acceptodds
Under review as a conference paper at ICLR 2027

Symmetric Composition of Anisotropic Operators for Global Subseasonal-to-Seasonal Climate Forecasting

Abstract

Accurate global Subseasonal-to-Seasonal (S2S) climate forecasting is critical for disaster preparedness and resource management, yet it remains challenging due to chaotic atmospheric dynamics. Despite advances in geometry-aware representations, existing methods do not specify how the zonal and meridional interactions that govern real atmospheric dynamics should be explicitly coordinated and composed, leaving an important architectural gap for S2S forecasting. In this paper, we propose **AnisoCast**, which explicitly models anisotropic global atmospheric dynamics for accurate S2S climate forecasting. It couples: (1) an Anisotropic Embedding strategy that tokenizes the global grid into latitudinal rings, preserving the integrity of zonal periodic structures; and (2) stacked Aniso Blocks that arrange latent Zonal and Meridional Operators in a weight-shared palindromic composition inspired by symmetric operator splitting. Extensive experiments on the ERA5 reanalysis dataset demonstrate that AnisoCast establishes a new state-of-the-art, significantly outperforming existing methods in both forecasting accuracy and computational efficiency.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Subseasonal-to-Seasonal Forecasting, Climate forecasting, Neural operators

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_eU46OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces AnisoCast, which encodes global meteorological fields into latitude-ring tokens and processes them through latent-channel spectral mixing and gated convolutions between neighboring latitudes. These operations are arranged in a weight-shared symmetric composition to directly predict mean fields for weeks 3–4 and weeks 5–6. The reported experiments show improved performance on several variables and lower inference costs among the evaluated models.

Critical strengths
  1. The architecture addresses directional differences while maintaining computational efficiency. Latitude-ring representations separate within-ring feature processing from cross-latitude interactions and avoid the quadratic cost of global attention, providing a practical design for direct S2S forecasting.

  2. The paper compares several model families and evaluates multiple variables, latitude bands, and pressure levels. Spatial error maps, component ablations, and efficiency measurements provide supporting evidence for the overall approach.

Critical weaknesses
  1. AnisoCast uses latitude-ring tokenization similar to CirT, which the authors acknowledge as an inspiration. I think the paper would benefit from a more direct comparison explaining what AnisoCast keeps from CirT, what it changes, and why those changes are needed for the problem studied here. This would make it easier to understand the contribution beyond the shared tokenization scheme. It would also help to connect these differences to the existing ablation results, so I can see which design choices account for the additional performance gains and how much each contributes. The paper already includes relevant experiments, but the link between the motivation for these changes and the evidence supporting them could be clearer.

  2. The FFT in the Zonal Operator acts on learned channels with no inherent longitude ordering, so its connection to physical zonal waves is unclear. Likewise, the linear Strang-splitting analysis does not establish that symmetric ordering reduces long-horizon errors in the actual network. Testing a non-palindromic ordering with the same operator counts and shared weights would help assess this specific claim. Without such evidence, these interpretations remain plausible explanations rather than demonstrated mechanisms.

Questions for authors
  1. What are the main differences between AnisoCast and CirT beyond the shared latitude-ring tokenization, and which existing ablation results support the need for these changes?

  2. What evidence connects the FFT over learned latent channels to physical zonal waves and longitudinal periodicity?

  3. Have you compared symmetric ordering with a non-palindromic ordering using the same operator counts and weight-sharing scheme to isolate its effect on long-horizon errors?

Reviewer_B5j2OpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces AnisoCast for global subseasonal-to-seasonal forecasting. It represents atmospheric fields as latitude-ring tokens and combines zonal and meridional operators through a weight-shared symmetric composition inspired by operator splitting. The model directly predicts mean fields for weeks 3–4 and 5–6. Experiments on ERA5 report competitive accuracy and computational efficiency against operational forecasting systems and data-driven baselines, supported by component ablations and robustness analyses.

Critical strengths
  1. The latitude-ring representation and directional operators provide a well-motivated inductive bias for atmospheric anisotropy, combining global within-ring mixing with local interactions across neighboring rings.

  2. The evaluation covers multiple meteorological variables and both forecast windows. Comparisons with representative baselines, component ablations, and analyses across random seeds, seasons, and spatial resolutions provide useful evidence of effectiveness and robustness.

  3. Directly predicting the target-window means aligns the model output with the forecasting objective and offers a practical formulation for extended-range prediction without requiring intermediate forecasts.

Critical weaknesses
  1. Appendix E.3 specifies a meridional kernel size of (k=7) without explaining its selection. Since this choice controls the range of cross-latitude interactions, a brief justification or sensitivity analysis would strengthen the architectural rationale.

  2. Figure 7 omits several details needed to assess the latency comparison, including warm-up, measurement repetitions, and the operations included in timing. These details matter for reproducibility and fair comparison.

Questions for authors
  1. How was (k=7) selected, and how sensitive is performance to nearby kernel sizes on representative variables?

  2. Could the authors specify the timing protocol for Figure 7, including warm-up, measurement repetitions, and whether preprocessing and data transfer are included?

1 comment

Thanks for the response. My concerns have been adequately addressed. After reading the other reviewers's comments, I support its acceptance and will keep my original score.

Reviewer_FekCOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AnisoCast, a lightweight model for global S2S forecasting. Each latitude ring is compressed into a single token, and the resulting ring tokens are processed by stacked blocks. Each block alternates a Fourier-domain filter applied along the channel dimension (the Zonal Operator) with a gated convolution along latitude (the Meridional Operator), in a weight-shared, symmetric arrangement inspired by Strang splitting. The model predicts the two-week mean fields for weeks 3–4 and weeks 5–6 from a single-day initial state and is trained with a latitude-weighted MSE loss.

Critical strengths
  1. The work rests on a physically well-motivated premise. Large-scale atmospheric dynamics are inherently anisotropic: the zonal direction is periodic and approximately invariant under rotation about the polar axis, whereas the meridional direction exhibits no such symmetry. Building this asymmetry into the architecture is a reasonable inductive bias.

  2. The architecture is notably simple, yet on the reported benchmark it achieves competitive or superior performance against more complex baselines, with reduced inference latency.

  3. The authors acknowledge the limits of their interpretation and provide a practical lightweight model for S2S forecasting.

Critical weaknesses
  1. The Aniso Block is essentially a MetaFormer-style combination of token mixing and channel mixing, with channel mixing implemented in the Fourier domain. The existing ablation that removes the Zonal Operator also adds a second Meridional Operator, so the contribution of the latent mixer remains unclear. Replacing it with an identity or a standard channel mixer of comparable size would isolate this effect.

  2. Two-week mean fields are relatively smooth and low-dimensional, which may favor this design. The appendix analyses of daily outputs, precipitation, and extremes are limited, so it is unclear whether the advantage holds for less smooth targets.

  3. According to the paper, CirT also uses latitude-ring tokens, so the main differences appear to lie in the token mixer, channel mixer, normalization, and block arrangement. Experiments isolating which of these drives the gain over CirT are needed.

  4. The Anisotropic Embedding maps each ring to channels through a learned linear map, so the channel index is not constrained to correspond to longitude or zonal wavenumber. The Zonal Operator’s Fourier transform along this axis therefore acts as structured channel mixing rather than a spatial spectral filter. This makes its interpretation as modeling zonal wave dynamics questionable.

  5. For several variables, AnisoCast’s weeks 5–6 scores are comparable to or better than its weeks 3–4 scores. Other direct models are similarly flat, whereas the NWP systems degrade with lead time. This suggests the direct models may be operating close to climatology, and that part of their advantage may come from comparing mean-regressing predictions against deterministic NWP forecasts. Please clarify whether the NWP scores come from control forecasts or ensemble means, and include climatology and persistence baselines.

Questions for authors
  1. The variant with unshared Zonal Operators performs consistently worse, even though the shared configuration is a special case of the unshared one. This suggests that the gain may come from the regularizing effect of weight tying rather than from the symmetric splitting structure itself. Could the authors comment on this?

  2. The zonal and meridional directions are indeed orthogonal, but extending this to an “orthogonal decomposition” and “orthogonal physical processes,” as in Figure 1, is not mathematically justified, since no inner product or decomposition of the state is defined. Furthermore, operators acting on orthogonal subspaces would commute, which conflicts with the non-commutativity used to motivate the symmetric ordering. Please clarify or revise this terminology.

  3. To verify that the model captures zonal structure, could the authors examine how a longitudinal rotation of the input is reflected in the latent representation? Alternatively, could they compare against a variant that applies the Fourier transform along the longitude axis?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.