acceptodds
Under review as a conference paper at ICLR 2027

StreamDA: Recurrent Spatiotemporal Modeling for Hourly Global Atmospheric Data Assimilation

Abstract

Data assimilation, as the core of numerical weather prediction, is a computationally complex and extremely time-consuming process. In recent years, deep learning-based methods have substantially reduced the computational cost of data assimilation. However, existing deep learning methods still exhibit limitations in the temporal modeling of multi-temporal observations and the efficient fusion of multi-source heterogeneous observations. We propose StreamDA, a recurrent spatiotemporal modeling framework for hourly global data assimilation, powered by a Memory-based Temporal Fusion (MTF) module and a Coverage Modulation for Observation Fusion (CMOF) module. MTF retrieves temporal context from a sequence of historical latent analysis states and propagates it across time, achieving effective integration of multi-temporal observations under hourly cycling assimilation. CMOF combines local coverage with source-specific learnable weights to dynamically modulate observational contributions, improving spatial feature fusion of multi-source observations. In global assimilation experiments, StreamDA reduces mean relative RMSE across variables by 18.2% compared with the primary baseline model, with lower errors in 67 of 69 output variables. Furthermore, the model exhibits greater long-term stability and generates hourly, high-refresh analysis fields, providing robust support for real-time weather forecasting.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Data assimilation, Recurrent latent state modeling, Multi-source observation fusion, AI for weather

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_bSWXOpenReview ↗
Rating
2(−1)
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies an important problem of data assimilation in numerical weather prediction. The authors propose StreamDA, a recurrent spatio-temporal modeling for hourly global data assimilation, featuring memory-based temporal fusion and coverage modulation. Generally, this paper is well-written and easy to follow. The method is technically sound. In addition, experiments show the effectiveness of the proposals.

Critical strengths
  1. This paper is well-written and easy to follow. The structure is clear.
  2. A recurrent spatio-temporal modeling framework is proposed for hourly global data assimilation in numerical weather prediction, consisting of memory-based temporal fusion, coverage modulation, and observation fusion. Extensive experiments show the effectiveness of the proposals.
Critical weaknesses
  1. Although data assimilation is very important in numerical weather prediction, many new researchers may not know it clearly and may not know why it is important. It would be better to include more background at the beginning of the abstract and introduction.
  2. Efficiency plays an important role in weather prediction. It is suggested to include a comparison of the efficiency, i.e., training time, inference time, FLOPs, and memory usage, of the baselines and the proposed StreamDA. It would be better to include more baselines as follows.

[1]. Climate model configurations of the ECMWF Integrated Forecasting System (ECMWF-IFS cycle 43r1) for HighResMIP, Geoscientific model development 2018.

[2]. Learning skillful medium-range global weather forecasting, Science 2023.

[3]. Accurate medium-range global weather forecasting with 3D neural networks, Nature 2023.

[4]. Scaling transformer neural networks for skillful and reliable medium-range weather forecasting, NeurIPS 2024.

Questions for authors

Please see the weaknesses.

9 comments
AuthorsOpenReview ↗

Response to Reviewer bSWX

We thank the reviewer for the positive assessment and constructive suggestions. We have revised the Abstract and Introduction to more clearly explain the role and importance of data assimilation in numerical weather prediction.

We have also expanded the computational efficiency comparisons. StreamDA shows advantages in peak inference memory usage (GiB), computational cost per analysis (TFLOPs), forward-pass time per analysis (seconds), and training cost (GPU-hours). Detailed results and measurement settings are provided below.

Q1. Background on data assimilation

Thank you for your suggestion. We have added background on data assimilation at the beginning of the Abstract and Introduction. We explain how it combines observations with a background field to produce an analysis field for forecast initialization, and why analysis quality matters for forecast performance. We also describe the alternating assimilation and forecast cycle, along with the computational cost and update-frequency limitations of traditional assimilation methods. These additions help readers unfamiliar with the field understand the motivation for our work.

Q2. Computational efficiency and baselines

Thank you for your suggestion. We have added a comparison of the training and inference costs of StreamDA and HealDA. The first table reports training time, inference time, FLOPs, peak inference memory, GPU-hours, and hardware configurations. StreamDA requires less inference time, computation, and peak GPU memory. This difference partly stems from how the two methods process historical observations. HealDA processes the full observation window at each analysis step, repeatedly encoding observations shared by adjacent windows. StreamDA retains historical information through recurrent states and memory and incorporates new observations at each update. This reduces repeated processing and helps lower the cost of continuous assimilation.

MetricStreamDAHealDA
Inference time (s/analysis)0.3270.550
FLOPs (TFLOP/analysis)6.4823.393
Peak inference memory (GiB)2.388.69
Training time (hours)63.32~200
Training cost (GPU-hours)253.28~1,600
Training hardware4 × H8008 × H100

We have also added downstream forecast evaluations with Aurora and Stormer alongside Pangu-Weather. The second table reports the mean relative RMSE increase compared with ERA5-initialized forecasts, with equal weight given to all 69 variables. For all three forecast models, StreamDA initialization gives a smaller gap to the ERA5-initialized reference than HealDA* initialization. These results further demonstrate StreamDA's applicability across different forecast models.

Lead timePangu: HealDAPangu: StreamDAAurora: HealDAAurora: StreamDAStormer: HealDAStormer: StreamDA
1 day53.4832.5754.8636.7950.5527.83
3 days25.6917.6121.6215.0322.4513.61
5 days16.0311.7913.249.7313.679.10
7 days10.217.798.556.219.006.30
10 days5.104.133.942.795.383.71

Thanks for the response. However, it is not enough to only compare StreamDA with HealDA. It is required to show the comparisons between StreamDA and all baselines.

AuthorsOpenReview ↗

Response to Reviewer bSWX

Dear Reviewer bSWX,

Thank you for your further feedback. Following your suggestion, we have added FuXi-Weather to the comparison. We now report the computational costs of StreamDA, HealDA, and FuXi-Weather, as well as the accuracy of downstream forecasts initialized from their analysis fields.

1. Computational efficiency (Appendix B.3, Table 6)

The table below compares the training and inference costs of the three assimilation methods. StreamDA requires the fewest FLOPs per analysis and the least peak inference memory.

MetricStreamDAHealDAFuXi-Weather
Inference time (s/analysis)0.3270.5500.594
FLOPs (TFLOP/analysis)6.4823.3939.130
Peak inference memory (GiB)2.388.695.59
Training time (hours)63.32~200~80
Training cost (GPU-hours)253.28~1,600~320
Training hardware4 × H8008 × H1004 × A100

2. Downstream forecasts (Appendix B.2, Figures 7–8)

We initialized Pangu-Weather, Aurora, and Stormer with analysis fields from each of the three assimilation methods. For each forecast model, we compared all three initializations at the same forecast start times. The table below reports the mean relative increase in forecast RMSE (%) compared with ERA5 initialization. Lower values are better.

Forecast modelInitializationDay 1Day 3Day 5Day 7Day 10
Pangu-WeatherStreamDA32.5717.6111.797.794.13
HealDA53.4825.7016.0310.205.11
FuXi-Weather63.7739.3727.0017.318.09
AuroraStreamDA36.8015.039.726.202.80
HealDA54.8821.6213.248.553.94
FuXi-Weather67.6031.9220.9013.185.84
StormerStreamDA27.8313.619.096.293.71
HealDA50.5622.4513.679.005.39
FuXi-Weather65.7538.2326.8518.7311.22

We have incorporated these comparisons into the revised manuscript. Thank you again for helping us improve the paper.

Best regards,
The Authors

Reviewer_2Rn1OpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a weather data assimilation method that incorporates observations from multiple sources (satellites, stations, etc.) through a Coverage Modulation for Observation Fusion (CMOF) module: per-source point features are mapped onto the grid with learned distance-dependent weights, and fused according to local grid occupancy to better handle sparse coverage. A memory-based temporal fusion (MTF) module provides a latent background from past analysis states. The method cycles hourly: a latent background is predicted sequentially and combined with the encoded observations, then passed to a shared Swin Transformer backbone that updates the latent analysis state. The authors report an 18.2% RMSE reduction over HealDA and improved robustness to missing observations.

Critical strengths

CMOFCompared with uniform source fusion as in HealDA, weighting each source by its local coverage substantially reduces errors for geopotential and mean sea-level pressure (≈13%, Table 2), and markedly improves robustness when sources are removed. This matters because missing sources are frequent in operational settings. Note that the gain comes from the coverage term CtsC_t^s rather than the per-source scaling αs\alpha_s (Table 8).

Sequential assimilation beats joint window processing

The Window-6h experiment (Sec. 4.4.1) compares sequential hourly updates with joint processing of the same six-hour observation window, with no additional history and the same observation encoder. Sequential processing reduces RMSE by 7.6% (8.4% during observation gaps). This is important because the amount of information is held fixed: the gain comes from integrating observations through successive hourly state updates rather than pooling the whole window in a single pass.

Critical weaknesses

Main comparison with HealDA is not reliable

The headline result (18.2% RMSE reduction) relies on a comparison with HealDA that is not controlled: HealDA is evaluated 1-2 years after the end of its training period, with different observation sources, unspecified inputs, and a regridding that is not described. The HealDA errors reported in Table 1 (e.g. Z500 RMSE 74.99 m2/s2m^2/s^2) are much higher than those reported in the HealDA paper itself that were much closer to StreamDA’s own value. The authors acknowledge that this comparison does not isolate their contribution (Sec. 4.4.1). Moreover, the evaluation is much thinner than in HealDA: RMSE against ERA5 only, without error bars, spectral analysis, comparison to an operational analysis, or ensemble/probabilistic downstream metrics. The downstream forecasts use only Pangu-Weather, which the HealDA paper reports to be unstable with 1° initial conditions, and no ERA5-initialized reference is provided. More generally, since the end goal of data assimilation is to initialize forecasts, and the paper claims to support real-time weather forecasting, forecast skill should be a central part of the evaluation rather than an appendix experiment: it should involve standard forecast models (e.g. GraphCast, FCN3 or Aurora), and the paper shows no qualitative examples of analysis fields or forecasts. As a result, the paper’s main quantitative claims cannot be compared with published numbers.

Limited contribution and positioning

Cycling with a background was already shown to improve accuracy and robustness to missing observations in FuXi Weather, and the HealDA authors report a similar gain (∼10%\sim 10\%) from adding a background. HealDA also already assimilates heterogeneous multi-source observations with sensor-specific encoders. The specifically new components bring small gains: memory retrieval improves over mean pooling by ∼1%\sim 1\% (Table 7), and αsα_s has no consistent effect (Table 8). The method remains a deterministic point estimator trained with an L1 loss, without uncertainty quantification, while the paper cites generative DA approaches (SDA, DiffDA, FlowDA) without discussing or justifying this choice. Finally, the source-specific encoders make it unclear how new instruments or changes in existing ones could be handled without retraining: robustness to source removal is tested, but not to the addition of new sources. These limitations, central for operational use, are barely discussed, and the conclusion is very short.

Questions for authors
  1. I would like to see whether the authors can reproduce HealDA’s published results on its own test year (2022) with their evaluation pipeline. Which observations did you feed to HealDA for 2023–2024, and how did you regrid its HEALPix outputs to the 1° lat-lon grid? Without this, I find it hard to trust the main comparison.

  2. Could you report error bars across several training seeds, or ensemble/probabilistic metrics (e.g. CRPS of downstream ensemble forecasts, as in HealDA)? Many of the reported differences, especially in the ablations, are below 1%, and I cannot tell whether they are significant.

  3. How would the model handle a new observation source, or a change in an existing instrument (e.g. a new satellite platform)? Would retraining only the source-specific encoder be enough? You show robustness to removing sources, but in an operational setting, adding new ones seems at least as important.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 2Rn1(Part 1)

We thank the reviewer for recognizing the effectiveness of StreamDA’s temporal modeling and observation fusion modules. MTF improves performance by about 7.6%, reaching 8.4% during observation gaps, while observation fusion reduces errors in geopotential and mean sea-level pressure by about 13%. In response to your concerns, we added a controlled comparison with HealDA using the same observation sources, observation types, and test period, but less training data. StreamDA* achieved mean relative RMSE reductions of 8.18% on Grid and 6.92% on HPX64. The added significance tests further support the consistent gains in our ablation studies, which also help guide future research. We also added downstream forecast evaluations with Aurora and Stormer, along with power spectrum comparisons of the analysis fields. Our detailed responses follow.

Q1. Reproducing HealDA and documenting the evaluation

Thank you for this suggestion. Using our evaluation pipeline, we reproduced HealDA’s published results for 2022, as shown in the table below. For HealDA inference, we used observation data downloaded from its official data source. We have also added the details of mapping HealDA’s HEALPix outputs to the latitude–longitude grid in Appendix H.1. For a controlled comparison, we instead retrained StreamDA using the same GDAS observation source and types as the official HealDA setup (excluding IASI, HIRS, and SSMI). Notably, StreamDA was trained on only six years of data (2016–2021),which is far less than HealDA’s 22-year training set (2000–2021). We refer to this retrained model as StreamDA*.

The retrained StreamDA outperforms the official HealDA baseline on the 2022 test set, reducing overall ERA5-referenced RMSE by 8.18% ( latitude–longitude grid ) and 6.92% ( HPX64 grid ). For July 2023–June 2024, these ERA5-referenced reductions further widen to 11.96% and 10.47%, respectively, providing additional evidence of StreamDA’s ability to generalize over longer time horizons. Full experimental details and results have been added to the appendix of the revised manuscript.

Variable (unit)Grid: HealDAGrid: StreamDA*HPX64: HealDAHPX64: StreamDA*HealDA (published, HPX64)
Z500 (m²/s²)44.6340.9244.8539.2046.60
T850 (K)0.790.790.700.730.70
U500 (m/s)1.811.801.681.671.68
T2m (K)1.030.790.700.69—
U10m (m/s)0.860.680.650.58—

Q2. Training seeds and small ablation gains

Thank you for your suggestion. To assess whether the improvements are statistically significant, we added ablation significance tests using 10,000 paired bootstrap resamples of individual six-hourly analysis times on the main-paper test set. The full model achieved a mean relative RMSE reduction of 1.10% compared with mean pooling, with a 95% confidence interval of [1.077%, 1.128%], and 0.21% compared with the version without source-level weights αs\alpha_s, with a 95% confidence interval of [0.178%, 0.238%]. Both improvements are statistically significant.

Q3. New observation sources and instruments

We agree that integrating new observation types into operational weather forecasting systems is challenging. For example, Goldberg et al. (2013) [1] noted that assimilating data from a new instrument into NOAA’s operational forecast models typically required roughly two years of software development and testing. Similarly, incorporating entirely new observation types into AI-based assimilation systems demands representative training data and model adaptation. We thank the reviewer for raising this point and have added a discussion of these challenges and our model’s corresponding limitations to the Conclusion of the revised manuscript.

In addition, to evaluate zero-shot adaptation to a new satellite platform carrying an existing sensor type, we conducted an additional experiment incorporating ATMS observations from NOAA-21, which became available after August 2023 and were therefore absent from our training set. Without any retraining or fine-tuning, adding NOAA-21 observations reduces overall RMSE by an average of 0.53% compared to the baseline without them. This demonstrates StreamDA’s ability to immediately leverage data from a previously unseen platform.

Goldberg. 2013. Joint Polar Satellite System The United States next generation civilian polar orbiting [1]

AuthorsOpenReview ↗

Response to Reviewer 2Rn1(Part 2)

Q4. Downstream forecast evaluation

We thank the reviewer for raising this point. We have expanded our downstream evaluation to include Aurora and Stormer. The table below reports the mean relative RMSE increase (%) across 69 variables compared with ERA5-initialized forecasts. Lower values indicate a smaller gap to the ERA5-initialized reference. StreamDA initialization consistently narrows this gap compared with HealDA, supporting its effectiveness beyond Pangu-Weather.

We agree with HealDA that downstream forecast models exhibit model-dependent sensitivity to the characteristics of their initial analysis fields. Understanding this interaction is an important direction for future research. We also acknowledge that forecasts initialized from either HealDA or StreamDA analysis fields still show a performance gap compared to ERA5-initialized baselines, leaving room for further improvements in both analysis quality and downstream forecast skill. We have added more discussion in the Conclusion section.

Lead timePangu: HealDAPangu: StreamDAAurora: HealDAAurora: StreamDAStormer: HealDAStormer: StreamDA
1 day53.4832.5754.8636.7950.5527.83
3 days25.6917.6121.6215.0322.4513.61
5 days16.0311.7913.249.7313.679.10
7 days10.217.798.556.219.006.30
10 days5.104.133.942.795.383.71

Q5. Analysis fields and power spectra

We have added case comparisons of the two models’ analysis fields in Appendix G.1 and power spectrum comparisons in Appendix G.2. Both models capture the main large-scale structures. Over some high-wavenumber ranges, StreamDA’s power spectra are closer to ERA5, providing additional evidence of its representation of fine-scale features.

Q6. Module contributions

Thank you for your comments. As you noted in your review, the experiments show clear performance gains from MTF and CMOF. Tables 8 and 9 test individual designs within these modules, so their gains do not represent the full benefits of MTF and CMOF. We have added significance tests to these ablation studies, confirming that these designs provide consistent improvements. These experiments also help guide future research. For example, Table 8 compares different ways of combining historical information. The results show that beyond keeping historical information, better ways to select and use it can further improve temporal modeling.

Q7. Deterministic estimation

Thank you for the suggestion.

Uncertainty quantification has substantial practical value, but deterministic analyses also remain an important foundation of operational weather forecasting. For example, HRES and GFS both use deterministic analysis fields obtained through data assimilation to initialize deterministic forecasts. SDA, DiffDA, and FlowDA offer valuable explorations of generative data assimilation, but their reported experiments primarily use idealized systems or simulated observations derived from reanalysis data. They have not yet demonstrated the continuous ingestion and assimilation of real-world, heterogeneous observation streams considered in this work. We therefore believe that deterministic data assimilation for this setting remains valuable. Nevertheless, we acknowledge that StreamDA has not yet been extended to ensemble data assimilation and cannot provide uncertainty estimates for its analysis fields. We will explicitly state this limitation in the Conclusion and identify ensemble data assimilation and uncertainty quantification as directions for future research.

Reviewer_2Rn1OpenReview ↗

Response to the Authors by Reviewer 2Rn1

Dear authors,

I would first like to thank you for your detailed response and for the substantial improvements made to the manuscript following my comments. However, I believe there are still a few major concerns.

Q1. Controlled comparison

Training StreamDA* on data comparable to HealDA’s is, in my view, what makes it possible to assess the actual contribution of the method, including CMOF and MTF. The original comparison could not separate the benefits of your method from those of more recent training data and additional observation sources. I thank you for this controlled experiment, which largely mitigates my concern and which also shows that the large gap in the original comparison stems from a degradation of HealDA beyond its training period. However, this experiment shows an RMSE reduction of only 8.18% (latitude–longitude grid) and 6.92% (HPX64 grid) relative to HealDA, whereas the abstract and main text still report the uncontrolled result (18.2%). I would ask the authors to move the controlled comparison to the main text and abstract, and to present the full-system comparison as a secondary result.

Q3. New instruments

I thank the authors for the NOAA-21 experiment and the added discussion in the conclusion. As a side note, modular learned approaches offer an interesting alternative here: with a generative prior and explicit observation operators (e.g. score-based DA), a new instrument can in principle be incorporated without retraining the prior, provided an observation operator and error model are available. More generally, the revised conclusion mentions uncertainty estimation only briefly as future work; a short discussion of how generative or ensemble DA approaches relate to StreamDA would strengthen the positioning of the paper.

Q4. Downstream forecasts

The added experiments with Aurora and Stormer and the ERA5-initialized reference are appreciated. However, the comparison still pits the full StreamDA (more sources, trained until 2023) against HealDA* evaluated well beyond its training period, so the size of the advantage reflects the same confounders as the main comparison. Since StreamDA* is already available, I would ask the authors to report downstream forecasts for StreamDA* vs HealDA*, which only requires additional inference.

Q5. Spatial structure

The analysis maps, power spectra and verification against independent MISR winds address my concern about the evaluation relying only on ERA5.

In light of the controlled comparison and the additional evaluations, I raise my soundness score to 2 and my overall rating to 2 (weak reject), as the methodological contribution remains limited relative to prior work and the main result is still framed around the uncontrolled comparison.

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

In data assimilation for climate science, we want to use observations to predict an analysis field from which to forecast future weather movements. Since observations come in many modalities and are not sampled at the same rate, the authors propose a novel method of fusing observations, by computing point features and then aggregating them over a grid, using MLPs to parameterize/guide the process. They also suggest using an attention mechanism to retrieve past analysis states, so that they may be used to help compute future forecasts. The proposed method achieves lower RMSE than all tested benchmarks, and ablation studies imply that the techniques are sound.

Critical strengths
  1. The paper clearly responds to actual, pressing problems in the field.
  2. At a high-level, the proposed methods are physically interpretable. Both the coverage modulation and memory-based fusion methods have clear spatial/temporal/physical meaning, and the overall pipeline is much less opaque than usual deep learning methods.
Critical weaknesses

Although interpretable at a high-level, the solution pipeline contains many components and is quite complicated. It seems plausible that a simpler method could achieve similar results. A number of components are lacking mathematical and physical explanation. See the questions for examples.

Questions for authors

Can you justify or elaborate certain mathematical choices? Particularly:

  • The weight mapping function is not clearly motivated. Why is an (I am assuming MLP) used instead of just a weighted softmax?

  • Explain the normalization term in the observation field expression, the square root denominator expression is not justified.

  • Why the convolution terms in the memory-based temporal fusion expression for Ht−1H_{t-1}?

  • The H in the H * cos term in the training loss will cancel with the H in the H*W denominator outside the sum. Is this correct?

  • What is the Patch Projection you mentioned in the appendix? Can you explain the other terms in this expression? Why the log terms in the concatenation?

The number of compared models in experiments seems limited.

In general, the approach needs much more motivation and justification. The methods section is currently just a list of things done, with little to no explanation and no theory.

3 comments
AuthorsOpenReview ↗

Responses to Reviewer swGt

We thank the reviewer for recognizing the importance of the problem we address. Since deep learning–based data assimilation baselines for real-world observations remain limited, we hope this work encourages more researchers to explore this direction. Meanwhile, as you rightly noted, StreamDA is a physics-inspired deep learning framework. Accordingly, we have revised the Methods section to clarify the motivation and design rationale behind each module. We have also added statistical significance tests to verify the robustness of our ablation gains, especially controlling experiments in each submodule. More importantly, we hope these hierarchical ablation results provide researchers with deeper insights, helping pinpoint the most critical directions for future improvement.

Q1. Weight mapping and softmax

Thank you for the suggestion. Weight mapping allows information from an observation point to reach surrounding grid cells, rather than only the cell containing it. This is particularly useful when observations are sparse, as cells without direct observations can still benefit from nearby observations. Since an observation’s influence on surrounding cells may depend on its content and relative position, we use an MLP to learn this relationship rather than prescribe fixed weights. Softmax can normalize the resulting scores, but it does not determine how an observation influences surrounding cells and therefore cannot directly replace the learned weight mapping.

Our implementation uses softplus followed by normalization. We will clarify this motivation in Methods.

Q2. The square root normalization term

Thank you for the suggestion. Adding features from multiple observation sources can increase their variance, leading to larger fluctuations in the fused features. For example, summing SS independent features with equal variance increases the variance by a factor of SS; dividing the sum by S\sqrt{S} restores the original variance. This motivates using square-root normalization to control the scale of the fused features.

Q3. Convolution terms in temporal fusion

Thank you for the question. The two residual terms provide historical information and recent changes, respectively: Rt−1−St−1R_{t-1}-S_{t-1} represents the difference between the retrieved historical state and the current state, while St−1−St−2S_{t-1}-S_{t-2} represents the state change over the past hour. We use convolutions to incorporate information from neighboring grid cells and transform these differences into corrections to the current state, rather than simply adding the raw differences. Since the two terms convey different information, we use separate convolutions to learn how each should influence the next prediction. We will clarify the meanings of these residual terms and this design motivation in Methods.

Q4. Cancellation of H in the loss

Thank you for pointing this out. Yes, the two H factors cancel. We will simplify the expression accordingly; this does not change the latitude-weighted loss.

Q5. Patch Projection and the remaining terms

Thank you for the question. This design provides the backbone with both local observation content and sampling information. The weighted-average feature OtsO_t^s represents the observation content, but does not distinguish between missing, sparse, and dense observations. We therefore also include the valid-observation indicator ItsI_t^s, observation count NtsN_t^s, and sum of weights before normalization WtsW_t^s, allowing the model to use this information when processing the observation features. Patch Projection then combines neighboring grid features into patch features for the backbone, enabling it to jointly process observations and their spatial distribution within a local region. It is implemented as a 2×22 \times 2 convolution with stride 2, mapping each non-overlapping grid patch to a feature vector. We will add this motivation, the architectural details, and the symbol definitions to the appendix.

Q6. Log terms in the concatenation

When mapping observations onto the grid, we compute a weighted average of the observation features mapped to each grid cell, yielding OtsO_t^s. However, this operation loses information about the observation count and total weight. For example, a single observation and multiple observations may produce the same averaged feature. We therefore concatenate the observation count N and the sum of weights before normalization W , allowing the model to distinguish these cases and use this information when processing the averaged features. Since these statistics can vary greatly across grid cells, we apply log⁡(1+x)\log(1+x) to compress large values and prevent statistics from a few densely observed cells from dominating subsequent computations. This transformation preserves their ordering and maps zero to zero when no observations are available. We will add this explanation to the revised manuscript.

Dear Reviewer swGt,

Thank you again for your time and constructive feedback. We have provided point-by-point responses and additional experimental results, and uploaded a revised manuscript with changes highlighted in blue. We hope these updates address your concerns.

As the discussion deadline approaches, we would appreciate it if you could review our replies when your schedule permits. Please let us know if anything remains unclear or requires further clarification. We would be happy to continue the discussion and further improve our manuscript. If our responses have addressed your concerns, we would be grateful if you could re-evaluate our work.

Best regards,
The Authors

Thanks to the authors for the rebuttal and clarifications for questions. These explain some aspects but my major concerns still remain that the approach needs more motivation and theoretical justifications for its design and synthesis.

Reviewer_e8k8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents StreamDA, an hourly recurrent model for global atmospheric data assimilation. It combines CMOF, which puts thirteen observation sources onto a common grid, with a recurrent memory module that carries information between hourly cycles. The model is trained on ERA5 and evaluated with year-long cycling, ablations, observation outages, and Pangu forecast experiments.

Critical strengths
  1. This is an ambitious experiment. Running the system hourly over a full held-out year, with thirteen observation sources and 69 output variables, is not trivial. The continuous cycling test is especially useful.

  2. I also found the ablations reasonably helpful. In particular, the six-hour control and the memory-length experiments make it easier to see what the recurrent part is adding.

Critical weaknesses
  1. My main concern is that Table 1 is not really an apples-to-apples comparison. HealDA was trained with different years, observations, preprocessing, and cycling choices, and FuXi-Weather is run at a different resolution. So the reported 18.2% gain may come partly from the data pipeline rather than StreamDA itself. The Window-6h experiment helps, but it does not fully settle this.

  2. I am also not fully convinced by using ERA5 as the only verification target. Many of the input observations already contribute to ERA5, so lower RMSE can mean better imitation of the reanalysis rather than a better estimate of the atmosphere. A check against withheld observations, or at least innovation statistics, would make the result much more convincing.

  3. The uncertainty analysis feels rather thin for a year-long recurrent experiment. It looks like the results come from one checkpoint, and the hourly samples are strongly correlated. The outage test is also quite narrow: one removal order and outages of only eight hours. Finally, CMOF seems to learn coverage plus a source-level weight; I do not see how this supports the broader claim about changing observation reliability.

Questions for authors
  1. Could the authors train at least one baseline with the same years, observations, resolution, and preprocessing? Without that, it is hard to tell how much of the gain comes from StreamDA.

  2. Is it possible to evaluate against observations that were not used in ERA5, and to report uncertainty with either multiple runs or a block bootstrap over time?

  3. I may be missing something here, but how exactly does CMOF distinguish unreliable observations from areas with sparse coverage? Do the outage results change with a different removal order or a longer outage?

  4. There are also a few small wording issues: "for each variables" after Equation 13, "Algorithms 2" in Section 4.2, and the unclear "we select its outputs" in the same paragraph. Could these be cleaned up?

7 comments
AuthorsOpenReview ↗

Response to Reviewer e8k8(Part 1)

We thank the reviewer for recognizing our proposed recurrent temporal modeling approach, the technical challenges involved, and our substantial effort. We greatly appreciate your constructive feedback and address each point below.

Q1. Matched baseline training

Thank you for this suggestion. Retraining HealDA on our dataset under its original configuration would require over 1,600 GPU-hours, exceeding our rebuttal compute budget. For a controlled comparison, we instead retrained StreamDA using the same GDAS observation source and types as the official HealDA setup (excluding IASI, HIRS, and SSMI). Notably, StreamDA was trained on only six years of data (2016–2021), which is far less than HealDA’s 22-year training set (2000–2021). We refer to this retrained model as StreamDA*.

The retrained StreamDA outperforms the official HealDA baseline on the 2022 test set, reducing overall ERA5-referenced RMSE by 8.18% ( latitude–longitude grid ) and 6.92% ( HPX64 grid ). For July 2023–June 2024, these ERA5-referenced reductions further widen to 11.96% and 10.47%, respectively, providing additional evidence of StreamDA’s ability to generalize over longer time horizons. Full experimental details and results have been added to the appendix of the revised manuscript.

Variable (unit)Grid: HealDAGrid: StreamDA*HPX64: HealDAHPX64: StreamDA*
Z500 (m²/s²)44.6340.9244.8539.20
T850 (K)0.790.790.700.73
U500 (m/s)1.811.801.681.67
T2m (K)1.030.790.700.69
U10m (m/s)0.860.680.650.58

Q2. Independent observations and uncertainty

We have added evaluations against both conventional and independent MISR observations [1] to the appendix, and applied a paired bootstrap to quantify the statistical uncertainty of the performance differences between the two models. Key details are provided below.

We have added a bootstrap analysis to quantify the statistical uncertainty of the performance differences between StreamDA* and HealDA against ERA5. On the 2022 test set, we performed 10,000 paired bootstrap resamples. StreamDA* achieved a mean relative RMSE reduction of 8.18% over HealDA, with a 95% confidence interval of [8.02%, 8.34%]. Improvements were statistically significant for 59 of 69 variables.

Evaluations against conventional observations assimilated into ERA5 showed consistently lower RMSE for StreamDA* in 2022. Following the ECMWF study by Lean et al. (2026), we also evaluated both models against independent MISR wind observations that were not assimilated into ERA5. StreamDA* showed consistent improvements across all altitude ranges. Detailed results are provided below. Bold indicates lower error; † indicates significance at the 95% level based on 10,000 paired bootstrap resamples.

Observation typeVariable (unit)HealDAStreamDA*RMSE reduction
Surface (SURF)MSL (hPa)3.4113.379†0.92%
Surface (SURF)T2m (K)2.5132.439†2.93%
Surface (SURF)U10m (m/s)2.0101.957†2.65%
Surface (SURF)V10m (m/s)1.9421.882†3.10%
Radiosonde (TEMP)T (K)1.2191.185†2.74%
Radiosonde (TEMP)U (m/s)2.7922.677†4.11%
Radiosonde (TEMP)V (m/s)2.7772.447†11.90%
Altitude (km)RMSVD: HealDARMSVD: StreamDA*U RMSE: HealDAU RMSE: StreamDA*V RMSE: HealDAV RMSE: StreamDA*
0–23.6433.609†2.2152.190†2.8932.868†
2–45.5555.525†3.7083.691†4.1364.112†
4–65.6225.576†3.4673.437†4.4264.390†
6–86.5516.455†4.0233.958†5.1705.099†
8–107.8717.729†4.8184.712†6.2246.126†
10–1510.42810.342†5.3175.234†8.9708.920
0–15 (Overall)4.5064.465†2.7932.766†3.5363.506†

Reference [1]: Global reanalysis from observations alone with machine learning. https://arxiv.org/pdf/2607.07879v1

Responses to Q3 and Q4 follow in Part 2.

AuthorsOpenReview ↗

Response to Reviewer e8k8(Part 2)

Continued from Part 1.

Q3. Observation reliability and outage robustness

CMOF accounts for both local observation sparsity and overall differences in source contributions. Specifically, spatial coverage Cs,tC_{s,t} measures the fraction of grid cells within each patch containing valid observations, characterizing local data availability and dynamically modulating source features during fusion. In contrast, the learned source-level weight αs\alpha_s, shared across locations and times, controls each source’s overall relative contribution.

Regarding observation reliability, we clarify that αs\alpha_s represents learned source-level importance rather than an explicit estimate of observation uncertainty. While αs\alpha_s implicitly reflects both a source’s informational value and its typical error characteristics under the training distribution, it cannot dynamically detect or respond to abrupt error spikes caused by anomalous or corrupted data. Quality control (QC) is still necessary to filter out invalid observations. After QC, Cs,tC_{s,t} adjusts the fusion weights to reflect the remaining valid coverage, while αs\alpha_s provides source-level scaling.

To make this distinction explicit, we have replaced “observation reliability” with “observation importance” in the revised manuscript and clarified its meaning and limitations.

We extended the outage experiments to 100 random removal orders across dates spanning all 12 months, with outages lasting 8, 24, and 48 hours. Averaged over dates and removal orders, StreamDA consistently showed a smaller relative RMSE increase than the variant w/o CMOF. Detailed results are provided in Appendix D.2, Figure 10.

Q4. Wording and references

Thank you for pointing these out. We have corrected these wording issues in the revised manuscript and carefully polished the entire paper.

Dear Reviewer e8k8,

Thank you again for your time and constructive feedback. We have provided point-by-point responses and additional experimental results, and uploaded a revised manuscript with changes highlighted in blue. We hope these updates address your concerns.

As the discussion deadline approaches, we would appreciate it if you could review our replies when your schedule permits. Please let us know if anything remains unclear or requires further clarification. We would be happy to continue the discussion and further improve our manuscript. If our responses have addressed your concerns, we would be grateful if you could re-evaluate our work.

Best regards,
The Authors

Comments on the paper

2 comments

Dear Reviewers,

As the discussion period is coming to an end, I kindly ask all reviewers who have not yet responded to the authors' rebuttal to carefully review their responses and provide feedback or update their reviews accordingly.

Thank you for your time and efforts!

We thank all four reviewers for their careful evaluation of StreamDA and their constructive feedback. We especially appreciate their recognition of the following strengths.

  1. The Memory-based Temporal Fusion (MTF) module explicitly models the temporal information in observations and improves analysis accuracy. In a controlled experiment using the same observation encoder and the same six-hour observation set, StreamDA recursively updates the analysis state on an hourly basis. Compared with jointly processing the entire observation window, this approach reduces RMSE by 7.6%, further demonstrating the importance of temporal modeling for observations. (2Rn1, e8k8)

  2. The Coverage-Modulated Observation Fusion (CMOF) module improves analysis accuracy and robustness to observation outages. Compared with the corresponding baseline, CMOF adjusts the contribution of each observation source according to its local coverage. This reduces errors in geopotential and mean sea-level pressure by approximately 13% and substantially limits error increases during observation outages, thereby improving the system’s robustness in operational settings. (2Rn1, swGt)

  3. StreamDA maintains its accuracy advantage throughout year-long continuous hourly assimilation and demonstrates greater stability during periods of elevated error. In the controlled experiment with aligned observation types, StreamDA* reduces the mean relative RMSE by 8.18% compared with the baseline during the 2022 evaluation period. This reduction further increases to 11.96% on the subsequent test set used in the paper, providing additional evidence of the system’s stability during long-term operation. (e8k8, bSWX)

Following the reviewers’ suggestions, we added the following experiments and revised the manuscript accordingly.

  1. Controlled comparison and clearer model naming. We removed IASI, HIRS, and SSMI and retrained StreamDA* using the observation types aligned with HealDA. On the paper’s test set, StreamDA* reduces the mean relative RMSE on the latitude–longitude grid by 11.96% compared with HealDA. This controlled result is now reported in the abstract and Table 1. In response to the issue raised by 2Rn1 during the final round of discussion, we also standardized the model names and superscripts throughout the main text, appendices, figures, and tables so that readers can directly identify the configuration associated with each result. (e8k8, 2Rn1)

  2. Observation-based verification and spatial structure. We evaluated the analysis fields against surface stations, radiosondes, and independent MISR satellite wind observations that were not assimilated into ERA5. Across all altitude layers in the MISR evaluation, StreamDA achieves significantly lower root-mean-square vector differences (RMSVD) and lower RMSE for both zonal and meridional winds than HealDA and FuXi-Weather. The added analysis-field case studies illustrate the spatial distribution of errors, while the power spectra assess how well each model represents small- and medium-scale atmospheric variability. (e8k8, 2Rn1)

  3. Observation outages and significance tests. The experiments cover all 12 months, 100 random removal orders, and outages lasting 8, 24, and 48 hours. Across all three durations, StreamDA exhibits a smaller mean error increase than the variant without CMOF. We also conducted significance tests for the baseline comparisons and key ablation experiments. (e8k8, 2Rn1)

  4. Observation fusion and zero-shot transfer. Coverage reflects local sampling density, whereas source-level weights are learned during training. We also clarified observation encoding, fusion normalization, and memory updates. The NOAA-21 ATMS experiment demonstrates that StreamDA can use observations from the same instrument type on a new satellite platform without retraining. (swGt, e8k8, 2Rn1)

  5. Downstream forecasts. We added Aurora and Stormer alongside Pangu-Weather. At identical forecast start times, we compared forecasts initialized from the analysis fields of StreamDA, HealDA, and FuXi-Weather, using ERA5-initialized forecasts as a reference. (2Rn1, bSWX)

  6. Assimilation background and computational efficiency. In the abstract and introduction, we further explained the role of analysis fields in forecast initialization. Table 6 now compares the inference time, computational cost, and peak memory usage of the three data assimilation systems. (bSWX)

The experimental results and the locations of the corresponding manuscript revisions are detailed in our point-by-point responses. We again thank all four reviewers for helping us improve the paper. We welcome any further questions and would be pleased to continue the discussion and provide additional clarification.

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.