acceptodds
Under review as a conference paper at ICLR 2027

AtmosWeave: Learning Atmospheric Dynamics for Global and Regional Weather Forecasting

Abstract

Data-driven weather forecasting learns atmospheric evolution directly from large-scale historical data and has achieved strong forecasting performance. Despite this progress, existing machine learning models still struggle to capture local interactions across spatial hierarchies, preserve directional variation, and efficiently model long-range dependencies. We introduce AtmosWeave, a graph-based weather forecasting framework that combines multiscale mesh propagation with efficient grouped linear attention. Specifically, Scale-Decoupled Directional Mesh Propagation separates within-level normalization from cross-level fusion and incorporates directional deviations into local propagation, while Geography- and State-Aware Grouped Linear Attention enables selective long-range interactions through source grouping and target-dependent routing. For high-resolution regional forecasting, we further introduce CrossWarp to couple global and regional representations through learned feature warping and selective global information injection, enabling the regional model to better leverage global context. Experiments demonstrate that AtmosWeave achieves strong global forecasting performance while effectively leveraging global information for high-resolution regional forecasting. Code is included with the submission and will be released publicly upon acceptance.

Primary area:
learning on time series and dynamical systems
Keywords:
weather forecasting, graph neural networks, global-to-regional forecasting, spatiotemporal dynamics

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_JTSmOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes AtmosWeave, a graph-based framework for global and regional weather forecasting. The global model introduces Scale-Decoupled Directional Mesh Propagation (SD-DMP) to improve multiscale and directional message passing, and Geography- and State-Aware Grouped Linear Attention (GS-GLA) to capture long-range interactions efficiently. For regional forecasting, the authors further propose CrossWarp, which adaptively transfers features from a frozen global model to a high-resolution regional model.

Experiments on ERA5 demonstrate competitive global forecasting performance and consistent improvements in regional forecasting. Ablation studies also show the effectiveness of the proposed components.

Critical strengths

1.Well-motivated architectural design. The paper identifies meaningful limitations of graph-based weather forecasting, including multiscale message aggregation and long-range information exchange. The proposed SD-DMP and GS-GLA provide reasonable solutions to these challenges.

2.Effective global-to-regional coupling. CrossWarp introduces an interesting approach to transferring global atmospheric information into regional forecasting. The results in Table 3 show clear improvements over both regional-only forecasting and fixed geographic alignment.

3.Competitive forecasting performance. The experiments cover global and regional forecasting across multiple lead times. AtmosWeave achieves competitive global results and consistently strong regional forecasting performance.

Critical weaknesses
  1. The individual contributions of scale decoupling and directional modeling are not sufficiently validated.

SD-DMP introduces two distinct mechanisms: scale-decoupled message aggregation and directional modeling using graph Laplacian eigenvectors. However, Table 3 only evaluates the removal of the entire module.

Although the results demonstrate the overall effectiveness of SD-DMP, they do not establish whether both mechanisms are necessary. In particular, it remains unclear whether spectral directional features provide advantages over simpler geometric directional encodings.

It could be useful separately ablating scale decoupling and directional modeling, and comparing the proposed spectral representation with conventional geometric direction encodings.

  1. The necessity of geography- and state-aware grouping remains unclear.

GS-GLA uses both geographic coordinates and atmospheric states to construct source groups and target routing weights. However, the current ablation only compares the full model with a variant without grouping.

According to Table 3, removing the grouping mechanism increases RMSE from 0.2654 to 0.2721, while ACC decreases only slightly from 0.8875 to 0.8868. This suggests that grouping provides some improvement, but the individual contributions of geographic and state information remain unclear.

  1. CrossWarp lacks sufficient ablations to justify its multi-stage coupling design.

CrossWarp combines adaptive feature warping at the encoder, orthogonal-component coupling in the processor, and feature correction at the decoder. While Table 3 demonstrates that adaptive coupling outperforms fixed geographic alignment, it does not establish the necessity of each coupling stage.

In particular, it is unclear whether orthogonal projection provides advantages over direct residual feature fusion, or whether independently recomputing feature correspondence at the decoder is necessary.

Minor:

1.How is global context generated during autoregressive regional forecasting? Appendix A.2 describes the pairing of regional inputs with global atmospheric context, while Section 4.4 uses features extracted from a frozen global forecasting model.

However, it is not entirely clear how global context is generated and updated during multi-step regional forecasting. Specifically, does the regional model use recursively predicted global states at each forecast step, or global features derived from ERA5 fields at the corresponding future times?

2.The paper could benefit from further language polishing and structural reorganization to improve its clarity and readability. In particular, the two research objectives could be more clearly distinguished. The current presentation makes it somewhat difficult to follow the overall methodological framework and the relationship between its components.

Questions for authors

Please see weaknesses

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

This paper proposes a graph-based framework for global and high-resolution regional weather forecasting. The proposed framework incorporates three main components: 1) Scale-Decoupled Directional Mesh Propagation (SD-DMP), which captures multiscale local atmospheric dynamics while preserving directional information; 2) Geography- and State-Aware Grouped Linear Attention (GS-GLA), which efficiently models long-range spatial dependencies through grouped attention; and 3) CrossWarp, which adaptively integrates global atmospheric information into high-resolution regional forecasting through learned feature correspondence and selective information injection. Experiments on global ERA5 data and a high-resolution regional dataset over East Asia demonstrate competitive forecasting performance against existing methods, while ablation studies support the effectiveness of the proposed components.

Critical strengths
  1. The paper is generally well written, clearly organized, and easy to follow.
  2. The paper studies an interesting and important problem, particularly how to effectively leverage global atmospheric background information to improve high-resolution regional weather forecasting. This is a practically valuable direction.
  3. The proposed method is well motivated, and its overall design appears reasonable.
Critical weaknesses
  1. A closely related work, S2Transformer [1], is missing from the paper. S2Transformer also addresses the challenge of balancing local and global spatial information in meteorological forecasting. Specifically, it introduces a spatial structured attention mechanism that captures local dependencies through intra-subgraph attention and global dependencies through inter-subgraph attention, together with a multiscale architecture for modeling spatial interactions at different scales. These ideas are closely related to the motivation behind the proposed SD-DMP and GS-GLA modules. Although S²Transformer focuses on station-based forecasting rather than gridded global and regional forecasting, the conceptual similarities warrant a discussion.
  2. The implementation details of the learned low-rank directional fusion operator (FdirF_{\mathrm{dir}}) in Section 4.2 are not sufficiently clear. In particular, the method uses the first (K) nontrivial eigenvectors of the normalized mesh Laplacian as graph coordinates to construct directional features. However, it is unclear how these eigenvectors are computed, whether they are precomputed and reused, and how their computation affects the overall efficiency of the method. Although Section 4.5 provides a complexity estimate for SD-DMP, it does not explicitly account for the cost of computing the Laplacian eigenvectors. Since spectral decomposition can be computationally expensive for large meshes, the authors should clarify the implementation of (F_{\mathrm{dir}}), explain the eigenvector computation procedure, and provide a more complete computational cost analysis, including any preprocessing overhead.
  3. The experimental evaluation mainly focuses on conventional forecasting metrics, such as RMSE and ACC. While these metrics are useful for measuring overall predictive accuracy, they do not necessarily reflect a model's ability to capture fine-grained meteorological structures, which is particularly important for high-resolution regional forecasting. For meteorological applications, one of the key motivations for increasing spatial resolution is to better resolve localized weather phenomena and extreme events. It would therefore be valuable to investigate whether the proposed global-to-regional coupling improves the prediction of such phenomena, rather than merely reducing average forecasting errors.

[1] S2Transformer: Scalable Structured Transformers for Global Station Weather Forecasting, TMLR 2026.

Questions for authors
  1. Please justify the design of low-rank directional fusion operator and the spectral decomposition of mesh graph.
  2. Can the author clarify the relationship between the proposed method and the reference [1]?
  3. Can the authors evaluate the model on extreme weather cases or analyze its ability to preserve fine-scale spatial structures and localized extremes? Such experiments would provide stronger evidence of the practical value of the proposed high-resolution regional forecasting framework.
Reviewer_f5bCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

AtmosWeave is a graph-based framework for global and regional weather forecasting. Its global processor combines scale-decoupled directional mesh propagation with grouped linear attention for long-range interactions. CrossWarp transfers features from a frozen global model to a regional model through adaptive spatial correspondence and gated residual updates. Experiments evaluate global forecasting at 1.5° resolution and regional forecasting over East Asia at 0.25° resolution.

Critical strengths
  1. GS-GLA makes a useful modification to linear attention by separating source assignment from target routing and conditioning both on atmospheric state and geography.

  2. CrossWarp provides a reasonable way to let regional features select relevant global context rather than relying entirely on fixed geographic alignment.

  3. The results are competitive, and the existing ablations support the contributions of the main modules, the grouping mechanism, and adaptive global-to-regional coupling.

Critical weaknesses
  1. The main text leaves important parts of the method difficult to follow. The weights in Eq. (7), the directional signals and fusion in Eq. (8), and the correspondence scores in Eq. (15) require substantial reference to the appendix. Section 4.4 also introduces a separate regional model without clearly describing its backbone or its relationship to the preceding sections.

  2. The relationship to prior methods needs clearer discussion. The spectral smoothing and derivative operators are closely related to Directional Graph Networks (ICML 2021), and STCast (CVPR 2026) already studies adaptive global-to-regional weather coupling. The paper should explain more precisely what its designs add beyond these approaches.

  3. The ablations establish module-level benefits but leave two important mechanisms unresolved. Removing SD-DMP changes both scale aggregation and directional correction, so the contribution of the latter remains unclear. Likewise, the CrossWarp results do not establish whether orthogonal injection improves on directly injecting the routed global features.

  4. The regional multi-step inference procedure is insufficiently specified. The appendix describes pairing regional states with concurrent global context, but does not clearly explain how this context is updated during a ten-day rollout. This matters for interpreting the long-range regional results.

Questions for authors
  1. What architecture does the regional backbone use, and does its processor include SD-DMP and GS-GLA? Please clarify the complete two-model pipeline and move the essential operator definitions into the main text.

  2. What are the specific methodological advances over Directional Graph Networks and STCast?

  3. Do the gains remain when scale-decoupled aggregation is retained without the directional correction, and when CrossWarp directly injects the routed global features without orthogonal projection?

  4. What exactly is retained in “w/o Adaptive CrossWarp”? During multi-step inference, is all subsequent global context generated from the initial state, without reading future ERA5 observations?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.