acceptodds
Under review as a conference paper at ICLR 2027

UniGIO: Unified Generative Global In-situ Weather Modeling from Spatiotemporal Incomplete Observations

Abstract

Global In-situ Observation (GIO) provides fine-scale, pointwise weather measurements beyond gridded satellite products, supporting numerical weather prediction and disaster prevention. However, severe spatiotemporal incompleteness in GIO leads existing methods to rely on AI-ready preprocessing or massive global multi-source data, limiting accurate, timely, and efficient in-situ weather services. To address this gap, we propose UniGIO, a unified generative framework that exploits station-centered observational complementarity and decouples local extremes from steady evolution, unifying forecasting, imputation, and generation directly from incomplete GIO. UniGIO progressively builds complementarity from discrete station interactions to continuous local fields. It captures missing-aware sequential evolution in a shared state space, while separate expert representations support extreme-pattern adaptation and global pattern recognition. These mechanisms form a lightweight CVAE suitable for distributed stations, explicitly modeling missing-state uncertainty for extreme-event assessment and generation. Experiments on Weather-5K, the largest up-to-date GIO dataset, demonstrate SOTA performance, with average gains of 11%, 12%, and 5% in accuracy, fidelity, and extreme-event capture, advancing unified weather modeling from incomplete station observations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
In-situ weather observation, Data-driven weather system modeling, Time series forecasting, imputation, generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1aoCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose UniGIO, a small conditional VAE for global weather-station data interpolation and forecasting. Forecasting, imputation and station/variable generation are framed as a mask-conditioned completion problem. Its main components are an Observation Mixer (variable attention, station graph attention and a latent field grid) and an Adaptive Temporal Mixer (reliability-weighted bidirectional Mamba + a geography-aware mixture-of-experts). The authors test their approach on Weather-5K and reports strong gains in accuracy. They compare to generative, station-forecasting and foundation-model baselines and include a comparison to raw HRES, an operational ECMRWF product.

Critical strengths

I want to highlight the following strengths of the submission

  • Overall, the paper is pretty well written and easy to follow; the methodology is outlined comprehensively (though sometimes lacking a bit of detail).
  • The problem this submission tackles is obviously highly relevant. AI for climate and weather modeling is becoming increasingly important with accelerating climate change.
  • The idea of disentangling (global) regular time series forecasting from (local) extreme event modeling is very interesting and intuitively a valuable thing to try out. It's both well motivated practically and an interesting approach methodologically.
Critical weaknesses

The submission has the following shortcomings, ordered roughly from, in my opinions, most to least severe:

  • There's an overall lack of contextualization with existing work in sparse / masked spatio-temporal modeling. Application wise, there's so much related work on sparse and continuous spatio-temporal observations from which the goal is to obtain a (dense) map and to forecast; from remote sensing to ecology. The paper could benefit from a related work section to place its contributions within existing work beyond just AI for weather and climate. Methods wise, I feel like several important baselines are missing in the experiments. From simple interpolation baselines to discrete models (temporal MAEs etc.) and continuous models (Gaussian Processes, location encoders, neural processes,...so much work focused on exactly this problem!). Furthermore, there's a whole area of literature focused on spatio-temporal sensor networks [1,2,3]. Why is there no comparison to any of these methods arguably in the same area?
  • The claim that UniGIO "outperforms operational HRES" seems overstated, given that is not the case for wind direction and sea-level pressure.
  • There are no error bars / standard errors given thoughout the work.
  • If I understand correctly, extreme events are simulated by sampling (and masking) random stations; this seems a bit simplistic when actually known locations / areas that exhibit regular extreme events could be chosen for masking?
  • The figures in the paper are illegible (way too small); you have to zoom in a lot in order to be able to see details. If you're printing out the paper to read it you could simply not understand the figures. They are also way too overcrowded, so that they don't actually act as a valuable visual helper to get key messages across. Figures should be heavily simplified and reworked with a focus on legibility.
  • Am I missing something or is space modeled twice? The field-former + the graph network. If so, why is it modelled twice? And what is the effect of each of the components?

References:

[1] Cini, Andrea, Ivan Marisca, and Cesare Alippi. "Filling the g_ap_s: Multivariate time series imputation by graph neural networks." ICLR 2022

[2] Wu, Yuankai, et al. "Inductive graph neural networks for spatiotemporal kriging." AAAI 2021

[3] Liu, Mingzhe, et al. "Pristi: A conditional diffusion framework for spatiotemporal imputation." ICDE 2023

Questions for authors

Please see comments above for questions; specifically, please share your thoughts on my perceived lack of baseline comparisons + the questions on extreme event simulation and double space modeling.

8 comments
AuthorsOpenReview ↗

Response to Reviewer 1aoC (Weakness 1/6): Related Work and Additional Spatiotemporal Baselines

Response: Related Work and Additional Spatiotemporal Baselines

Revision note: Major revisions are highlighted in blue in the revised manuscript. Newly added baseline methods are highlighted in blue in the revised manuscript, including Tables 1–4 and Appendix Tables F.1–F.3. The corresponding experimental analysis is provided in the main text (Lines 366–372). The expanded Related Work is provided in Appendix A.2 and A.4.

Thank you for this valuable suggestion. We have substantially expanded the Related Work and strengthened the experimental comparisons to systematically address the methodological categories highlighted by the reviewer. These additions provide a broader context for positioning UniGIO and further clarify its distinctive advantages in global in-situ observation (GIO) modeling.

1. Expanded Related Work. We have extended our discussion to cover sparse and masked spatiotemporal learning, including classical interpolation and Gaussian Processes, discrete station-based modeling and masked autoencoders, continuous spatial modeling and neural processes, graph-based spatiotemporal sensor networks, and generative imputation. We further discuss representative methods such as Kriging, location encoders, GRIN, SPIN, and IGNNK. These additions establish clearer connections to the broader literature and distinguish the general objectives of sparse spatiotemporal reconstruction from the specific requirements of GIO modeling.

2. Additional Baseline Comparisons. Following the reviewer's suggestion, we have incorporated recent representative advances in the methodological directions highlighted by the reviewer, alongside classical approaches:

  • Classical interpolation and Gaussian Processes: ST-Interp (Spatiotemporal Interpolation) and ST-GP (Spatiotemporal Gaussian Process), representing interpolation and probabilistic spatiotemporal inference.
  • Discrete spatiotemporal modeling: KITS (AAAI 2025), representing inductive modeling of unobserved stations.
  • Continuous spatial modeling: SConvCNP (NeurIPS 2025), representing neural-process-based continuous spatial prediction.
  • Spatiotemporal sensor networks: GSLI (AAAI 2025), representing graph-based learning of heterogeneous spatiotemporal dependencies.
  • Generative spatiotemporal modeling: PriSTI (ICDE 2023), representing conditional diffusion-based imputation.
  • Masked spatiotemporal modeling: Our original Deterministic baseline already adopts the Masked Autoencoder (MAE) paradigm.

These comparisons cover the methodological families identified by the reviewer while incorporating recent representative advances. The updated results show that UniGIO achieves the best MAE/MSE across all 20 variable–metric–mask combinations at 120h and the best F1-SEDI in 15 of 16 extreme-event comparisons, demonstrating consistent advantages across these diverse approaches.

More importantly, the additional baselines reveal a common weakness in extreme-event reconstruction. Under curated masks at the 99.5th percentile, their F1-SEDI scores range from only 0.05–0.14 for temperature, 0.01–0.06 for wind speed, and 0.01–0.12 for sea-level pressure, compared with 0.55, 0.22, and 0.79 for UniGIO, respectively. These results suggest that generic spatiotemporal modeling does not necessarily capture the localized abrupt extremes characteristic of weather observations, even when it effectively models broader spatiotemporal dependencies.

3. Distinction from Existing Sparse Spatiotemporal Modeling. Unlike approaches primarily designed for individual tasks such as interpolation, reconstruction, or imputation, UniGIO unifies forecasting, imputation, and generation from severely incomplete station observations. Beyond task unification, GIO modeling must address the mismatch between discrete station observations and continuous weather processes, as well as the coexistence of smooth temporal evolution and localized abrupt extremes.

Accordingly, UniGIO combines discrete station interactions with continuous local-field representations to integrate fragmented evidence across space and time, while evolution–extreme decoupling preserves stable temporal dynamics and enables adaptation to localized extremes. These complementary mechanisms distinguish UniGIO from existing task-specific spatiotemporal modeling approaches.

We believe the expanded literature coverage, additional baselines, and corresponding experimental analysis substantially address the reviewer's concerns and more clearly establish UniGIO's contributions to sparse spatiotemporal modeling.

AuthorsOpenReview ↗

Response to Reviewer 1aoC (Weaknesses 2/6): Clarification of the HRES Comparison

Response: Clarification of the HRES Comparison

Thank you for pointing this out. We have revised the original statement to explicitly specify the variables for which UniGIO achieves lower forecasting errors than operational HRES.

Revision note: Major revisions are highlighted in blue in the revised manuscript. We have revised the corresponding statement in the main text (Lines 417–419) to clarify the scope of this comparison.

Revised text:

It also outperforms the observation and time-series foundation models ESFM and Chronos2 on most evaluated metrics, while achieving lower forecasting errors than operational HRES for temperature, dewpoint, and wind rate.

AuthorsOpenReview ↗

Response to Reviewer 1aoC (Weaknesses 3/6): Statistical Variability and Error Bars

Response: Statistical Variability and Error Bars

Thank you for pointing this out. We agree that reporting statistical variability is important for assessing the reliability of our results.

Revision note: Major revisions are highlighted in blue in the revised manuscript. We have added a new subsection, Inference Stability, in Appendix H.3 of the revised manuscript, including error-bar visualizations in Fig. H.1 and detailed numerical results in Table H.3.

Specifically, we repeat the evaluation with five different inference seeds for each method, while keeping the trained checkpoints, test split, mask settings, and other evaluation parameters fixed. We report the mean and sample standard deviation (SD) for MAE, MSE, and SEDI.

The results show that UniGIO exhibits consistently low inference variability, with small variations in MAE and SEDI across the five seeds. The model rankings also remain consistent with the reported comparisons. These findings support the stability of our conclusions with respect to inference randomness.

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

UniGIO unifies forecasting, imputation, and conditional generation from incomplete weather-station observations. Its CVAE combines information across stations and local fields with temporal modeling and geographically conditioned experts. Experiments on Weather-5K examine several tasks, missingness patterns, time windows, and regional transfer.

Critical strengths
  1. The paper addresses a useful and difficult problem. Treating several tasks within one framework is appealing, and the use of complementary station observations is well motivated. The design is closely tied to the structure of the data.
  2. The experiments are a particular strength. The reported improvements are broad, with curated-mask temperature MAE improving from 2.05 for SportsTraj to 1.45 (Table 1). Task comparisons, regional transfer, and ablations add useful evidence. The compact 0.9M-parameter variant also makes the method attractive (Appendix H).
Critical weaknesses
  1. The native-mask evaluation would benefit from a clearer description of which entries have reference values and how they enter the scores. The construction of complete sequences for posterior training also needs clarification (Appendices D.4 and E.2-E.4).
  2. The extreme-event results are encouraging, but the risk interpretation goes beyond the sampling-sensitivity evidence in Figure 5(b). Calibration results or more modest wording would help. Appendix D.7 already acknowledges this distinction.
Questions for authors
  1. How are native missingness, reference availability, and complete posterior-training sequences handled in practice?
  2. Can the authors provide probabilistic validation for extreme-event risk, or limit this interpretation to the observed sampling sensitivity?
5 comments
AuthorsOpenReview ↗

Response to Reviewer ajbZ (Weaknesses 1/2): Native-mask References and Posterior Training

Response: Native-mask References and Posterior Training

Thank you for this question. All native-mask positions in Weather-5K have corresponding ground-truth references in the quality-controlled dataset.

Weather-5K provides complete in-situ observation sequences constructed through manual quality control of raw station records, with reference to ERA5 and other observational sources. It preserves the original missing or invalid positions, which we use as native masks during evaluation. Our model reconstructs these originally unavailable observations and is evaluated against the corresponding quality-controlled references. Specifically, native-mask positions with available references contribute to the evaluation metrics, while observed positions are excluded. This benchmarks automated reconstruction against manually curated references.

Accordingly, the posterior branch directly takes the unmasked, quality-controlled Weather-5K sequences as complete observations during training, allowing the Event Aligner (EA) to construct event-aligned posterior representations.

Furthermore, our evaluations with curated masks provide complementary evidence of the model's ability to generate missing values from originally valid observations under controlled missingness patterns.

Further details are available in the Weather-5K paper:

Han, T., Wen, Z., Chen, Z., Du, D., Guo, S., and Bai, L. (2026). Benchmarking Physics-Informed Time-Series Models for Operational Global Station Weather Forecasting. In Proceedings of the 43rd International Conference on Machine Learning (ICML), PMLR 306:40098--40132.

AuthorsOpenReview ↗

Response to Reviewer ajbZ (Weaknesses 2/2): Extreme-event Risk Interpretation

Response: Extreme-event Risk Interpretation

Revision note: We have refined the discussion of latent variance in the main text (Lines 479–485) and added a statistical analysis in revised Fig. 6 to further examine its relationship with potential extreme-event risk.

As explicitly stated in Appendix D.7, latent variance is not a calibrated extreme-event probability. This is a deliberate distinction between qualitative risk assessment and quantitative probability estimation. Our method captures spatially and temporally varying sensitivity to extreme-event dynamics, providing interpretable indications of potential extreme-event risk and enabling controllable scenario generation. However, we do not claim to estimate calibrated event probabilities.

To further examine this relationship statistically, we divided samples, including extreme-event cases, into five quantile groups (Q1–Q5) according to the magnitude of their historical variations, which serves as a proxy for potential extreme-event risk. For each group, we measured the median absolute prediction change as the CVAE sampling coefficient increased from 0 to 1.5, with error bars illustrating within-group variability.

The results in revised Fig. 6 show an overall increasing trend across all five meteorological variables. In particular, wind direction exhibits a clear and nearly monotonic increase from approximately 10° in Q1 to 31° in Q5, while wind speed increases from approximately 0.10 to 0.26 m/s. Temperature and dewpoint also show generally increasing responses, whereas sea-level pressure exhibits a weaker upward trend with minor fluctuations. This positive association provides further empirical evidence that latent sampling sensitivity can reflect potential extreme-event risk, without implying calibrated event probabilities.

We have refined the corresponding descriptions to ensure this distinction is consistently maintained throughout the paper.

AuthorsOpenReview ↗

Response to Reviewer ajbZ (Questions 1/2): Native Missingness, Reference Availability, and Posterior Training

Response: Native Missingness, Reference Availability, and Posterior Training

Thank you for this important question. We distinguish between native missingness in the original station records and reference availability in the quality-controlled Weather-5K dataset. All native-mask positions have corresponding reference values, allowing the posterior branch to use complete sequences during training.

Native missingness. Weather-5K preserves the original missing or invalid positions in raw station observations as native masks. These masks represent realistic observational incompleteness and are used to construct incomplete model inputs and evaluate reconstruction performance.

Reference availability. Weather-5K provides quality-controlled reference values at these positions, obtained through manual quality control with reference to ERA5 and other observational sources. Thus, native missingness does not imply the absence of evaluation targets in the benchmark. Our model reconstructs these originally unavailable observations and is evaluated against the corresponding curated references.

Complete posterior-training sequences. During training, the Event Aligner directly uses the complete, unmasked Weather-5K reference sequences to construct event-aligned posterior representations. These reference values are available only to the training-time posterior branch, not to the conditional prior used for inference. During inference, the posterior branch is removed, and UniGIO generates missing observations solely from incomplete inputs and their observation masks.

Additionally, curated-mask experiments evaluate reconstruction on originally valid observations that are artificially withheld, providing complementary evidence beyond native missingness.

Further details on dataset construction are provided in Han et al. (2026), Benchmarking Physics-Informed Time-Series Models for Operational Global Station Weather Forecasting, ICML 2026.

Reviewer_BtzoOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes UniGIO, a unified generative framework with station-centered observation mixing, reliability-weighted temporal modeling, and geographically conditioned experts, which enables forecasting, imputation, and generation from incomplete weather-station observations. Experiments demonstrate the effectiveness of UniGIO.

Critical strengths
  1. The integration of reliability-weighted temporal modeling and geographically conditioned expert routing within a unified generative framework is a interesting design.
Critical weaknesses
  1. The methodological novelty is limited, and the work appears incremental. UniGIO mainly integrates established techniques, including CVAE, attention, and MoE routing. The authors should clarify the novel contributions beyond this integration.

  2. Some reported MAE/MSE pairs appear mathematically inconsistent. For example, Table 1 reports MAE/MSE of 10.3/76.5 for TimeGAN, violating MSE ≥ MAE² under identical evaluation settings. The authors should clarify the aggregation procedure or correct the results.

Questions for authors
  1. How does the training-time Event Aligner handle genuinely missing target values when constructing posterior representations?
6 comments
AuthorsOpenReview ↗

Response to Reviewer Btzo ( Weaknesses 1/2): Contribution and Novelty

Response: Contribution and Novelty

Revision note: Major revisions are highlighted in blue in the revised manuscript, primarily in the Introduction (Lines 041–087).

Our contributions lie in two main aspects: (1) establishing a station-centered generative paradigm for unified global in-situ weather modeling, addressing the high data and computational demands of global multi-source modeling and the loss of fine-scale dynamics caused by subsequent local interpolation in in-situ service; and **(2) developing task-driven methodological innovations to address severe spatiotemporal missingness under limited resources while efficiently modeling the coexistence of smooth temporal evolution and localized abrupt extremes in local chaotic weather systems.

1. Station-centered Generative Paradigm

Our first contribution is establishing a station-centered generative paradigm as an alternative to global multi-source observation foundation models for GIO modeling. While general spatiotemporal methods can be adapted to individual subaspects of this task, they are typically not designed to jointly meet its full requirements, such as extreme-event awareness.

Compared with Observation FM, UniGIO directly generates complete observations from severely incomplete station measurements, without relying on massive global multi-source data or explicit spatial interpolation. By exploiting local observational complementarity, it preserves fine-scale weather dynamics while reducing data and computational demands. This enables unified forecasting, imputation, and generation through lightweight, distributed station-level modeling, particularly benefiting observation-sparse, timely, and resource-constrained in-situ services, such as disaster management.

2. Task-driven Methodological Innovations

Our second contribution lies in methodological innovations addressing the unique challenges of incomplete GIO. We have clarified these challenges in the Introduction and connected them to the motivations of our designs.

Station-centered complementarity. We progressively integrate evidence across variables and neighboring stations through a station-to-field-to-station mechanism, which aggregates fragmented observations into continuous local-field representations and propagates regional context back to missing station states. This captures spatial continuity without dense global grids or explicit interpolation.

Evolution–extreme decoupling. We separate stable temporal evolution from localized abrupt extremes. Reliability-weighted state-space modeling preserves sequential dependencies while suppressing uncertainty propagation from missing observations, whereas pattern-decoupled experts enable global geography-conditioned pattern sharing and local extreme adaptation. This reduces interference between coexisting fundamentally different weather dynamics.

Lightweight probabilistic generation. The CVAE framework explicitly models mean and variance at missing states, preserving observation constraints while capturing uncertainty for extreme-event assessment and pointwise controllable generation. Its lightweight architecture also facilitates distributed deployment across resource-constrained stations.

3. Clarification of the Novelty Criterion

We respectfully disagree with evaluating methodological novelty primarily based on the use of established architectural paradigms. CVAE, Attention, and MoE Routing are general modeling paradigms, not inherently indicators of incremental innovation. By this criterion, any work built upon CVAE, Transformer, or MoE architectures could be considered incremental, regardless of its specific contributions.

Our innovations are coherent, task-driven modeling principles rather than arbitrary combinations of existing modules. CVAE, Attention, and MoE are the means of implementation, not the basis on which novelty should be judged. We believe novelty should be assessed through the problems addressed, the specific technical mechanisms introduced, and the capabilities enabled by their integration.

AuthorsOpenReview ↗

Response to Reviewer Btzo (Weaknesses 2/2): MAE/MSE inconsistency

Response: MAE/MSE Inconsistency

Revision note: We have corrected the MAE/MSE results for TimeGAN and TimeVAE throughout the affected tables. F1-SEDI and fidelity metrics already incorporated observation restoration and were correctly computed; therefore, these results remain unchanged. No changes were required to UniGIO or the evaluation results of other baselines.

We carefully examined all reported MAE/MSE results and identified the exact cause. The issue was strictly limited to the evaluation postprocessing of two baselines, TimeGAN and TimeVAE, where an observation-restoration step was inadvertently omitted. UniGIO, its evaluation, and all other baseline evaluations remain unaffected. The issue did not involve model training or the definitions of MAE/MSE. After correcting the two affected baselines, our main experimental findings, performance advantages, and conclusions remain unchanged.

1. Cause of the inconsistency

Our evaluation computes the sum of absolute or squared errors divided by the number of masked positions. Before evaluation, the intended procedure restores observed values using:

y_hat = y_hat * (1 - mask) + y_observed * mask

This ensures that prediction errors at observed positions are zero, so only masked positions contribute to MAE/MSE.

However, this step was inadvertently omitted for TimeGAN and TimeVAE. Consequently, their error numerators included all positions, while the denominators counted only masked positions.

Let N denote the total number of positions and M the number of masked positions, with M<N. Under this incorrect implementation, the Cauchy--Schwarz inequality only guarantees:

\begin{equation} \mathrm{MAE}^{2} \leq \frac{N}{M}\mathrm{MSE}, \end{equation}

rather than the standard inequality:

\begin{equation} \mathrm{MAE}^{2}\leq \mathrm{MSE}. \end{equation}

For example, consider a mask [1,1,0,0] and absolute errors [2,2,2,2]. The incorrect evaluation gives MAE =4 and MSE =8, yielding 4^2=16>8. After restoring observed values, the errors become [0,0,2,2], giving MAE =2 and MSE =4, which satisfies the expected inequality.

2. Correction and verification

We systematically checked all reported results and identified 32 inconsistent MAE/MSE pairs, exclusively involving TimeGAN and TimeVAE: 5 in Table 1, 4 in Table 5, 9 in Table 7, and 14 in Appendix Table F.1.

Since the generated outputs had been saved, we corrected the observation-restoration step and recomputed the affected metrics without retraining.

For example, in Table 1, the native-mask Temperature MAE/MSE of TimeGAN changed from 10.3/76.5 to 3.25/19.3, while its Dewpoint results changed from 6.98/41.5 to 2.69/16.2.

We also corrected the 72h/240h results in Appendix Table F.1. For example, TimeGAN's 240h native-mask Temperature MAE/MSE changed from 11.2/88.8 to 3.49/22.3.

After correction, all 32 previously inconsistent pairs satisfy the expected mathematical inequality. F1-SEDI and fidelity results were independently verified and remained unchanged, as their evaluation procedures already included observation restoration.

3. Impact on the conclusions

This implementation error was confined to the evaluation postprocessing of TimeGAN and TimeVAE. The UniGIO model, its reported results, and the evaluations of all other methods were unaffected.

We have updated the affected tables and re-evaluated the comparisons using the corrected results. UniGIO retains its leading accuracy performance across the corrected comparisons, and the central conclusions remain supported. The correction does not change the overall performance ranking or the main conclusions of the paper.

We sincerely thank the reviewer for identifying this issue, which helped us improve the correctness and reproducibility of our evaluation.

AuthorsOpenReview ↗

Response to Reviewer Btzo ( Questions 1/1): Handling Genuinely Missing Target Values in the Event Aligner

Response: Handling Genuinely Missing Target Values in the Event Aligner

Thank you for this question. Within the Weather-5K benchmark, native missing positions have corresponding manually quality-controlled reference values, allowing the training-time Event Aligner (EA) to construct posterior representations using complete reference sequences. This benchmark-level reference availability does not imply that ground truth is naturally available for missing observations in real-world deployment.

Specifically, Weather-5K provides quality-controlled in-situ observation sequences constructed from raw station records with reference to ERA5 and other observational sources. It preserves the original missing or invalid positions as native masks while providing corresponding reference values in the quality-controlled product. Accordingly, the posterior branch directly uses the complete, unmasked Weather-5K sequences during training removed when inference.

Our native-mask evaluation provides a direct test of transfer to realistic missingness patterns: models trained on curated observations are evaluated on the original missingness distribution without additional training. This complements curated-mask evaluation, which measures reconstruction performance under controlled missing patterns with originally valid observations.

For further details on dataset construction, please refer to Han et al. (2026), Benchmarking Physics-Informed Time-Series Models for Operational Global Station Weather Forecasting, ICML 2026.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Note to the Reviewers and the Editor: Please consider our responses together with the revised manuscript

We sincerely thank the reviewers and the Editor for their valuable time, constructive feedback, and thoughtful suggestions. We have uploaded a thoroughly revised version of the manuscript in pdf, in which major modifications are highlighted in blue for ease of reference.

Due to the substantial nature of the revisions, many changes involve rewriting or adding entire paragraphs and sections, while the expanded experimental comparisons require modifications across multiple tables. Consequently, it is difficult to reproduce all revised content in the individual rebuttal responses without excessive duplication. We have therefore provided specific explanations and references to the corresponding sections, figures, and tables throughout our responses.

We sincerely encourage all reviewers and the Editor to consider our point-by-point responses together with the revised manuscript (please download the pdf), as the full extent of our revisions and additional contributions may not be fully conveyed by the responses alone.

We have made every effort to address the concerns raised, strengthen the methodological contributions, expand the experimental evidence, and improve the clarity and rigor of the manuscript. We genuinely hope that these revisions adequately resolve the reviewers' concerns and that the improvements will be reflected in your final assessments and evaluations.

Thank you again for your consideration and for giving us the opportunity to improve our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.