Pith. sign in

REVIEW 3 major objections 5 minor 43 references

PEAR: Equal Area Weather Forecasting on the Sphere

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A weather transformer trained natively on the equal-area HEALPix sphere grid beats the same architecture on the latitude–longitude grid, and matches or beats a model almost eight times larger out to ten days.

desk verdict A genuinely new HEALPix-native transformer for weather forecasting with strong, self-consistent results—but the equal-area grid is not yet isolated as the cause of the gains. read the letter →

arxiv 2505.17720 v3 pith:C4FHRVIF submitted 2025-05-23 cs.LG physics.ao-ph

Hampus Linander , Tage Tykesson , Pietro Rosso , Christoffer Petersson , Daniel Persson , Jan E. Gerken This is my paper
classification cs.LGphysics.ao-ph
keywords weatherforecastingHEALPixequal-areasphericalgridtransformermedium-rangeERA5pixelization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that the equiangular latitude–longitude grid used by most learned weather forecasters introduces a bad inductive bias: its cells shrink toward the poles, so models spend capacity on oversampled poles and require latitude weighting. The authors build PEAR, a transformer that runs entirely on HEALPix, a spherical pixelization in which every cell covers the same surface area. They train it on ERA5 reanalysis data resampled to HEALPix and report that it outperforms an otherwise comparable Pangu-style model on the equiangular grid at forecast horizons up to ten days, with better average ACC and RMSE on nearly all variables, while being smaller and faster. If the comparison is taken at face value, this shows that an equal-area discretization alone can improve learned weather forecasting without any extra compute.

What carries the argument

The load-bearing object is the HEALPix grid with its two index orders. In the nested ordering, blocks of four consecutive pixels correspond to one coarser pixel, so patch embedding, window partitioning, and downsampling become contiguous tensor reshapes; in the ring ordering, a cyclic roll of the pixel list rotates the sphere around the polar axis, giving shifted-window attention with masks at the poles. Because the cells have equal area, the model can share a single learned relative-position embedding across all windows, removing the need for latitude-dependent weighting. The architecture is a volumetric SWIN-style transformer whose input, latent, and output tensors all live on the $12 n_{\mathrm{side}}^2$ HEALPix pixels.

What would settle it

Train PEAR on the equiangular latitude–longitude grid using the same simplified relative positional embedding and the same masked ring-shift attention, keeping hyperparameters identical; if that model matches PEAR's ACC and RMSE, then the grid is not the driver. Alternatively, retrain the equiangular baseline with HEALPix resampling and the original positional embedding to see whether the advantage disappears.

Watch

Extended reading notes

Core claim

The paper's central claim is that the geometry of the discretization matters for learned medium-range weather forecasting: replacing the equiangular latitude–longitude grid with HEALPix, and keeping the model on HEALPix throughout, yields a transformer that outperforms the same architecture on the equiangular grid and matches or beats a larger equiangular model at five days and beyond. The reported numbers show PEAR with 4.3 million parameters beating the 11.4-million-parameter Pangu baseline on most variables and often beating the 33.7-million-parameter Pangu-Large baseline, at 1.5 and 3.2 times faster inference respectively. The paper also shows that the equal-area grid makes latitude weighting in the loss and evaluation metrics unnecessary, because each pixel already represents the same physical area.

Load-bearing premise

The headline comparison isolates the grid only if PEAR and the equiangular baselines differ solely by discretization and the expected resampling; the paper itself notes that PEAR uses a simplified learned relative positional embedding and relies on a third-party baseline implementation, so the accuracy gap could in principle come from those choices rather than from equal-area cells.

Editorial extensions

If this is right

  • If the grid alone drives the gain, then next-generation HEALPix-native weather data, such as the planned digital twin data, should improve learned forecasting further because no resampling artifacts are introduced.
  • Equal-area discretization lets future models drop spatial weights and simplify evaluation, since every pixel represents the same physical area on the sphere.
  • The reported speed and parameter efficiency suggest that equal-area grids could make high-resolution learned forecasting cheaper, not just more accurate.
  • PEAR's improvements persist and grow at longer lead times, which matters for medium-range forecasts where error accumulation is the main challenge.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A cleaner test of the paper's thesis would train PEAR on the equiangular grid with the same simplified relative-positional embedding and the same masked ring-shift attention, so that grid choice is the only difference; the paper does not report this ablation.
  • The abstract promises a check on climate-model emulation, but the main text does not report such experiments, so that part of the claim should be sought in the released code or future work before being weighed.
  • If the equal-area advantage is real, it should also hold for probabilistic and ensemble forecasting models, where pole oversampling currently inflates both compute and variance estimates.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces PEAR, a transformer-based weather forecasting model that operates natively on the HEALPix spherical pixelization. The authors argue that equal-area HEALPix cells remove the unphysical latitude-dependent resolution of equiangular grids used by models like Pangu-Weather. PEAR is compared against a reimplementation of Pangu and a larger Pangu-Large baseline on the ERA5-lite dataset, with reported ACC and RMSE improvements at lead times up to 10 days, along with lower parameter counts and faster inference. The paper also claims experiments on equivariance and climate model emulation, and provides a GitHub repository for the implementation.

Significance. If the central claim is established, the work would provide a practical demonstration that equal-area spherical discretization can benefit learned weather forecasting, potentially motivating further use of HEALPix in operational-style models. The paper is transparent about its computational constraints and makes code available, which supports reproducibility. However, the current evidence does not isolate the effect of the HEALPix grid from other architectural changes, and the baseline is a third-party reimplementation whose fidelity is not verified. The significance is thus contingent on additional controlled experiments.

major comments (3)
  1. [§4.2, §5, Table 1] The central claim that PEAR outperforms the 'corresponding model on an equiangular grid' due to the HEALPix discretization is not supported because PEAR differs from the Pangu baseline in several dimensions. Section 4.2 explicitly states 'In contrast to Pangu [9], we use a simplified learned relative positional embedding' and notes that this embedding 'accounts for most of the parameter savings compared to Pangu in Table 1.' Table 1 shows PEAR has 4.3M parameters versus 11.4M for Pangu. Since the paper's thesis is that the equal-area grid is responsible for the accuracy gain, the absence of an ablation that isolates the grid choice from the positional-embedding change is load-bearing. A minimal fix would be to train an equiangular-grid version of PEAR with the same simplified positional embedding, or a HEALPix version with the original Pangu-style embedding, and show that the accuracy gap persists when only the grid changes.
  2. [§5, ref [39]] The baseline 'Pangu' is not the original Pangu-Weather model or its official implementation, but a third-party reimplementation from the WeatherLearn repository. The paper does not provide evidence that this reimplementation reproduces the original Pangu-Weather's behavior or accuracy on ERA5-lite. If the reimplementation is weaker than the official model, the comparison would be biased in favor of PEAR. The authors should either use the official Pangu code (or the authors' own faithful reimplementation) and report its performance, or validate WeatherLearn against published Pangu-Weather results on the same dataset and protocol.
  3. [Abstract, main text] The abstract states that the authors 'perform numerical experiments on the equivariance properties of our setup and verify the performance of PEAR on climate model emulation,' but the manuscript as provided contains no such experiments or results. The main text and appendix discuss window shifting and masking but do not report equivariance experiments, and there is no section on climate model emulation. These promised contributions are missing and must either be added or removed from the abstract.
minor comments (5)
  1. [§3.2] The text uses 'ECWMF' which should be 'ECMWF' (European Centre for Medium-Range Weather Forecasts).
  2. [Figure 1 and Figure 3 captions] The phrase 'course-graining' should be 'coarse-graining.'
  3. [§5] The citation 'Pangu [39]' refers to the WeatherLearn repository, not the original Pangu-Weather paper [9]. This is confusing because [9] is the original Pangu-Weather publication. Please clarify that the baseline is a reimplementation and cite both.
  4. [Appendix, Eqs. (1) and (2)] The typesetting of the RMSE and ACC equations appears garbled in the provided version; the summation limits and fraction bars should be checked for correctness in the final PDF.
  5. [§5, Figure 4] The claim that PEAR 'outperforms Pangu-Large' at longer lead times is stronger than what Figure 4 shows: for t2m, PEAR appears comparable or slightly below Pangu-Large across the lead times. The text later says 'better (msl, u10, v10, t, u, v, z) or comparable (t2m),' which is more accurate. Consider harmonizing the wording.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: PEAR's accuracy claims rest on held-out ERA5 evaluation and an external Pangu baseline, not on inputs that reduce to the outputs.

full rationale

PEAR's central comparison is a standard empirical pipeline: the model is trained on ERA5-lite years 2007-2017 and evaluated on 2019 using RMSE and ACC against reanalysis targets (Section 5 and Appendix A). The reported metrics are measured against independent validation data; no fitted parameter is renamed as a prediction, and the accuracy numbers are not identities or rearranged training objectives. The only apparent self-citation, HEAL-SWIN [15], is used for the ring-shifting strategy in the windowed attention implementation (Section 4.2); it is a reproducible architectural detail and is not invoked to justify PEAR's forecasting skill. The paper's own statement that PEAR uses a 'simplified learned relative positional embedding' while Pangu shares hyperparameters is a legitimate attribution concern: the comparison does not isolate the HEALPix grid from the positional-embedding change. That is a confound or external-validity issue, not circular reasoning, because the evaluation remains grounded in independent ERA5 data rather than being derived from the paper's own assumptions. I therefore find no circular step.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The model's performance claim requires the chosen grid resolution, loss weighting, training data years, and the equality of all architectural components except the grid; the last is explicitly contradicted by the simplified positional embedding, so the ledger includes it as an ad hoc assumption.

free parameters (6)
  • Grid resolution nside = 64
    Sets the 12*64^2 = 49,152-pixel state; chosen so native ERA5 cells are finer than HEALPix pixels. Not fitted to the loss but determines all downstream resolution choices.
  • Surface loss weight = 1/4
    Weight applied to surface variable L1 loss during training, following Pangu; trades off surface versus upper-air accuracy.
  • Optimizer hyperparameters = AdamW, lr=5e-4, weight decay=3e-6
    Fixed without reported tuning; results could depend on these values.
  • Training split = 2007-2017 train, 2019 validation
    ERA5-lite subset with 4017 daily samples; a small training set that may inflate the measured performance gap.
  • Relative positional embedding = learned B tensor, shape (1, Nheads, (Wd*Whp)^2)
    A simplified shared positional embedding used instead of Pangu's; the paper says this accounts for most of PEAR's parameter savings and is a confound in the grid comparison.
  • Architecture hyperparameters (depths, heads, embeddings) = See Table A1 (e.g., 2+12+2 attention blocks, 48/96 dims, 6/12 heads)
    These choices follow Pangu and are not tuned here; they determine parameter counts and capacity.
assumptions (5)
  • domain assumption ERA5 reanalysis is the ground truth for weather state
    Used as target and reference for all metrics; assumes reanalysis has negligible error for this purpose.
  • domain assumption Resampling ERA5 from its native 0.25 degrees equiangular grid to HEALPix nside=64 introduces no significant loss
    The paper argues ERA5 cells are smaller than HEALPix pixels everywhere, but does not quantify resampling artifacts.
  • domain assumption The WeatherLearn implementation faithfully reproduces Pangu-Weather
    Baseline comparisons are against a third-party reimplementation (ref [39]), not the original Pangu code; any discrepancy changes the comparison.
  • ad hoc to paper PEAR and Pangu share the same architecture aside from the grid
    Section 4.2 shows PEAR uses a simplified relative positional embedding, so 'same architecture hyperparameters' (Section 5) does not mean identical architecture; the paper treats the grid as the only meaningful difference.
  • standard math Climatology over 11 daily years is a valid reference for ACC
    Standard practice, but with only 11 samples per calendar day, the climatology estimate has noise.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PEAR: Equal Area Weather Forecasting on the Sphere." pith.science (2026). https://pith.science/paper/C4FHRVIF

@misc{pith2026250517720,
  author       = {Pith},
  title        = {Pith review of: PEAR: Equal Area Weather Forecasting on the Sphere},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C4FHRVIF}},
  note         = {Machine review of arXiv:2505.17720}
}
read the original abstract

Artificial intelligence is rapidly reshaping the natural sciences, with weather forecasting emerging as a flagship AI4Science application where machine learning models can now rival and even surpass traditional numerical simulations. Following the success of the landmark models Pangu Weather and Graphcast, outperforming traditional numerical methods for global medium-range forecasting, many novel data-driven methods have emerged. A common limitation shared by many of these models is their reliance on an equiangular discretization of the sphere which suffers from a much finer grid at the poles than around the equator. In contrast, in the Hierarchical Equal Area iso-Latitude Pixelization (HEALPix) of the sphere, each pixel covers the same surface area, removing unphysical biases. Motivated by a growing support for this grid in meteorology and climate sciences, we propose to perform weather forecasting with deep learning models which natively operate on the HEALPix grid. To this end, we introduce Pangu Equal ARea (PEAR), a transformer-based weather forecasting model which operates directly on HEALPix-features and outperforms the corresponding model on an equiangular grid, and other baselines, without any computational overhead. Furthermore, we perform numerical experiments on the equivariance properties of our setup and verify the performance of PEAR on climate model emulation.

Figures

Figures reproduced from arXiv: 2505.17720 by the authors.

Figure 1
Figure 1. Left: Predicted surface level temperature from PEAR. Green lines show the HEALPix cell [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Shift and corresponding mask for windowed attention. Illustration of a scalar tensor with [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. PEAR architecture schematic. Violet slices correspond to the variables visualized on the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Mean anomaly correlation coefficient (ACC), higher is better, for the surface and upper [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 19 canonical work pages

  1. [39]

    Zhuoqun Li.Lizhuoq/WeatherLearn. Apr. 2025.URL: https://github.com/lizhuoq/WeatherLear n

    work page 2025

  2. [12]

    Advancing Parsimonious Deep Learning Weather Prediction Using the HEALPix Mesh

    Matthias Karlbauer et al. “Advancing Parsimonious Deep Learning Weather Prediction Using the HEALPix Mesh”. In:Journal of Advances in Modeling Earth Systems16.8 (2024), e2023MS004021. ISSN: 1942-2466.DOI:10.1029/2023MS004021. arXiv:2311.06253

    arXiv 2024

  3. [9]

    Accurate Medium-Range Global Weather Forecasting with 3D Neural Networks

    Kaifeng Bi et al. “Accurate Medium-Range Global Weather Forecasting with 3D Neural Networks”. In: Nature619.7970 (July 2023), pp. 533–538.ISSN: 1476-4687.DOI:10.1038/s41586-023-06185-3

    doi:10.1038/s41586-023-06185-3 2023

  4. [1]

    Deep learning for spatiotemporal forecasting in Earth system science: a review

    Manzhu Yu, Qunying Huang, and Zhenlong Li. “Deep learning for spatiotemporal forecasting in Earth system science: a review”. In:International Journal of Digital Earth17.1 (2024), p. 2391952

    work page 2024

  5. [2]

    Foundation Models for Geophysics: Review and Perspective

    Qi Liu and Jianwei Ma. “Foundation Models for Geophysics: Review and Perspective”. In:arXiv preprint arXiv:2406.03163(2024)

    arXiv 2024

  6. [3]

    A comprehensive review of methods for hydrological forecasting based on deep learning

    Xinfeng Zhao et al. “A comprehensive review of methods for hydrological forecasting based on deep learning”. In:Water16.10 (2024), p. 1407

    work page 2024

  7. [4]

    Learning Skillful Medium-Range Global Weather Forecasting

    Remi Lam et al. “Learning Skillful Medium-Range Global Weather Forecasting”. In:Science382.6677 (Dec. 2023), pp. 1416–1421.DOI:10.1126/science.adi2336. arXiv:2212.12794

    arXiv 2023

  8. [5]

    FuXi: A Cascade Machine Learning Forecasting System for 15-Day Global Weather Forecast

    Lei Chen et al. “FuXi: A Cascade Machine Learning Forecasting System for 15-Day Global Weather Forecast”. In:npj Climate and Atmospheric Science6.1 (Nov. 2023), pp. 1–11.ISSN: 2397-3722.DOI: 10.1038/s41612-023-00512-1. arXiv:2306.12873

    arXiv 2023

Show all 43 references
  1. [6]

    Kang Chen et al.Fengwu: Pushing the skillful global medium-range weather forecast beyond 10 days lead. 2023. arXiv:2304.02948

    2023

  2. [7]

    July 2023

    Marcin Andrychowicz et al.Deep Learning for Day Forecasts from Sparse Observations. July 2023. arXiv:2306.06079

    2023 arXiv

  3. [8]

    Scaling Transformer Neural Networks for Skillful and Reliable Medium-Range Weather Forecasting

    Tung Nguyen et al. “Scaling Transformer Neural Networks for Skillful and Reliable Medium-Range Weather Forecasting”. In:Advances in Neural Information Processing Systems. Ed. by A. Globerson et al. V ol. 37. Curran Associates, Inc., 2024, pp. 68740–68771. arXiv:2312.03876

    2024 arXiv

  4. [10]

    The ERA5 global reanalysis

    Hans Hersbach et al. “The ERA5 global reanalysis”. In:Quarterly journal of the royal meteorological society146.730 (2020), pp. 1999–2049

    2020

  5. [11]

    Analysis Issues for Large CMB Data Sets

    K. M. Gorski, E. Hivon, and B. D. Wandelt. “Analysis Issues for Large CMB Data Sets”. In:Arxiv eprints arXiv:astro-ph/9812350(1998). eprint:arXiv:astro-ph/9812350

    1998 arXiv

  6. [13]

    Vivek Ramavajjala.HEAL-ViT: Vision Transformers on a Spherical Mesh for Medium-Range Weather Forecasting. Feb. 2024. arXiv:2403.17016

    2024 arXiv

  7. [14]

    DeepSphere: Efficient spherical convolutional neural network with HEALPix sampling for cosmological applications

    N. Perraudin et al. “DeepSphere: Efficient spherical convolutional neural network with HEALPix sampling for cosmological applications”. In:Astronomy and Computing27 (Apr. 2019), pp. 130–146. ISSN: 2213-1337.DOI: 10.1016/j.ascom.2019.03.004 .URL: http://dx.doi.org/10.1016/j. as...

    2019 doi

  8. [15]

    HEAL-SWIN: A Vision Transformer On The Sphere

    Oscar Carlsson et al. “HEAL-SWIN: A Vision Transformer On The Sphere”. In:Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. July 2023, pp. 6067–6077. arXiv: 2307.07313

    2023 arXiv

  9. [16]

    The Destination Earth digital twin for climate change adaptation

    F. J. Doblas-Reyes et al. “The Destination Earth digital twin for climate change adaptation”. In:EGU- sphere2025 (2025), pp. 1–41.DOI: 10.5194/egusphere-2025-2198 .URL: https://egusphere. copernicus.org/preprints/2025/egusphere-2025-2198/

    2025 doi

  10. [17]

    Destination Earth – A digital twin in support of climate services

    Jörn Hoffmann et al. “Destination Earth – A digital twin in support of climate services”. In:Climate Services30 (2023), p. 100394.ISSN: 2405-8807.DOI: https://doi.org/10.1016/j.cliser.2023. 100394.URL:https://www.sciencedirect.com/science/article/pii/S2405880723000559

    2023 doi

  11. [18]

    https://www.ecmwf.int/en/newsletter/175/computing/migration-grib1-grib2-preparing-ecmwf- model-output-future

    Migration from GRIB1 to GRIB2: Preparing ECMWF Model Output for the Future. https://www.ecmwf.int/en/newsletter/175/computing/migration-grib1-grib2-preparing-ecmwf- model-output-future. Text. Apr. 2023

    2023

  12. [19]

    Hailong Shu et al.Forecasting the Future with Future Technologies: Advancements in Large Meteorolog- ical Models. Apr. 2024. arXiv:2404.06668

    2024 arXiv

  13. [20]

    Advances and Prospects of Deep Learning for Medium-Range Extreme Weather Forecasting

    Leonardo Olivetti and Gabriele Messori. “Advances and Prospects of Deep Learning for Medium-Range Extreme Weather Forecasting”. In:Geoscientific Model Development17.6 (Mar. 2024), pp. 2347–2358. ISSN: 1991-959X.DOI:10.5194/gmd-17-2347-2024. 9

    2024 doi

  14. [21]

    Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows

    Ze Liu et al. “Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows”. In:Pro- ceedings of the IEEE/CVF International Conference on Computer Vision. Aug. 2021, pp. 10012–10022. arXiv:2103.14030

    2021 arXiv

  15. [22]

    Simon Lang et al.AIFS – ECMWF’s Data-Driven Forecasting System. Aug. 2024. arXiv: 2406.01465

    2024 arXiv

  16. [23]

    FourCastNet: Accelerating Global High-Resolution Weather Forecasting Using Adaptive Fourier Neural Operators

    Thorsten Kurth et al. “FourCastNet: Accelerating Global High-Resolution Weather Forecasting Using Adaptive Fourier Neural Operators”. In:Proceedings of the Platform for Advanced Scientific Computing Conference. Davos Switzerland: ACM, June 2023, pp. 1–11.ISBN: 9798400701900.DO...

    2023

  17. [24]

    Spherical Fourier Neural Operators: Learning Stable Dynamics on the Sphere

    Boris Bonev et al. “Spherical Fourier Neural Operators: Learning Stable Dynamics on the Sphere”. In: Proceedings of the 40th International Conference on Machine Learning. PMLR, July 2023, pp. 2806–

    2023

  18. [25]

    Physics-Inspired Adaptions to Low-Parameter Neural Network Weather Forecast Systems

    Sebastian Scher and Gabriele Messori. “Physics-Inspired Adaptions to Low-Parameter Neural Network Weather Forecast Systems”. In:Artificial Intelligence for the Earth Systems3.1 (2024), e230046.DOI: 10.1175/AIES-D-23-0046.1. arXiv:2008.13524

    2024 arXiv

  19. [26]

    CirT: Global Subseasonal-to-Seasonal Forecasting with Geometry-Inspired Transformer

    Yang Liu et al. “CirT: Global Subseasonal-to-Seasonal Forecasting with Geometry-Inspired Transformer”. In:The Thirteenth International Conference on Learning Representations. 2025. arXiv:2502.19750

    2025 arXiv

  20. [27]

    Probabilistic Weather Forecasting with Hierarchical Graph Neural Networks

    Joel Oskarsson et al. “Probabilistic Weather Forecasting with Hierarchical Graph Neural Networks”. In:Advances in Neural Information Processing Systems. V ol. 37. Dec. 2024, pp. 41577–41648. arXiv: 2406.04759

    2024 arXiv

  21. [28]

    Probabilistic Weather Forecasting with Machine Learning

    Ilan Price et al. “Probabilistic Weather Forecasting with Machine Learning”. In:Nature(Dec. 2024), pp. 1–7.ISSN: 1476-4687.DOI:10.1038/s41586-024-08252-9

    2024 doi

  22. [29]

    Yingkai Sha et al.Improving AI Weather Prediction Models Using Global Mass and Energy Conservation Schemes. Jan. 2025. arXiv:2501.05648

    2025 arXiv

  23. [30]

    ClimaX: A Foundation Model for Weather and Climate

    Tung Nguyen et al. “ClimaX: A Foundation Model for Weather and Climate”. In:Proceedings of the 40th International Conference on Machine Learning. PMLR, July 2023, pp. 25904–25938

    2023

  24. [31]

    ACE: A Fast, Skillful Learned Global Atmospheric Model for Climate Prediction

    Oliver Watt-Meyer et al. “ACE: A Fast, Skillful Learned Global Atmospheric Model for Climate Prediction”. In:Tackling Climate Change with Machine Learning: Workshop at NeurIPS. arXiv, Oct

  25. [32]

    Cristian Bodnar et al.A Foundation Model for the Earth System. Nov. 2024. arXiv:2405.13063

    2024 arXiv

  26. [33]

    Analyzing and Exploring Training Recipes for Large-Scale Transformer-Based Weather Prediction

    Jared D. Willard et al. “Analyzing and Exploring Training Recipes for Large-Scale Transformer-Based Weather Prediction”. In:Artificial Intelligence for the Earth Systems4.2 (2025), p. 240061.DOI: 10. 1175/AIES-D-24-0061.1. arXiv:2404.19630

    2025 arXiv

  27. [34]

    John Schreck et al.Community Research Earth Digital Intelligence Twin (CREDIT). Nov. 2024. arXiv: 2411.07814

    2024

  28. [35]

    WeatherBench 2: A Benchmark for the Next Generation of Data-Driven Global Weather Models

    Stephan Rasp et al. “WeatherBench 2: A Benchmark for the Next Generation of Data-Driven Global Weather Models”. In:Journal of Advances in Modeling Earth Systems16.6 (2024), e2023MS004019. ISSN: 1942-2466.DOI:10.1029/2023MS004019. arXiv:2308.15560

    2024 arXiv

  29. [36]

    Improving Medium-Range Ensemble Weather Forecasts with Hierarchical Ensemble Transformers

    Zied Ben Bouallègue et al. “Improving Medium-Range Ensemble Weather Forecasts with Hierarchical Ensemble Transformers”. In:Artificial Intelligence for the Earth Systems3.1 (Jan. 2024), e230027.DOI: 10.1175/AIES-D-23-0027.1. arXiv:2303.17195

    2024 arXiv

  30. [37]

    Neural General Circulation Models for Weather and Climate

    Dmitrii Kochkov et al. “Neural General Circulation Models for Weather and Climate”. In:Nature 632.8027 (Aug. 2024), pp. 1060–1066.ISSN: 1476-4687.DOI: 10.1038/s41586-024-07744-y . arXiv: 2311.07222

    2024 arXiv

  31. [38]

    Swin Transformer V2: Scaling Up Capacity and Resolution

    Ze Liu et al. “Swin Transformer V2: Scaling Up Capacity and Resolution”. In:Proceedings of the IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022, pp. 11999–12009. arXiv:2111.09883

    2022 arXiv

  32. [40]

    The probability anomaly correlation and calibration of probabilistic forecasts

    Huug van den Dool et al. “The probability anomaly correlation and calibration of probabilistic forecasts”. In:Weather and Forecasting32.1 (2017), pp. 199–206

    2017

  33. [41]

    Oliver Watt-Meyer et al.ACE2: Accurately Learning Subseasonal to Decadal Atmospheric Variability and Forced Responses. Nov. 2024. arXiv:2411.11268

    2024 arXiv

  34. [42]

    Can Artificial Intelligence-Based Weather Prediction Models Simulate the Butterfly Effect?

    T. Selz and G. C. Craig. “Can Artificial Intelligence-Based Weather Prediction Models Simulate the Butterfly Effect?” In:Geophysical Research Letters50.20 (2023), e2023GL105747.ISSN: 1944-8007. DOI:10.1029/2023GL105747

    2023 doi

  35. [43]

    On Some Limitations of Current Machine Learning Weather Prediction Models

    Massimo Bonavita. “On Some Limitations of Current Machine Learning Weather Prediction Models”. In:Geophysical Research Letters51.12 (2024), e2023GL107377.ISSN: 1944-8007.DOI: 10.1029/ 2023GL107377. 10 Table A1: PEAR architecture overview. The windowed attention blocks contain ...

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.