] The climatological skill score is computed from MSE and can be negative; this is not discussed in the text. Please add a sentence explaining the interpretation and acceptable range.
  4. [Fig. 4] Panel (j) is not defined in the caption. Please specify what correlation is plotted and how it is computed.
  5. [Appendix A] The notation R_x, R_y, R_o and the derivation leading to Eqs. (4)–(5) are difficult to follow. Suggest rewriting with explicit coordinate arguments or adding a table of transformation rules for clarity.

Circularity Check

0 steps flagged

No significant circularity: the invariance derivation is self-contained and the empirical test protocol compares models with themselves; the Reverse Longitude inconsistency is a validity caveat, not a circular derivation.

full rationale

The paper's formal derivation is Appendix A, which is self-contained and parameter-free: it starts from the 2D incompressible Navier-Stokes equations with Coriolis forcing and shows that applying reversal/rotation operators together with the tilde transforms (Eqs. 7-9) recovers the original equations. No fitted constant or empirical target is imported. The empirical test protocol is also non-circular: the generalization error is computed as the difference between a model's own baseline forecast and its forecast on rotated/reversed inputs (Sec. 5.5), and the pass/fail criterion is a comparison with the same model's forecast error, not with any quantity fitted to the conclusion. The passages that could look circular are the admissions in Sec. 3.2 and Sec. 5.1 that Reverse Longitude is not fully physically consistent for GraphCast: time cannot be run backwards or made local, and the Coriolis parameter is learned and cannot be changed. These are genuine validity caveats — they mean the Reverse Longitude failure may be partly caused by the input inconsistency itself — but they are not circularity, because the paper does not define failure as the inconsistency; it argues the inconsistency exposes the model's time and coordinate dependence. Similarly, the use of SpeedyWeather, co-developed by an author, is supplemented by the independent analytic invariance proof, so the self-citation is not load-bearing. The NeuralGCM result is explicitly scoped to the encoder-decoder module (Sec. 3.3) and is not used to claim a full-model forecast failure. Overall, no step reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

No free parameters are fitted to make the result work; the decision threshold in Eq. 2 is qualitative and not assigned a value. The central claim rests on the spatial-invariance assumption and the self-consistency of the transformed inputs; the latter is violated for GraphCast's Reverse Longitude test and untested for NeuralGCM's full forecast loop. No new physical entities are introduced.

axioms (4)
  • domain assumption The atmospheric governing equations are invariant under the three spatial transformations when the specified sign changes, time shifts, and solar-forcing rotations are applied.
    Derived in Appendix A for incompressible 2D Navier-Stokes with periodic boundary conditions; assumed to extend to full primitive-equation models with moisture and radiation. Supported by SpeedyWeather experiment (Fig. 9), but not proven for every process in GraphCast/NeuralGCM.
  • ad hoc to paper The transformed input fields are physically self-consistent and complete for each ML model.
    Section 5.1/Table 2 describe transforming initial and boundary conditions, solar radiation, wind components, and time. For GraphCast Reverse Longitude, time and Coriolis cannot be adjusted, so this premise is knowingly false for that test case.
  • domain assumption Leaving coordinate inputs and embeddings unchanged while transforming fields is the correct way to expose learned spatial shortcuts rather than architecture artifacts.
    Section 5.1/Table 1 disallow changing coordinates or embeddings. This presupposes that a spatially generalizing model should be nearly equivariant to absolute coordinates even if trained with them.
  • domain assumption 2m-temperature is a sufficient diagnostic for the failure/success classification.
    Section 5.5 explains the choice via GraphCast's loss weighting and strong daily/seasonal cycle. Other variables could show different generalization-error magnitudes.

pith-pipeline@v1.3.0-alltime-deepseek · 14561 in / 13911 out tokens · 137557 ms · 2026-08-01T09:33:35.192795+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Spatial Generalization Tests for Machine Learning-based Weather Models to Assess Physical Consistency." pith.science (2026). https://pith.science/paper/WPONKNA6

@misc{pith2026260720716,
  author       = {Pith},
  title        = {Pith review of: Spatial Generalization Tests for Machine Learning-based Weather Models to Assess Physical Consistency},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WPONKNA6}},
  note         = {Machine review of arXiv:2607.20716}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Machine learning-based weather prediction is revolutionizing weather forecasting by learning from weather data in present-day climate. However, generalization to other climates remains a major challenge. With melting sea ice, land-use change, and increasing ocean temperatures, boundary conditions are changing. Therefore, generalization in time depends on generalization in space. Here, we present three test cases to evaluate whether machine learning-based weather and climate models generalize in space and apply them to GraphCast and NeuralGCM. We reverse or rotate the planet in longitude or latitude under the model's coordinate system and adapt all boundary conditions and forcings accordingly. Physics-based general circulation models simulate a rotated/reversed planet with only rounding errors, but GraphCast and NeuralGCM fail these tests. The analyses furthermore revealed unphysical variable mappings based on correlation rather than causation. We argue that machine learning-based climate models should be designed to pass generalization tests to prevent overfitting on present-day regional climate.

Figures

Figures reproduced from arXiv: 2607.20716 by Christian Schroeder de Witt, Hannah M. Christensen, Maren H\"over, Milan Kl\"ower.

Figure 1
Figure 1. Figure 1: Three spatial generalization test cases. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the generalization error calculation. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Spatially averaged generalization errors in 2m-temperature of [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Daily cycle learned from combined solar radiation, time and coordi [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Spatial distributions of errors in 2m-temperature forecasts of [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Spatial generalization of NeuralGCM’s encoder-decoder. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Example of a subset of GraphCast initial conditions for the three test cases, including sign changes for the wind components. In ERA5, top-of-atmosphere (TOA) incident solar radiation is accumulated, units here are integrated Wm−2 over one hour. Surface geopotential is orographic height times gravity. 5.2 Mathematics of rotation and reversal Rotate Longitude, Reverse Latitude, and Reverse Longitude mathema… view at source ↗
Figure 8
Figure 8. Figure 8: GraphCast generalization error as [PITH_FULL_IMAGE:figures/full_fig_p028_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: SpeedyWeather’s generalization error by test case [PITH_FULL_IMAGE:figures/full_fig_p029_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: RMSE in 2m-temperature for different daily cycle adaptations. [PITH_FULL_IMAGE:figures/full_fig_p030_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Effect of changing geopotential at surface on 2m-temperature. [PITH_FULL_IMAGE:figures/full_fig_p031_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 1 canonical work pages

  1. [1]

    Forecasting Global Weather with Graph Neural Networks (2022)

    Keisler, R. Forecasting Global Weather with Graph Neural Networks (2022). URL http://arxiv.org/abs/2202.07575. ArXiv:2202.07575 [physics]

    Pith/arXiv arXiv 2022

  2. [2]

    URL https://www.science.org/doi/10.1126/science.adi2336

    Lam, R.et al.Learning skillful medium-range global weather forecasting.Science (2023). URL https://www.science.org/doi/10.1126/science.adi2336

    doi:10.1126/science.adi2336 2023

  3. [3]

    URL https://www.nature.com/articles/ s41586-023-06185-3

    Bi, K.et al.Accurate medium-range global weather forecasting with 3D neural networks.Nature619, 533–538 (2023). URL https://www.nature.com/articles/ s41586-023-06185-3

    2023

  4. [4]

    URL https://www.nature.com/articles/s41586-024-08252-9

    Price, I.et al.Probabilistic weather forecasting with machine learning.Nature 637, 84–90 (2025). URL https://www.nature.com/articles/s41586-024-08252-9

    2025

  5. [5]

    URL https://www.nature.com/articles/ s41586-024-07744-y

    Kochkov, D.et al.Neural general circulation models for weather and cli- mate.Nature632, 1060–1066 (2024). URL https://www.nature.com/articles/ s41586-024-07744-y

    2024

  6. [6]

    Rasp, S., Pritchard, M. S. & Gentine, P. Weatherbench: A benchmark dataset for data-driven weather forecasting.Journal of Advances in Modeling Earth Systems 12, e2020MS002109 (2020)

    2020

  7. [7]

    URL http://arxiv.org/ abs/2202.11214

    Pathak, J.et al.FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators (2022). URL http://arxiv.org/ abs/2202.11214. 2202.11214. 33

    Pith/arXiv arXiv 2022

  8. [8]

    & LeCun, Y

    Mathieu, M., Couprie, C. & LeCun, Y. Deep multi-scale video prediction beyond mean square error (2016). URL http://arxiv.org/abs/1511.05440. 1511.05440

    Pith/arXiv arXiv 2016

  9. [9]

    URL https://onlinelibrary.wiley.com/doi/abs/10.1029/ 2023MS004021

    Karlbauer, M.et al.Advancing Parsimonious Deep Learning Weather Prediction Using the HEALPix Mesh.Journal of Advances in Modeling Earth Systems16, e2023MS004021 (2024). URL https://onlinelibrary.wiley.com/doi/abs/10.1029/ 2023MS004021

    2024

  10. [10]

    URL https://gmd.copernicus.org/articles/18/5575/2025/

    Mahesh, A.et al.Huge ensembles – Part 1: Design of ensemble weather forecasts using spherical Fourier neural operators.Geoscientific Model Development18, 5575–5603 (2025). URL https://gmd.copernicus.org/articles/18/5575/2025/

    2025

  11. [11]

    URL http://arxiv.org/abs/2310.02074

    Watt-Meyer, O.et al.ACE: A fast, skillful learned global atmospheric model for climate prediction (2023). URL http://arxiv.org/abs/2310.02074

    Pith/arXiv arXiv 2023

  12. [12]

    Ullrich, P. A.et al.Recommendations for comprehensive and indepen- dent evaluation of machine learning-based earth system models.Jour- nal of Geophysical Research: Machine Learning and Computation2, e2024JH000496 (2025). URL https://agupubs.onlinelibrary.wiley.com/doi/abs/ 10.1029/2024JH000496. E2024JH000496 2024JH000496

    doi:10.1029/2024jh000496 2025

  13. [13]

    URL https://arxiv.org/abs/2605.06944

    Henn, B.et al.Aimip phase 1: systematic evaluations of ai weather and climate models (2026). URL https://arxiv.org/abs/2605.06944. arXiv:2605.06944

    Pith/arXiv arXiv 2026

  14. [14]

    URL https://rmets.onlinelibrary

    Hersbach, H.et al.The ERA5 global reanalysis.Quarterly Journal of the Royal Meteorological Society146, 1999–2049 (2020). URL https://rmets.onlinelibrary. wiley.com/doi/10.1002/qj.3803

    doi:10.1002/qj.3803 1999

  15. [15]

    URL http://arxiv.org/abs/2409.18529

    Rackow, T.et al.Robustness of AI-based weather forecasts in a changing climate (2024). URL http://arxiv.org/abs/2409.18529. 34

    Pith/arXiv arXiv 2024

  16. [16]

    URL https://onlinelibrary.wiley.com/doi/abs/10

    Rasp, S.et al.WeatherBench 2: A Benchmark for the Next Generation of Data- Driven Global Weather Models.Journal of Advances in Modeling Earth Systems 16, e2023MS004019 (2024). URL https://onlinelibrary.wiley.com/doi/abs/10. 1029/2023MS004019

    2024

  17. [17]

    & Sukhatme, J

    Jalan, S. & Sukhatme, J. Intraseasonal Equatorial Kelvin and Rossby Waves in Modern AI-ML Models (2025). URL http://arxiv.org/abs/2507.07952

    Pith/arXiv arXiv 2025

  18. [18]

    & Barnes, E

    Diao, C. & Barnes, E. A. Assessing MJO Tropical-Extratropical Teleconnections in Deep Learning Weather Prediction Mod- els (2025). URL https://essopenarchive.org/users/631510/articles/ 1296512-assessing-mjo-tropical-extratropical-teleconnections-in-deep-learning-weather-prediction-models

    2025

  19. [19]

    URL https://www.nature.com/articles/s41612-025-00949-6

    Ba˜ no-Medina, J.et al.Are AI weather models learning atmospheric physics? A sensitivity analysis of cyclone Xynthia.Climate and Atmospheric Science8, 92 (2025). URL https://www.nature.com/articles/s41612-025-00949-6

    2025

  20. [20]

    & Messori, G

    Olivetti, L. & Messori, G. Do data-driven models beat numerical models in forecasting weather extremes? A comparison of IFS HRES, Pangu-Weather, and GraphCast.Geoscientific Model Development17, 7915–7962 (2024). URL https: //gmd.copernicus.org/articles/17/7915/2024/

    2024

  21. [21]

    J., Yang, W

    Meng, Z., Hakim, G. J., Yang, W. & Vecchi, G. A. Deep Learning Atmospheric Models Reliably Simulate Out-of-Sample Land Heat and Cold Wave Frequencies (2025). URL http://arxiv.org/abs/2507.03176

    arXiv 2025

  22. [22]

    & Engelke, S

    Zhang, Z., Fischer, E., Zscheischler, J. & Engelke, S. Physics-based models out- perform AI weather forecasts of record-breaking extremes.Science Advances12, eaec1433 (2026). URL https://www.science.org/doi/10.1126/sciadv.aec1433. 35

    doi:10.1126/sciadv.aec1433 2026

  23. [23]

    On Some Limitations of Current Machine Learning Weather Pre- diction Models.Geophysical Research Letters51, e2023GL107377 (2024)

    Bonavita, M. On Some Limitations of Current Machine Learning Weather Pre- diction Models.Geophysical Research Letters51, e2023GL107377 (2024). URL https://onlinelibrary.wiley.com/doi/abs/10.1029/2023GL107377

    doi:10.1029/2023gl107377 2024

  24. [24]

    & Craig, G

    Selz, T. & Craig, G. C. Can Artificial Intelligence-Based Weather Prediction Models Simulate the Butterfly Effect?Geophysical Research Letters50(2023). URL https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2023GL105747

    doi:10.1029/2023gl105747 2023

  25. [25]

    Hakim, G. J. & Masanam, S. Dynamical Tests of a Deep Learning Weather Prediction Model (2024). URL https://journals.ametsoc.org/view/journals/aies/ 3/3/AIES-D-23-0090.1.xml

    2024

  26. [26]

    URL https://joss.theoj.org/papers/10.21105/joss.06323

    Kl¨ ower, M.et al.SpeedyWeather.jl: Reinventing atmospheric generalcirculation models towards interactivity and extensibility.Journal of Open Source Software 9, 6323 (2024). URL https://joss.theoj.org/papers/10.21105/joss.06323

    work page doi:10.21105/joss.06323 2024

  27. [27]

    D¨ uben, P. D. & Palmer, T. N. Benchmark Tests for Numerical Weather Forecasts on Inexact Hardware.Monthly Weather Review142, 3809– 3829 (2014). URL https://journals.ametsoc.org/view/journals/mwre/142/10/ mwr-d-14-00110.1.xml

    2014

  28. [28]

    & D¨ uben, P

    Chantry, M., Thornes, T., Palmer, T. & D¨ uben, P. Scale-Selective Preci- sion for Weather and Climate Forecasting.Monthly Weather Review147, 645–655 (2019). URL https://journals.ametsoc.org/view/journals/mwre/147/2/ mwr-d-18-0308.1.xml

    2019

  29. [29]

    M.et al.HEALPix: A Framework for High-Resolution Discretization and Fast Analysis of Data Distributed on the Sphere.The Astrophysical Journal 622, 759 (2005)

    G´ orski, K. M.et al.HEALPix: A Framework for High-Resolution Discretization and Fast Analysis of Data Distributed on the Sphere.The Astrophysical Journal 622, 759 (2005). URL https://iopscience.iop.org/article/10.1086/427976. 36

    doi:10.1086/427976 2005

  30. [30]

    URL https://ams.confex.com/ams/103ANNUAL/meetingapp.cgi/Paper/415842

    ARCO-ERA5: An Analysis-Ready Cloud-Optimized Reanalysis Dataset. URL https://ams.confex.com/ams/103ANNUAL/meetingapp.cgi/Paper/415842

  31. [31]

    & Hamman, J

    Hoyer, S. & Hamman, J. xarray: N-D labeled Arrays and Datasets in Python. Journal of Open Research Software5(2017). URL https://openresearchsoftware. metajnl.com/articles/10.5334/jors.148

    doi:10.5334/jors.148 2017

  32. [32]

    R.et al.Array programming with NumPy.Nature585, 357–362 (2020)

    Harris, C. R.et al.Array programming with NumPy.Nature585, 357–362 (2020). URL https://www.nature.com/articles/s41586-020-2649-2

    2020

  33. [33]

    Hunter, J. D. Matplotlib: A 2D Graphics Environment.Computing in Sci- ence & Engineering9, 90–95 (2007). URL http://ieeexplore.ieee.org/document/ 4160265/

    2007

  34. [34]

    Cartopy: a cartographic{Python}library with a{Matplotlib}interface (2010)

    Office, M. Cartopy: a cartographic{Python}library with a{Matplotlib}interface (2010). URL https://cartopy.readthedocs.io

    2010

  35. [35]

    URL https://github.com/MakieOrg/GeoMakie

    MakieOrg/GeoMakie.jl (2026). URL https://github.com/MakieOrg/GeoMakie. jl. Original-date: 2019-08-04T11:06:51Z

    2026

  36. [36]

    N.et al.The JASMIN super-data-cluster (2012)

    Lawrence, B. N.et al.The JASMIN super-data-cluster (2012). URL http:// arxiv.org/abs/1204.3553. ArXiv:1204.3553 [cs.DC]

    Pith/arXiv arXiv 2012

  37. [37]

    URL https://www.jasmin.ac.uk/about/

    JASMIN Site. URL https://www.jasmin.ac.uk/about/. 37