acceptodds
Under review as a conference paper at ICLR 2027

Generative Hierarchical Mesh Network for Global Station Weather Probabilistic Forecasting

Abstract

Global Station Weather Forecasting (GSWF) is pivotal for energy, aviation, and agriculture, yet existing time series methods struggle with its core challenges: capturing hierarchical dependencies and inherent uncertainty of chaotic meteorological systems, while adapting to global weather station data’s uneven spatial distribution and limited size. To address these, we propose a novel Generative Hierarchical Mesh Network (GHMN). Specifically, we construct a hierarchical graph with bidirectional cross-level mapping, encoding bottom-layer local details into top-layer global features, and propagate global context downward for fine-grained predictions. For uncertainty modeling, we design a lightweight flow matching-based generative head that takes deterministic outputs as conditions to generate plausible ensemble predictions without preset priors. By decoupling forecasting into deterministic mean prediction and residual uncertainty estimation, we reduce learning complexity. Experiments show GHMN outperforms spatiotemporal baselines by up to 11%, with the advantage of probabilistic prediction.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatiotemporal Series Forecasting, Global Station Weather Forecasting, Hierarchical Mesh Network, Flow Matching

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_GqBFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GHMN, a framework the combines a hierarchical spherical mesh backbone with a flow-matching generative head for global station weather forecasting. Experiments on the NOAA GSOD dataset across six meteorological variables show consistent improvements over GNN-based, Transformer-based, and diffusion-based baselines in both deterministic and probabilistic settings.

Critical strengths

1.The paper addresses two real pain points in global station weather forecasting: partitioning-based methods suffer from unbalanced splits and boundary information loss, and jointly learning spatiotemporal patterns and uncertainty is data-hungry when stations are sparse and unevenly distributed. 2.The bidirectional propagation with independent parameters per level clearly differs from standard multi-scale GNNs that share one layer across scales, and the single-scale ablations confirm the benefit. 3.GHMN outperforms all baselines across variables and metrics, with particularly clear gains on extreme-event identification and probabilistic forecasting.

Critical weaknesses

1.The flow-matching head is essentially a standard conditional MLP with AdaLN conditioning and an ℓ2\ell_2 velocity loss, closely following existing work. Its novelty lies mainly in where it is applied (on deterministic predictions rather than latent features), not in the method itself. The paper should state more clearly what is genuinely new here versus a straightforward engineering choice. 2.Table 2 only reports CSDI, DiffSTG, and GHMN, and the CRPS numbers for the diffusion baselines are difficult to assess without the deterministic baselines' CRPS for reference. More broadly, standard probabilistic baselines (e.g., quantile regression or a Gaussian head on the same backbone) are missing, so it is unclear how much of the gain comes from flow matching specifically. 3.The paper evaluates only on GSOD (daily), while WEATHER-5K is merely cited as a reference, not used as a benchmark. Since hourly data is more relevant for many operational tasks, adding WEATHER-5K as an extra evaluation setting would better demonstrate the model's generality

Questions for authors
  1. What is genuinely new in the generative formulation? The flow-matching head closely follows existing work (a standard conditional MLP with AdaLN conditioning and an ℓ2\ell_2 velocity loss), with novelty mainly in where it is applied rather than in the method itself. Could the authors clarify what is genuinely new here versus a straightforward engineering choice?

  2. How much of the probabilistic gain comes from flow matching specifically? Table 2 only reports CSDI, DiffSTG, and GHMN, and the CRPS numbers are hard to assess without the deterministic baselines' CRPS for reference. Standard probabilistic baselines (e.g., quantile regression or a Gaussian head on the same backbone) are also missing. Can the authors add such baselines to isolate the contribution of flow matching?

  3. Can the model generalize to an additional dataset? The paper evaluates only on GSOD (daily), with WEATHER-5K merely cited as a reference. Since hourly data is more relevant for many operational tasks, could the authors add WEATHER-5K as an extra evaluation setting to better demonstrate generality?

  4. What is the inference cost of GHMN versus baselines? For the probabilistic setting with 10 samples × 50 steps in particular, reporting latency, FLOPs, and memory would make the efficiency claim more convincing.

2 comments

Q1: What is genuinely new in the generative formulation?

We agree that MLP, AdaLN, and flow matching are established components. Our contribution is a decoupled formulation for sparse, irregular station data: the hierarchical backbone learns multi-scale spatiotemporal dynamics, and the generative head estimates uncertainty conditioned on explicitly supervised deterministic forecasts rather than latent features. Jointly trained, this design reduces the complexity of learning dynamics and uncertainty together under limited data.


Q2: How much of the probabilistic gain comes from flow matching specifically? Standard probabilistic baselines (e.g., quantile regression or a Gaussian head on the same backbone) are also missing.

Deterministic baselines provide point forecasts rather than predictive distributions. If treated as point-mass distributions, their CRPS reduces to absolute error; thus, Table 1's MAE provides the corresponding deterministic CRPS reference. To isolate the probabilistic contribution of flow matching, we conduct controlled head ablations using the same GHMN backbone, identical head architectures, and a matched inference budget of 50 steps.

HeadMAEMSECRPS
Diffusion2.3511.771.12
Flow Matching2.2710.960.95

FM improves both point metrics and CRPS under the matched budget. We also evaluate a Gaussian residual baseline with a fixed isotropic variance estimated from the same backbone's training residuals, reporting interval coverage and Sample Spread Ratio (SSR):

Uncertainty modelCov@80%Cov@95%SSR
Gaussian residual baseline0.640.710.83
Diffusion head0.700.821.14
Flow Matching head0.680.801.09

Experimental Analysis:

  1. The Gaussian baseline suffers from under-dispersion and low coverage. Its homoscedastic assumption fails to capture spatiotemporal heteroscedasticity, meaning it cannot adaptively widen predictive intervals for high-risk locations or extreme events.
  2. The Diffusion Head is over-dispersed; while it achieves higher coverage, it sacrifices sharpness. Our Flow Matching head yields the best SSR closest to 1.0. This confirms that FM learns a more accurate, heavy-tailed probability density through site-specific modeling.

Q3: Can the model generalize to an additional dataset?

We have additionally evaluated GHMN on MADIS 2020, an independent dataset with hourly observations and a substantially denser station network. Following the dataset's quality-control protocol, we use 25,328 stations in a 2-hour-to-2-hour forecasting task. As shown in the table, GHMN consistently outperforms MIGN across all variables.

ModelDewpoint MSE / MAETemperature MSE / MAEWind Speed MSE / MAE
MIGN2.21 / 0.974.49 / 1.500.78 / 0.56
GHMN2.09 / 0.934.21 / 1.430.76 / 0.57

Q4: Reporting latency, FLOPs, and memory would make the efficiency claim more convincing.

Our hierarchical mesh has a fixed number of levels and nodes, so increasing the station count primarily scales the GNN encoder/decoder. With 10 samples and 50 integration steps, the backbone is evaluated once to obtain a deterministic condition reused throughout ensemble generation. The 500 subsequent evaluations involve only the MLP head, avoiding repeated hierarchical mesh computation. The profiling results below use a training batch size of 16 and an inference batch size of 1 with 10 prediction samples. Nearly tripling the station count yields sub-linear increases in time and memory, demonstrating practical scalability.

DatasetStationsTraining TimeInference TimeReported GPU Memory
GSODapproximately 9,0002.9 h3.9 s30 GB
MADIS25,3285.3 h6.3 s39 GB

Considering the novelty of the work, I maintain my original score.

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

This paper addresses Global Station Weather Probabilistic Forecasting (GSWF) by proposing the Generative Hierarchical Mesh Network (GHMN). To handle irregularly distributed stations and multi-scale atmospheric interactions, the authors map station observations onto a hierarchical spherical HEALPix mesh (L=3L=3 levels) via a learnable GNN encoder, process multi-scale features through a U-Net-like backbone equipped with Multivariate SpatioTemporal (MST) blocks, and project representations back to station coordinates to produce deterministic predictions XconX_{con} (supervised by an MAE loss L1L_1). To enable probabilistic forecasting without end-to-end diffusion overhead, GHMN attaches a lightweight MLP-based Conditional Flow Matching (CFM) head conditioned on XconX_{con} via Adaptive Layer Normalization (AdaLN) (supervised by a velocity matching loss L2L_2). All experiments are conducted on a single daily dataset (NOAA GSOD across 6 variables), evaluating MSE, MAE, CRPS, and 90th/10th percentile SEDI against deterministic and diffusion baselines.

Critical strengths

Effective Decoupling of Deterministic Spatiotemporal Modeling and Uncertainty Estimation on Irregular Stations:

End-to-end spatiotemporal diffusion models (e.g., CSDI, DiffSTG) struggle to jointly learn complex spatial correlations and stochastic noise when trained on irregular, sample-limited station observations (as evidenced by their severe degradation in Table 2). By delegating multi-scale spatial representation learning to a hierarchical HEALPix mesh backbone (L1L_1) and restricting the generative flow-matching head (L2L_2) to model post-prediction distribution, GHMN achieves stable convergence and strong empirical gains across point (MSE/MAE), probabilistic (CRPS), and extreme event (SEDI) metrics on the GSOD benchmark (Tables 1–2, Figure 3)

Critical weaknesses

1. Questionable Experimental Protocol and Identical MAE Across Deterministic (Table 1) and Probabilistic (Table 2) Evaluations:

A major concern regarding experimental soundness and fairness lies in the fact that GHMN reports the exact same MAE values across all six meteorological variables in both Table 1 (comparison with deterministic models: 2.05, 1.91, 1.86, 3.11, 1.92, 3.12) and Table 2 (comparison with probabilistic models: 2.05, 1.91, 1.86, 3.11, 1.92, 3.12).

2. Limited Methodological Novelty ("GNN + Flow Matching" Integration):

The technical contributions are largely incremental combinations of established paradigms. First, following GraphCast, the "Station-to-Mesh Encoder →\to Multi-scale Mesh GNN →\to Mesh-to-Station Decoder" pipeline is already standard in weather forecasting. Second, migrating generative frameworks to probabilistic weather forecasting (since GenCast) and specifically applying Conditional Flow Matching (CFM) to meteorological forecasting have already been explored (e.g., FlowCast at ICLR 2026: https://proceedings.iclr.cc/paper_files/paper/2026/hash/fe112a9042ed19ec515ed700d394a8bb-Abstract-Conference.html).

3. Unsubstantiated Claims of "Lightweight Design" and Missing Ablation / Benchmark Studies:

Lack of Efficiency Metrics:

The paper repeatedly emphasizes that its flow-matching head is "lightweight" and effective in "cutting inference costs" (Abstract, Sec. 1, Sec. 4.3), yet provides zero empirical evidence on parameter counts, FLOPs, GPU memory, training time, or inference latency (especially given the 500 MLP evaluations per forecast) compared to deterministic and diffusion baselines.

Single Dataset:

Although the authors explicitly compare the hourly WEATHER-5K benchmark with the daily GSOD dataset in Section 5.1 and Table 4, all experiments are restricted to the coarse daily GSOD dataset.

Questions for authors

See Weakness

2 comments

Q1: 1. Questionable Experimental Protocol and Identical MAE Across Deterministic (Table 1) and Probabilistic (Table 2) Evaluations:

The identical MAE values arise because Tables 1 and 2 evaluate the same deterministic output from GHMN's backbone. Our decoupled architecture directly provides a point forecast, while the FM head generates samples for probabilistic evaluation using CRPS. For CSDI and DiffSTG, which lack a separate deterministic output, we average generated samples to compute point metrics. We will clarify these prediction sources and distinguish point-forecast metrics from ensemble-based CRPS in the table descriptions.


Q2: 2. Limited Methodological Novelty ("GNN + Flow Matching" Integration):

Our primary contribution is a parsimonious framework for probabilistic spatiotemporal forecasting on graph data, developed here for the challenges of Global Station Weather Forecasting.

  1. Innovation in spatial modeling for irregular stations. Existing spherical weather models such as GraphCast primarily operate on dense, fixed-resolution reanalysis fields. GHMN instead learns directly from sparse, unevenly distributed station observations. HEALPix recursive subdivision serves as a physical bridge that automatically organizes scattered observations into a balanced, equal-area hierarchy, avoiding graph-partitioning choices such as predefined subgraph counts. This provides structured spatial support for learning global patterns and local details. The same construction principle could extend to urban traffic or logistics, using grids such as H3 or S2 as the spatial bridge.

  2. Innovation in decoupled probabilistic modeling. We recognize that generative weather forecasting and conditional flow matching have been explored. FlowCast jointly models spatiotemporal dynamics and predictive uncertainty, a demanding learning problem when station observations are sparse and training data are limited. GHMN uses the explicit deterministic predictions of all station nodes as the generative condition, decoupling the task into “deterministic mean prediction” and “residual uncertainty estimation.” The hierarchical backbone captures shared spatiotemporal structure, while a lightweight, jointly trained FM head estimates uncertainty around its predictions. The novelty lies in a unified architecture that simplifies hierarchical spatial modeling and uncertainty estimation for non-uniform graph data, rather than in the use of flow matching itself.


Q3: Lack of Efficiency Metrics and Single Dataset:

Similar concerns were raised by Reviewer GqBF. Please refer to our responses to Q3 for additional experiments on the hourly MADIS dataset and Q4 for runtime, GPU memory, and analysis of the sampling cost.

Thank you for the careful and detailed response. The explanation have addressed my main concerns in computation efficiency and single dataset. However, I still do not think that presenting deterministic and probabilistic prediction results in the same table is an appropriate choice. Regarding novelty, there are some atmospheric papers, adopting flow matching. The current response is still insufficient to convince me that the methodological contribution of this paper significantly exceeds a combination and improvement of existing techniques.

Taking all of the above into account, although the authors’ response has addressed some issues, my core concerns remain, and I decide to maintain my original score.

Reviewer_9RLcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets the problem of global station weather forecasting using machine learning. It proposes an architecture that uses a hierarchical graph neural network combined with a mesh encoder-decoder. To get probabilistic predictions the authors add a post-hoc flow matching head that provides samples conditioned on the deterministic prediction. Compared to both deterministic and probabilistic baselines, the method performs well.

Critical strengths

The hierarchical mesh is well-motivated, having achieved previous success for MLWP. The paper compares to a lot of baselines that (to the best of my knowledge) are relevant. The decoupled post-hoc uncertainty quantification is a cheap fix that saves compute during inference.

Critical weaknesses

The graph construction is very similar to Graph-EFM/Graph-FM [1] which it cites but doesn't position itself against. Although applied to full-state weather forecasting and not station forecasting, the graph is built to handle arbitrary inputs and outputs which it uses to condition on the boundary for limited area modeling. With this in mind, the paper reads like an application of Graph-FM to station weather forecasting with a flow matching head added for UQ. The idea of forecasting from observations is also very similar to the DOP-works [2] which is not cited. The experiments are also lacking in terms of details and probabilistic evaluations.

[1] Oskarsson, Joel, et al. "Probabilistic weather forecasting with hierarchical graph neural networks." Advances in Neural Information Processing Systems 37 (2024): 41577-41648.

[2] Pinnington, Ewan, et al. "AIFS-DOP: End-to-End Medium-Range Weather Prediction from Observations Alone with Machine Learning." arXiv preprint arXiv:2606.19093 (2026).

Questions for authors
  • How does the encoder/decoder handle missing station input?
  • Can it be used to forecast at stations not in input?
  • Does the MLP-based flow head act on each station separately or all of them jointly?
    • If it acts jointly, how does it handle missing station input?
    • If it acts separately, how does it encode location-specific information and ensure that the results are spatially coherent?
  • Can you position yourselves against Graph-FM and provide a comparison with the same graph for deterministic prediction?
  • Can you provide more details on the rollouts?
    • What is fed back into the model for auturegressive prediction?
  • The probabilistic results are marginally better than the deterministic ones, can you provide some more evidence to why the flow head helps?
2 comments

Q1: What distinguishes GHMN from Graph-FM/Graph-EFM and observation-driven approaches such as AIFS-DOP? Can you clarify the methodological novelty and provide a deterministic comparison using the same graph?

We acknowledge the connections to hierarchical graph weather models and observation-driven forecasting. Our contribution is a unified framework that simplifies hierarchical spatial modeling and uncertainty estimation for sparse, irregular station data. The main distinctions and supporting evidence are as follows.

  1. Innovation in spatial modeling for irregular stations. Existing spherical weather models such as Graph-FM primarily operate on dense, fixed-resolution reanalysis fields. GHMN instead learns directly from sparse, unevenly distributed station observations. HEALPix recursive subdivision serves as a physical bridge that automatically organizes scattered observations into a balanced, equal-area hierarchy, avoiding graph-partitioning choices such as predefined subgraph counts. This provides structured spatial support for learning global patterns and local details. The same construction principle could extend to urban traffic or logistics, using grids such as H3 or S2 as the spatial bridge.

  2. Decoupled probabilistic modeling. GHMN conditions a lightweight FM head on explicitly supervised deterministic station forecasts, separating shared spatiotemporal structure learning from local residual uncertainty estimation. Both components are trained jointly. This differs from Graph-EFM's hierarchical latent-variable generation: GHMN concentrates stochastic modeling in a station-wise head, reducing the complexity of learning dynamics and uncertainty together under sparse observations.

  3. Controlled comparisons. Direct comparisons with existing mesh forecasting models are challenging because their published configurations primarily target dense gridded fields, whereas our task uses sparse, irregular station observations. To compare the spatial modeling strategies under comparable settings, we implemented a GraphForecast-like multi-resolution variant within our framework, replacing explicit hierarchical pooling with multiple edge spans on a single high-resolution mesh while keeping the scales and hidden dimensions identical to GHMN. GHMN outperforms this variant on both MAE and MSE, supporting the benefits of an explicit hierarchy: independent parameters enable scale-specific updates, while bidirectional propagation balances global patterns and local details.

ModelMAEMSE
GraphForecast-like multi-resolution variant2.4012.23
GHMN2.2710.96

Q2: How does the encoder/decoder handle missing station input? Can it be used to forecast at stations not in input?

Our station-to-mesh-to-station architecture separates input and output locations. Variable-specific encoders map available observations onto the shared HEALPix mesh for joint spatiotemporal modeling, without padding missing stations as observed nodes. The decoder then maps mesh representations to target stations through location-based connections, with loss computed only for valid targets. This design can, in principle, forecast at stations absent from the input by connecting their locations to the mesh.


Q3: Does the MLP-based flow head act on each station separately or all of them jointly?

The FM head operates independently at each station with shared MLP weights, conditioned on that station's deterministic forecast. Location-specific information enters through the station's mesh connections and the backbone's hierarchical spatial message passing. The head models local residual uncertainty around this spatially informed prediction, but does not explicitly couple residuals across stations. Modeling cross-station residual dependence remains future work.


Q4: Can you provide more details on the rollouts and clarify what is fed back into the model for autoregressive prediction?

For the autoregressive point-forecast evaluation in Appendix B.1, we feed the backbone's deterministic station predictions back as inputs. Starting from observed history, the model predicts the next weather state; we then append this prediction to the input window, discard the oldest time step, and repeat with the same model. Future ground-truth observations are not used during the rollout. Figure 4 reports MSE at three successive daily forecast steps on GSOD, comparing GHMN with Corrformer, MIGN, and STELLA.


Q5: Can you provide some more evidence to why the flow head helps?

Please refer to our response to Reviewer GqBF, Q2, which provides controlled FM-versus-diffusion head ablations on the same backbone and additional calibration diagnostics against a Gaussian residual baseline. These results support the contribution of the FM head to probabilistic forecast quality.

I thank the authors for their answers.

The per station FM head is an adhoc fix which has no guarantee to yield spatially coherent ensemble members. The new numbers also show a marginal improvement, and the gaussian residual already does fairly well. Claims like "...reducing the complexity of learning dynamics and uncertainty together under sparse observations." and "This design can, in principle, forecast at stations absent from the input by connecting their locations to the mesh." are not properly demonstrated. I also agree with reviewer wfMY on limited methodological novelty.

With this in mind, I'm maintaining my score.

Comments on the paper

1 comment

We sincerely thank the reviewers for their careful evaluation and constructive feedback. We have revised the manuscript accordingly. The main changes are as follows:

  1. We clarified variable-specific station handling and the FM head’s station-wise conditioning, training, and sampling procedures (Sections 4.2–4.3).
  2. We revised the Introduction and Conclusions to make the novelty and contributions clearer.
  3. We added MST Block details and a theoretical explanation of mean-residual decomposition to support the method (Appendices A.1–A.2).
  4. We expanded Appendix C with Persistence and interpolated Pangu comparisons; coverage and SSR diagnostics against Gaussian and diffusion baselines; multi-resolution, VAE, and diffusion ablations; regional evaluations; hourly MADIS results; and runtime and memory measurements.

We have made every effort to address the comments and strengthen the manuscript. We appreciate the reviewers’ time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.