acceptodds
Under review as a conference paper at ICLR 2027

PEAR: Equal Area Weather Forecasting on the Sphere

Abstract

Artificial intelligence is rapidly reshaping the natural sciences, with weather forecasting emerging as a flagship AI4Science application where leading data-driven models now outperform traditional numerical methods across a broad range of global medium-range forecasting tasks. A common design choice in many of these models is an equiangular discretization of the sphere, which has much smaller grid cells 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. Motivated by the growing adoption of this grid in meteorology and climate science, we make the case for using HEALPix as the native grid for data-driven weather forecasting. To this end, we introduce Pangu Equal ARea (PEAR), an efficient volumetric transformer with equal-area inputs, intermediate representations, and outputs. We evaluate PEAR using ERA5-Lite, a reduced subset of the ERA5 atmospheric reanalysis dataset, for forecast times of one to nine days. PEAR remains competitive with the strongest evaluated baselines at short forecasting horizons and outperforms them at longer horizons. Among the evaluated models, PEAR combines the highest average forecasting accuracy with the fastest inference, while requiring less memory than its closest competitors in accuracy. Spectral analysis shows that PEAR has lower large-scale forecast errors at longer forecasting horizons compared to baselines. Furthermore, we quantify learned rotational equivariance for PEAR trained on ERA5-Lite and more densely sampled ERA5 data, and demonstrate single-step climate-model emulation, predicting surface temperature and precipitation from emission forcings.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
deep learning, weather forecasting, transformers, AI4Science

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_5a5AOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PEAR. PEAR is a volumetric transformer for data driven weather forecasting that operates directly on the HEALPix grid. Utilizing an equal area discretization of the sphere allows the model to avoid the polar oversampling issue inherent in standard latitude and longitude grids. The architecture incorporates patch embedding, windowed attention with ring shifting, and hierarchical downsampling and upsampling directly on the one dimensional HEALPix index structure. The authors train and evaluate PEAR on the ERA5 Lite dataset, which is a highly reduced subset of the full ERA5 archive. They compare it against scaled down versions of leading models such as Pangu, GraphCast, FengWu, and AFNO. The submission claims PEAR achieves competitive short term and superior long term forecasting accuracy alongside faster inference times and reduced memory usage. Additional experiments evaluate spectral errors, rotational equivariance, and single step climate model emulation.

Critical strengths
  1. The motivation to use equal area grids natively from input to output to solve the polar singularity and oversampling issue is well justified. This approach aligns perfectly with current trends in climate science like the Destination Earth initiative.

  2. The architectural adaptations required to make windowed attention work on the one dimensional HEALPix array are elegantly implemented. The authors effectively handle ring shifting and the masking of physically disjoint regions.

  3. The paper provides a thorough set of downstream analyses. These include spectral error analysis, rotational equivariance tests, and climate model emulation.

Critical weaknesses
  1. The paper trains and evaluates on ERA5 Lite. This dataset consists only of daily samples over 11 years of training data. Modern artificial intelligence weather models derive their predictive power from massive datasets containing hourly or six hourly samples spanning 40 years. Claiming superiority in long term forecasting based on an extremely low data regime is scientifically unsound. Scaling laws in weather forecasting are steep. Performance on ERA5 Lite does not reliably extrapolate to real operational settings.

  2. The authors drastically scaled down the baselines to match the exceptionally small parameter count of PEAR. For example, they reduced GraphCast to 4.3 million parameters and Pangu to 11.4 million. The original GraphCast has roughly 36.7 million parameters and uses a level 6 icosahedral mesh. Bottlenecking these architectures to a toy size inherently handicaps designs that were specifically optimized for larger capacities and higher resolutions. This creates a deeply flawed and unequal comparison of model architectures.

  3. The core innovation of applying shifted window attention to HEALPix was already established by recent prior work like HEAL SWIN. Extending this concept to three dimensions by adding an explicit vertical dimension and applying it to weather data is an incremental engineering step. It does not represent a fundamental methodological breakthrough suitable for ICLR.

  4. ERA5 data is intrinsically generated on a reduced Gaussian grid and distributed on a standard latitude and longitude grid. Projecting this raw data onto HEALPix introduces interpolation artifacts and spatial smoothing. The paper evaluates PEAR on its native HEALPix grid but completely glosses over the information loss incurred during the initial data projection. This omission calls into question the fairness of all native grid metrics.

  5. The claims regarding fastest inference and less memory usage are tautological consequences of the experimental setup. The authors compare a highly simplified model against baselines that have been awkwardly squeezed into the same tight parameter budget. These efficiency gains are highly unlikely to hold when the model is scaled up to actual operational resolutions.

  6. The climate emulation experiment in Section 5.2 is a single step and fully supervised image translation task. It completely fails to evaluate autoregressive stability or temporal dynamics. Those elements are the core challenges of actual weather forecasting. The inclusion of this experiment feels disjointed and fails to strengthen the main claims regarding forecasting superiority.

Questions for authors
  1. How does PEAR perform when trained on the full ERA5 dataset containing 40 years of high frequency samples compared to the original and unscaled versions of GraphCast and Pangu?

  2. GraphCast already utilizes an internal representation that shares properties with equal area grids through its icosahedral multimesh. Why is HEALPix fundamentally superior to the icosahedral mesh? This question is especially critical given that your GraphCast baseline was severely restricted to a level 4 mesh instead of its native level 6 mesh.

  3. How did you quantify and mitigate the interpolation errors introduced when converting the raw ERA5 data to the HEALPix grid before training began?

  4. Could the superior performance at longer lead times simply be an artifact of PEAR having a smoother and more regularized representation? This smoothing from interpolation and a low parameter count might artificially penalize high frequency noise rather than demonstrating a genuinely better understanding of atmospheric dynamics.

2 comments

We thank the reviewer for recognising the motivation for a natively equal-area model and its connection to initiatives such as Destination Earth. We are glad that the reviewer found the ring shifting and masking of physically disjoint regions to be elegantly implemented, and appreciate that the thorough set of downstream analyses was highlighted.

Weaknesses

  1. We have further clarified that we evaluate PEAR and baselines in a controlled setting that is distinct from full-scale ERA5, see lines 23, 280–281 and 481–483 in the updated rebuttal revision (changes in red).

  2. GraphCast-I was chosen from a ten-configuration width and depth sweep by matching PEAR's training-step time. GraphCast-P matches PEAR's parameter count. Our Pangu baseline has 11.4M parameters, 2.6 times larger than PEAR. The mesh level follows the data resolution. The original GraphCast pairs a level-6 mesh with about a million grid points, roughly 25 grid points per mesh node. Our level-4 mesh on 49,000 grid points gives about 19. We have additionally trained GraphCast-P on a level-5 mesh at the same parameter count. It takes 4.3 times the inference time and 3.7 times the peak memory of PEAR, see lines 307–309 and Appendix A.5 (Table A.4). With this mesh GraphCast-P leads PEAR at shorter lead times and PEAR leads at 9 days (PEAR's RMSE is 4.9 % higher at day 1 and 2.2 % lower at day 9).

  3. HEAL-SWIN introduced a shifted-window transformer on HEALPix for spherical vision tasks, and we use its ring shifting along the spherical directions (line 215). The following architectural elements are specific to PEAR (Section 4.2, lines 198–245, Figures B.1 and B.2).

    • Volumetric tokens. The patch embedding groups 16 HEALPix pixels and two neighbouring pressure levels, with a separate embedding for the surface variables. This gives eight vertical token positions.
    • Volumetric windows. Attention is performed over windows of 64 horizontal tokens and two vertical positions.
    • Shifting and masking in three dimensions. The ring shift is combined with a cyclic shift in the vertical direction. This brings together voxels from the north and south poles, and from the lowest and highest levels. The attention mask separates these regions.
    • A learned relative position bias over the volumetric window that is shared between all windows, in contrast to Pangu.
    • Native HEALPix inputs and outputs, with patch embedding and patch recovery by convolutions over the nested index.

    Appendix B.5 (Table B.8) quantifies these: PEAR has 45.6 / 35.5 / 14.4 % lower RMSE at 1 / 5 / 9 days than a variant without shifting, 6.4 / 4.2 / 1.5 % lower than a variant without relative position bias, and 20.4 / 12.9 / 4.9 % lower than a variant with levels as channels.

  4. We agree that we introduce interpolation artifacts when reprojecting the ERA5 data to the HEALPix grid. The 157 × 314 equiangular grid of the baselines is produced from the same 0.25° data by the same bilinear interpolation. This step is common to all models. See lines 290–291 and Appendix C.6 (Table C.9) for a quantitative estimate. To further strengthen the results, we also included results when evaluated on a common equiangular grid, see Appendix C.2 and Tables C.6 and C.7 in the manuscript, where the ranking of models persists.

  5. We compare inference time and peak memory against adapted versions of the baselines that match internal resolution, parameter count or inference cost. See also our response to weakness 1.

  6. We agree that this experiment is not about forecasting and does not test autoregressive stability. We do not claim that this section supports superior forecasting, but rather that it shows the applicability of the model to climate emulation for direct temperature and precipitation prediction from forcing variables.

Questions

  1. A single Pangu-Weather model requires about 74,000 GPU-hours, which exceeds our allocation for a year, and a fair comparison requires several models and seeds. We now also compare at 16 times higher resolution (HEALPix nside 256, one seed per model), where PEAR has 0.9 / 8.7 / 6.5 % lower RMSE at 1 / 5 / 9 days than Pangu-Weather-Lite retrained with the same training settings (Appendix H, Table H.1).
  2. The cells of a refined icosahedral mesh are not equal in area, while HEALPix cells are. HEALPix pixels lie on iso-latitude rings in a nested hierarchy, which reduces window partitioning and patching to reshaping of contiguous index ranges. We would also like to point out that the GraphCast mesh level was chosen to match the internal resolution of the models (2,562 mesh nodes at level 4 compared to 3,072 internal locations for PEAR). See our response to weakness 2 for a level-5 mesh.
  3. See our response to weakness 4.
  4. PEAR forecasts have about twice the fine-scale spectral power of GraphCast forecasts at days 1 and 9 (Appendix E, Figure E.1). The results at longer lead times are therefore not explained by smoother PEAR forecasts.

Dear reviewer,

Thank you again for your time and effort reviewing the manuscript. As the discussion phase is coming to an end, we wanted to check if our responses have addressed your main concerns. Let us know if there are any further clarifications needed.

Reviewer_SF3hOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PEAR combines a Pangu-style volumetric transformer with native HEALPix inputs, latent representations and outputs. It retains vertical tokens, uses spherical shifted windows and shares positional biases across windows. Five-seed experiments on ERA5-Lite compare one-to-nine-day forecasts against scaled weather-model baselines, with native- and common-grid evaluation, architectural controls and spectral diagnostics. The model offers a strong accuracy–inference-cost trade-off in this reduced-data regime. ClimateSet experiments provide supplementary evidence of applicability beyond weather forecasting.

Critical strengths
  1. The 4.316M-parameter model takes 13.60 ms per step versus 28.19 ms for parameter-matched GraphCast-P, with 0.253 versus 0.823 GiB peak memory (Table A.2). Five-seed results favor PEAR broadly at longer horizons. This is a useful systems contribution even without claiming full-scale operational state of the art.
  2. Appendix C evaluates both common grids, including nearest-cell transfer to the equiangular grid, where PEAR still leads in 444/621 RMSE comparisons. Latitude-weighted Pangu and shared-position Pangu controls retain substantial gaps. Spectral analysis distinguishes better large-scale skill from fine-scale performance, where GraphCast can be stronger. These checks make the central empirical observation more persuasive.
Critical weaknesses
  1. HEAL-SWIN supplies spherical shifted windows, DLWP-HPX already uses native HEALPix forecasting, and HEAL-ViT applies a transformer on an internal HEALPix mesh [1–3]. The manuscript acknowledges these distinctions, but neither close weather predecessor is evaluated. The vertical-patch-size-one ablation retains vertical tokens and therefore does not test the central difference from representing levels as channels. Existing ablations show that PEAR's components help its own implementation, but leave the value of native input/output plus explicit vertical tokenization only partially established.
  2. GraphCast-P is trained for 500 epochs and PEAR for 1,000; the main GraphCast and FengWu losses are spatially uniform on the latitude–longitude grid, whereas uniform HEALPix loss weights physical area. The weighted-Pangu control addresses this confound for Pangu, not GraphCast-P. Figure A.1 shows substantial flattening, so unequal epochs alone do not prove undertraining. Nevertheless, neither matched training cost nor area-weighted GraphCast-P is reported. This limits attribution to equal-area representation and the breadth of the Pareto claim, especially when the surface-field margins are small.

[1] Carlsson et al. HEAL-SWIN: A Vision Transformer On The Sphere. CVPR 2024. [2] Karlbauer et al. Advancing Parsimonious Deep Learning Weather Prediction Using the HEALPix Mesh. JAMES 2024. [3] Ramavajjala. HEAL-ViT: Vision Transformers on a spherical mesh for medium-range weather forecasting. 2024 preprint.

Questions for authors
  1. Does PEAR retain its advantage over GraphCast-P when both optimize an area-weighted loss and are compared at matched training compute, or after documented convergence under model-appropriate tuning? Existing checkpoint curves at common budgets would already help distinguish this from a mere request for longer training.
  2. Can you isolate explicit vertical tokens using a parameter- and compute-controlled HEALPix variant that treats levels as channels, or provide a suitably matched HEAL-ViT/DLWP-HPX comparison? Which empirical gain is specific to PEAR's contribution beyond established spherical processing?
  3. Would selecting checkpoints by standardized, rather than raw physical-unit, RMSE change the ranking against GraphCast-P? Equation A.2 averages quantities with different units; a check using saved checkpoints would clarify whether a few large-magnitude variables drive selection.
2 comments

We thank the reviewer for highlighting the accuracy and inference-cost trade-off as a useful systems contribution. We are glad that the common-grid evaluations, the Pangu controls and the spectral analysis were found to strengthen the central observation.

Weaknesses

  1. We thank the reviewer for the suggestion of ablating levels as channels. We have trained a model with the 13 pressure levels folded into the channels of a single token per HEALPix patch, i.e. without any vertical attention. This version has 4.5M parameters compared to 4.3M for PEAR and was trained with the same 1000 epoch budget. PEAR outperforms this model at all lead times, with a larger gap at shorter horizons (20.4 / 12.9 / 4.9 % lower RMSE at 1 / 5 / 9 days, in all 25 seed pairs). See Appendix B.5 (Table B.8, see line 304) in the updated rebuttal revision (changed in red). We have also included DLWP-HPX in our baseline comparison (PhysicsNeMo implementation, one seed). PEAR has 23.0 / 29.7 / 14.8 % lower RMSE at 1 / 5 / 9 days (Figures 1, 3, 4 and 5, Table 1 and Appendix Table A.3). These results are provisional, with more seeds to be included in a revised version. HEAL-ViT has no published code or weights, nor is the paper published. In accordance with the ICLR guidelines we have not included it as a baseline.
  2. We thank the reviewer for pointing this out. We have now retrained GraphCast-P with area-weighted loss, with each seed trained until its validation score plateaued (500 to 600 epochs). This version of GraphCast-P leads PEAR at shorter lead times and PEAR leads at 9 days (PEAR's RMSE is 4.6 % higher at day 1, 0.5 % higher at day 5 and 1.2 % lower at day 9). We have revised the statements pertaining to the more compute expensive GraphCast-P (2x inference time and 3x peak memory) in the updated manuscript, see lines 25–27, 316–317 and 494–498, and Appendix C.5 (Table C.8).

Questions

  1. With both models trained until their validation score plateaued and GraphCast-P optimizing an area-weighted loss, PEAR only leads at longer forecast lead times (PEAR's RMSE is 4.6 % higher at day 1, 0.5 % higher at day 5 and 1.2 % lower at day 9). GraphCast-P with 2x inference time and 3x peak memory has better forecasting accuracy at shorter lead times. For GraphCast-I, which is matched to PEAR's training-step time, the same area-weighted control (five seeds, 1000 epochs) leaves PEAR ahead at all lead times, with 2.6 / 3.6 / 1.9 % lower RMSE at 1 / 5 / 9 days (5.1 / 5.4 / 3.9 % against the uniform-loss GraphCast-I), see Appendix C.5 (Table C.8). We have revised the corresponding statements in the updated manuscript, see lines 25–27, 316–317 and 494–498.
  2. We have trained a HEALPix variant of PEAR that treats the 13 pressure levels as channels of a single token per patch. This version has 4.5M parameters compared to 4.3M for PEAR, 0.84 times the training-step time and the same 1000 epoch budget. PEAR has 20.4 / 12.9 / 4.9 % lower RMSE at 1 / 5 / 9 days in all 25 seed pairs, which isolates the gain from explicit vertical tokens (Appendix B.5, Table B.8). We have also included DLWP-HPX in our baseline comparison (PhysicsNeMo implementation, one seed), where PEAR has 23.0 / 29.7 / 14.8 % lower RMSE at 1 / 5 / 9 days. These results are provisional, with more seeds to be included in a revised version. HEAL-ViT has no published code or weights.
  3. Selecting by standardized RMSE does change the selected checkpoints (33 of 45 seeds) but leaves the ranking unchanged. We include standardized selection results in Appendix A.2 (Table A.3), see line 296.

Dear reviewer,

Thank you again for your time and effort reviewing the manuscript. As the discussion phase is coming to an end, we wanted to check if our responses have addressed your main concerns. Let us know if there are any further clarifications needed.

Reviewer_7sShOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PEAR, a volumetric weather-forecasting transformer that operates natively on the HEALPix equal-area grid. It combines spherical shifted-window attention with an explicit vertical dimension and hierarchical spatial processing. Experiments on ERA5-Lite show competitive short-range accuracy, stronger longer-range performance, and favorable inference efficiency compared with retrained forecasting baselines. Additional experiments examine spectral errors, learned rotational equivariance, and climate-model emulation.

Critical strengths
  • The integration of native HEALPix representations with volumetric attention is well motivated. Preserving the vertical structure while supporting efficient window operations and hierarchical processing provides a useful architectural contribution.

  • The evaluation includes five training seeds, parameter- and speed-matched GraphCast variants, and common-grid comparisons. PEAR’s longer-range gains and favorable inference cost provide encouraging evidence within the evaluated setting.

Critical weaknesses
  • The training comparison does not sufficiently establish that the baselines received comparable optimization opportunities. PEAR is trained for 1,000 epochs, compared with 500 for GraphCast-P and 300 for Pangu, with additional batch-size differences. Validation curves and checkpoint selection are helpful, but do not establish whether the ranking would persist under matched training compute or stronger baseline tuning. Total training costs and hardware are also insufficiently documented.

  • The contribution of equal-area discretization is not fully isolated from other architectural changes. PEAR and Pangu differ in embedding width, head count, and positional-bias design. The latitude-weighted-loss and shared-position controls partially address this concern, but a more closely matched comparison is needed to attribute the gains specifically to native HEALPix processing. Several architectural ablations also use the earlier PEAR-HD8 configuration rather than the final model.

  • The evidence remains limited to a relatively small data and resolution regime: roughly 49,000 horizontal grid points, daily training samples, and one test year. This supports the proposed design in a controlled setting, but leaves its accuracy and efficiency advantages at higher resolutions and larger training scales unresolved. Comparisons with scaled-down architectures should therefore be distinguished clearly from comparisons with their full-scale counterparts.

Questions for authors
  • Could you report training hardware and GPU-hours, and provide matched-compute results or extended training and tuning for the strongest baselines?

  • Can a closer PEAR–Pangu control match width, depth, head count, positional biases, and training settings to better isolate the effect of the grid?

  • Do the gains persist at higher spatial resolutions, with denser training data, or across additional test years?

2 comments

We thank the reviewer for recognising the motivation behind combining native HEALPix representations with volumetric attention. We also appreciate the positive assessment of the evaluation with matched GraphCast variants and common-grid comparisons.

Weaknesses

  1. All seeds have been continued until their validation score plateaued (less than 0.25 % improvement over 100 epochs, at most 1000 epochs) with model ranking persisting for the baselines of the submission. The area-weighted GraphCast-P control requested by Reviewer SF3h leads PEAR at shorter lead times (PEAR's RMSE is 4.6 % higher at day 1 and 1.2 % lower at day 9) with 2x inference time and 3x peak memory, see Appendix C.5 (Table C.8). The updated compute costs are reported in Appendix A.2 (Table A.1: final epochs and GPU-hours per seed, Table A.3: results), see lines 281–282 in the updated rebuttal revision (changed in red).
  2. PEAR-HD8 is the matched control against Pangu-Weather with the same embedding width and head counts. Against Pangu trained with latitude weighted loss and shared positional bias PEAR-HD8 has 12.5 / 15.3 / 5.1 % lower RMSE at 1 / 5 / 9 days. We have clarified the relationship between the best PEAR hyperparameters and the matched controls (lines 303–304 and Appendix B.3). For completeness, we now also include the ablations relative to the final PEAR architecture, see Appendix B.5 (Table B.8).
  3. We agree and have clarified our controlled setting in relation to their full-scale counterpart. See lines 23, 280–281 and 481–483.

Questions

  1. All baselines have been continued until their validation score plateaued, with model ranking persisting. Training hardware, final epochs and GPU-hours per seed are reported in Appendix A.2 (Tables A.1 and A.3).
  2. PEAR-HD8 is this control. It matches the embedding width, depth and head counts of our Pangu baseline and uses the same optimizer settings. Against Pangu trained with latitude weighted loss and shared positional bias PEAR-HD8 has 12.5 / 15.3 / 5.1 % lower RMSE at 1 / 5 / 9 days (lines 303–304 and Appendix B.3).
  3. We now also compare at 16 times higher resolution (HEALPix nside 256, 786,432 cells, one seed per model). Against Pangu-Weather-Lite retrained on the same data with the same training settings, PEAR has 0.9 / 8.7 / 6.5 % lower RMSE at 1 / 5 / 9 days (Appendix H, Table H.1, see line 483).

Dear reviewer,

Thank you again for your time and effort reviewing the manuscript. As the discussion phase is coming to an end, we wanted to check if our responses have addressed your main concerns and, if so, you would consider updating your score. Let us know if there are any further clarifications needed.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.