acceptodds
Under review as a conference paper at ICLR 2027

STBridge: A Spatio-Temporal Neural Bridge for Global Weather Forecasting

Abstract

Medium-range weather forecasting requires preserving atmospheric structure over long lead times while controlling cross-scale error accumulation, maintaining spatial transport, and keeping meteorological variables coherently coupled. Existing models often rely on spherical graphs, multi-scale operators, temporal cascades, broad receptive fields, or variable-aware representations, but they typically model scale transfer, spatial transport, temporal evolution, and cross-variable interactions in separate pathways. To address these limitations, we propose STBridge, a Spatio-Temporal Neural Bridge that organizes global weather forecasting around a shared grid state. The model aggregates the encoded atmospheric history across multiple resolutions and returns the fused multi-grid state to the forecast grid, then learns directional spatial neighborhoods and temporal states for forecast evolution. The shared state supports spatial-semantic message exchange, followed by grid-wise coupling of the six atmospheric variables within each cell; a variableconditioned decoder produces the future fields. Experiments on the ERA5 dataset show that STBridge achieves the best average ACC and the lowest average RMSE among the compared foundation and baseline models, while ablation and visualization experiments confirm the effectiveness and accuracy of its core modules. These results demonstrate that a shared grid state can coordinate multi-scale context, spatio-temporal relations, spatial-semantic routing, and cross-variable coupling for global medium-range forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
Weather forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_tBbvOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

STBridge predicts six atmospheric variables at 6–168-hour leads from two historical ERA5 states. It combines multiscale feature transfer, directional neighborhoods and temporal states, spatial–semantic routing, and within-cell variable coupling on a common forecast grid. Training uses field error and structural regularizers. Evaluation on 256 initializations from 2022 reports aggregate ACC/RMSE improvements over four weather models, supplemented by ensemble forecasts, component ablations, and visual analyses.

Critical strengths
  1. The architecture has a coherent organizing principle. Multiscale, spatial, temporal, and variable-specific information retain a common grid correspondence. Area-weighted restriction and periodic longitude handling are also specified, providing a concrete basis for integrating these pathways.
  2. The main evaluation covers multiple variables and forecast leads. All methods are evaluated on a common verification grid using 256 initializations and 28 forecast leads. Across this broad evaluation, the paper shows an overall forecasting advantage for STBridge, providing empirical support for its ability to predict multiple variables across forecast horizons.
Critical weaknesses
  1. Concerns about the central novelty and methodological motivation. The main contribution is framed as resolving “separate pathways” through a shared grid state, but the paper seems not to make clear the resulting representational or forecasting bottleneck or the distinctive capability provided by Eqs. (3)–(9) beyond conventional fusion. Relevant precedents exist for multigrid modeling and variable aggregation, including SMgNO [1] and ClimaX [2]. Could the authors elaborate on the specific additional mechanism introduced along these directions? Moreover, the coupling loss in Eq. (29) can approach zero when attention concentrates on the target variable itself, as D.2 acknowledges. A low value of this loss therefore does not by itself demonstrate cross-variable information use. I remain concerned that the component-removal experiments do not sufficiently distinguish the proposed coordination mechanism from additional feature transformations, and am not yet convinced by the mechanism interpretation.
  2. There are formatting problems.Adjacent paragraphs in the main text follow almost at ordinary line spacing, without clear paragraph separation. This differs from the half-line spacing specified by the official template and substantially weakens readability and argumentative structure. Meanwhile, both pp. 7 and 9 have visible blank space at the bottom, with approximately five to six lines of text space remaining after the conclusion on p. 9. This leaves the manuscript's presentation and informational completeness below an acceptable standard.

[1] Hu, Y., Yin, F., Zhang, W., Ren, K., Song, J., Deng, K., and Zhang, D. Spherical multigrid neural operator for improving autoregressive global weather forecasting. Scientific Reports, 15, 11522, 2025.doi:10.1038/s41598-025-96208-y.

[2] Nguyen, T., Brandstetter, J., Kapoor, A., Gupta, J. K., and Grover, A. ClimaX: A foundation model for weather and climate. Proceedings of the 40th International Conference on Machine Learning, PMLR 202:25904–25938, 2023.

Questions for authors
  1. I remain uncertain about the specific bottleneck caused by “separate pathways” that existing multigrid or variable-aggregation methods struggle to address. Could the authors explain which operation in STBridge resolves it and what capability it adds beyond standard concatenation, addition, or attention-based fusion?
  2. Given the self-variable attention degeneracy in Eq. (29), I am concerned that reducing this loss may not reflect cross-variable information use. Could the authors explain how this objective encourages such interaction and provide a matched loss-only ablation or off-diagonal evidence to help distinguish cross-variable exchange from self-reconstruction?
  3. I remain uncertain about the attribution of routing and coupling gains: B.1 describes different initialization protocols for reduced variants and the complete system, while B.4/F.1 refer to six specialists. Could the authors show how much improvement over conventional fusion remains when all other settings are held constant, to help assess the role of the proposed coordination mechanism?
Reviewer_98FBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes STBridge, a neural architecture for global medium-range weather forecasting. The model combines four components: Adaptive Multi-Grid Bridging for cross-resolution information transfer, Spatio-Temporal Learnable Neighborhood for spatial and temporal representation learning, Dynamic Spatial Graph Routing for spatial-semantic message exchange, and Grid-Wise Multi-Element Coupling for interactions among meteorological variables. A compound objective additionally encourages spatial structure preservation, cross-scale consistency, and variable coupling.

Critical strengths
  1. Strong practical motivation and promising empirical performance. Global medium-range weather forecasting requires modeling cross-scale spatial structures, temporal evolution, and interactions among meteorological variables. STBridge addresses these challenges in a unified architecture and reports competitive results across six variables and forecast leads up to seven days.
  2. Comprehensive system design and empirical analysis. The paper investigates multiple architectural components, neighborhood configurations, and forecast visualizations. Its modular design provides a plausible framework for integrating multi-scale representations, spatial information exchange, and variable interactions, and the reported results suggest that the overall system is effective.
Critical weaknesses

The paper reads as overly engineering-driven rather than a principled methodological advance. STBridge combines multi-resolution feature fusion, learned neighborhood aggregation, gated spatial-semantic routing, variable-wise attention, and auxiliary structural losses. These components are reasonable, but the paper does not clearly establish which design is fundamentally novel beyond their integration. This point is also reflected by Figure 2, which packs too many elements into a single panel and easily causes readers to lose focus on the main contribution of this paper. If the authors can foreground the core innovation and its motivation, and restructure the presentation so that the contribution no longer reads as an engineering pipeline, I would be willing to raise my score.

Questions for authors
  1. What is the central methodological contribution of STBridge beyond combining multi-scale fusion, graph routing, temporal modeling, and variable-wise attention?
  2. How does STBridge compare with other baselines in parameter count, training cost, inference latency, and memory usage? Do the reported accuracy improvements justify the additional architectural complexity?
  3. How does STBridge perform under extreme weather conditions, and does the shared-grid design provide any measurable advantage in such settings?
Reviewer_Yw7xOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes STBridge, a global weather forecasting model integrating multi-scale feature fusion, spatio-temporal neighborhood learning, graph routing, and cross-variable attention. Experiments on ERA5 report superior performance over Pangu-Weather, FuXi, OneForecast, and Aurora for forecasts up to 168 hours.

Critical strengths
  1. The paper addresses an important problem in medium-range global weather forecasting and considers spatial, temporal, multi-scale, and cross-variable interactions.

  2. The architecture organizes multiple information-processing mechanisms within a shared grid representation.

  3. The evaluation covers multiple meteorological variables and forecast horizons, with additional ablation studies and visualizations.

Critical weaknesses
  1. The four modules primarily combine established operations, including pooling, interpolation, attention, gating, convolution, and GRU. Their specific motivations and necessity are not sufficiently justified. For example, it is unclear why Adaptive Multi-Grid Bridging is preferable to conventional multi-scale CNNs, or why additional graph routing is necessary after multi-scale spatial aggregation. Similarly, the benefits of combining temporal convolution and GRU for long-range atmospheric evolution remain unclear. More fundamentally, sequentially stacking these modules within a shared grid representation does not necessarily constitute a novel unified modeling mechanism. The paper lacks a convincing explanation of why this particular combination and organization are necessary.

  2. STBridge reportedly uses only two years of ERA5 training data (2019–2020) and eight training epochs, yet outperforms several large-scale pretrained weather forecasting models. While not inherently impossible, these results require substantially stronger reproducibility evidence. Important details remain insufficiently specified, including complete training configurations, auxiliary loss hyperparameters, training curves, random seeds, and checkpoint selection procedures. In particular, Appendix B.4 states that the reported results come from specialist checkpoints, while Appendix F.1 refers to six specialists. However, their training procedures, parameter sharing, and checkpoint selection are not clearly explained. This ambiguity makes it difficult to determine the actual model configuration and total training cost, particularly given the claimed shared representation and cross-variable coupling.

  3. The ablation studies in Tables 2 and 3 raise concerns about experimental fairness. Appendix B.1 states that the complete model uses the selected main-results system, while the ablation variants follow independent initialization protocols. This makes it difficult to attribute the reported performance differences solely to the proposed components. The issue is particularly evident in Table 3, where K=3K=3 substantially outperforms K=1,2,4K=1,2,4, despite the latter configurations showing similar performance. More controlled experiments with consistent training and checkpoint selection procedures are needed to support these conclusions. Another important issue is the forecasting strategy. STBridge employs lead-conditioned multi-horizon prediction, unlike the autoregressive or cascaded approaches used by several baselines. The experiments do not isolate the contribution of this strategy from that of the proposed modules. A simpler architecture trained with the same multi-horizon objective would provide a more informative comparison. Finally, the probabilistic evaluation raises concerns about metric interpretation. The signed SME in Eq. (14) is treated as a lower-is-better metric, although lower values do not necessarily indicate better spread-error agreement. Similarly, the RQE in Eq. (15) measures relative absolute deviations of predicted quantiles but does not directly assess quantile calibration.

Questions for authors
  1. Can the authors provide complete from-scratch training and evaluation scripts, configurations, training curves, checkpoints, and logs? Can the reported performance be reproduced across multiple random seeds under the stated two-year training dataset and eight-epoch budget? Can the authors repeat Tables 2 and 3 under consistent training and checkpoint selection protocols and report results across multiple random seeds? Can the contributions of temporal convolution, GRU, and auxiliary losses be evaluated separately?

  2. What exactly are the six specialists mentioned in the manuscript? Are they independently trained models or components of a jointly trained system?

  3. How much of the reported improvement comes from the lead-conditioned multi-horizon forecasting strategy itself? Can the authors compare against a simpler architecture using the same prediction objective and investigate temporal consistency across forecast leads? Please also clarify the probabilistic metric definitions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.