acceptodds
Under review as a conference paper at ICLR 2027

RISDA: Reliability-Informed Score-Based Data Assimilation of Multi-Source Atmospheric Observations

Abstract

Score-based atmospheric data assimilation separates a learned atmospheric prior from observation guidance, making it naturally flexible to changing observations. However, this flexibility breaks down for real-world heterogeneous multi-source observations, where the main bottleneck shifts to the guidance itself: different observations cannot always be directly matched to atmospheric states, and the estimated guidance can vary greatly in reliability across sources, regions, and latent channels. Therefore, we introduce RISDA (Reliability-Informed Score-based Data Assimilation), a framework designed for the assimilation of real-world multi-source observations. First, an atmospheric encoder and source-specific observation encoders map atmospheric states and heterogeneous observations into a shared latent space, bridging variable mismatches and enabling direct guidance for posterior sampling. Second, empirical diagnostics of latent errors motivate Variance-Calibrated Latent Guidance (VCLG), which constructs more accurate observation guidance by accounting for channel-specific uncertainties within the shared latent space. Third, to reasonably combine guidance from multiple sources during diffusion, we introduce Priority-Preserving Gradient Fusion (PPGF), which merges source-specific gradients according to reliability-informed priorities, ensuring that guidance from more reliable sources is preserved during sampling. Assimilation experiments incorporating five types of satellite observations with the surface observation demonstrate that RISDA achieves superior analysis accuracy and higher inference efficiency compared to baseline methods.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Score-based Data Assimilation, posterior sampling, generative models, gradient fusion

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_fZciOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents score-based DA approach for multi-sourced weather analysis. The method combines an error-correcting unified latent representation (over states and observations) with a gradient-conflict approach (from multi-task learning) for improved guidance.

Critical strengths
  • The multi-source setup of DA is very practical, so the need for a method that deals with this setup is important.

  • The experimental results with the recent diffusion model approach show improvements in terms of both run-time (Table 2) and the problem-specific metrics (Table 1).

  • Other results demonstrate the impact of every component of RISDA.

Critical weaknesses

Weaknesses/Questions/ Suggestions:

  • Minor: The main target of the paper is multi-source but the acronym RISDA does not account for that.

  • If the multi-source setup in Sec 3.1.2 is in preliminaries, then can authors cite which previous methods have dealt with this?

  • In equation 4, shouldn't the guidance terms be weighted?

  • For training the O2L operator, why do you use the ERA-5-pretrained decoder? This seems to be ad-hoc.

  • What is MMPS?

  • Why there needs to be mask in O2L of Sec 3.2.1?

  • Set \Omega is unnecessarily defined twice.

  • What is MHS?

  • Italic and non-Italic font is used for R^i.

  • Equations 8 and 11 have very strange super script on the left of s.

  • Major: The paper is very empirical (which is fine). Therefore, there needs to be an empirical support of why one would purse score-based methods and not others. Therefore, comparison with an ML-based supervised method and a non-ML method (e.g., numberical methods) is needed. If compute and time are not available for standard numerical methods, then the authors could at least share their results and compare even if the compute HW is not the same. For non-weather-analysis audiance, this would be valuable.

  • Missing recent literature such as [A,B]

[A] ForcingDAS: Unified and Robust Data Assimilation via Diffusion Forcing, ICML 2026 FoGen Workshop

[B] FlowDAS: A Stochastic Interpolant-based Framework for Data Assimilation, NeurIPS 2025

Questions for authors

See Weaknesses.

6 comments

We sincerely thank the reviewer for the insightful, detailed, and constructive feedback. We are greatly encouraged by the reviewer's recognition of our practical problem formulation ("...deals with this setup is important"), strong empirical performance and efficiency ("...show improvements in terms of both run-time (Table 2) and the problem-specific metrics (Table 1)"), and thorough ablation validation ("Other results demonstrate the impact of every component of RISDA"). We appreciate the critical suggestions regarding manuscript presentation and notation, the training of the O2L operator, baseline comparisons, and missing recent literature. We have carefully checked and revised the manuscript accordingly, and provide a detailed point-by-point response below.

Acronym and Title Issue (Weakness 1):

We have updated the acronym to RISDA-M throughout the title and manuscript to explicitly reflect the multi-source setting.

Preliminaries for Multi-Source Observations (Weakness 2):

Thank you for pointing this out. We have added relevant citations in Section 3.1.2 referencing established multi-source data assimilation.

Gradient Weighting in Equation 4 (Weakness 3):

Equation 4 adds the source-wise likelihood gradients without additional manually chosen scalar weights. Each gradient already accounts for source-specific uncertainty through the Gaussian likelihood in Equation 2, whose covariance incorporates the corresponding observation-error covariance together with the uncertainty in estimating the clean state from the noisy latent. Thus, differences in observation uncertainty are already reflected in the individual guidance terms.

Rationale for the ERA5-Pretrained Decoder in O2L Training (Weakness 3):

We use the frozen ERA5-pretrained atmospheric decoder as a common reference for training the source-specific observation encoders. The atmospheric prior is learned in the latent space of the pretrained atmospheric autoencoder. During O2L training, each observation encoder is optimized through this fixed decoder to reconstruct the corresponding ERA5 atmospheric variables. This reconstruction objective encourages the observation representations to be compatible with the pretrained atmospheric representation, providing implicit alignment through supervision in physical space. Keeping the decoder fixed ensures that the observation encoders adapt to a shared reconstruction mapping, rather than allowing each source to change that mapping during training. This design connects heterogeneous observations to the existing atmospheric representation and provides a common basis for subsequent latent-error calibration and observation guidance.

Definition of MMPS (Weakness 4):

We have added the full name—Moment Matching Posterior Sampling (Rozet et al., 2024; Andry et al., 2025)—upon its first occurrence in the manuscript.

Necessity of Masks in O2L in Section 3.2.1 (Weakness 5):

Polar-orbiting meteorological satellites scan the Earth along narrow orbital swaths, leaving vast geographic regions without valid observations within a short time window. While deep learning encoders are capable of extrapolating or "guessing" values for unobserved regions, such extrapolations carry high uncertainty and potential hallucinations. If left unmasked, these uncertain imputed values could generate spurious likelihood gradients, injecting false guidance signals into the latent space and degrading overall sampling accuracy. Therefore, spatial masks are strictly applied to restrict guidance signals exclusively to genuinely observed locations.

Redundant Definition of Set Omega (Weakness 6):

We have removed the second, redundant definition of \Omega in the revised manuscript.

Definition of MHS (Weakness 7):

We have added full names upon first appearance for all instrument acronyms throughout the text and included a reference (Eyre et al., 2022) for further background.

Font Consistency for R^i (Weakness 8):

We have carefully checked the entire manuscript and standardized the font formatting of R^i for consistency.

Left-Superscript Notation in Equations 8 and 11 (Weakness 9):

We have moved the superscript to the upper right position in Equations 8 and 11 to follow standard mathematical notation.

Extension of Baseline Comparisons (Weakness 10):

We thank the reviewer for this constructive suggestion. Below, we clarify our motivation for pursuing the score-based paradigm over classical numerical and deterministic supervised ML methods, as well as the practical constraints and experimental scope of our study: Traditional variational and ensemble frameworks (e.g., 4D-Var, EnKF) require solving full-physics atmospheric dynamical equations and tangent-linear/adjoint models, incurring prohibitive computational costs and long run-times. Meanwhile, while deterministic supervised ML DA models (e.g., Aardvark Weather, FuXi-Weather) achieve fast inference via direct input-to-target regression, they suffer from architectural rigidity—adding, removing, or modifying observation instruments requires retraining the entire large backbone network. In contrast, RISDA-M enables fast posterior sampling alongside plug-and-play modularity, allowing new observation streams to be integrated via lightweight observation encoders without retraining the prior diffusion backbone (see related work). Operational numerical weather prediction (NWP) systems assimilate orders of magnitude more global observations than the restricted satellite subset used in our controlled study. Therefore, direct accuracy comparisons against operational numerical analysis fields are unfair. Furthermore, classical numerical methods (such as 4D-Var) require high-vertical-resolution atmospheric states to run physical radiative transfer models (e.g., RTTOV) for satellite radiance assimilation, whereas our compact 13-pressure-level representation cannot natively interface with these physical observation operators. Indeed, precisely due to the lack of differentiable physical operators in compact neural state spaces, existing score-based DA literature has largely focused on synthetic or direct in-situ observations (where observation operators are trivial or unnecessary), leaving real-world satellite radiance assimilation unaddressed. Our study represents the exploration of real-world multi-source satellite radiance assimilation within the score-based DA paradigm. Accordingly, our experiments follow the standard protocols of the generative DA community and focus on evaluating the effectiveness of our framework against representative score-based DA baselines under identical real-world multi-source observation settings.

To broaden the comparison within generative assimilation, we implement an adapted LO-SDA baseline by augmenting our existing Appa implementation with a latent optimization stage at every diffusion sampling timestep. The adapted baseline uses the same L2O observation operators and observation inputs as Appa, with all other experimental settings kept unchanged. The results are reported in Table 1.

Table 1. RMSE (↓) comparison including the adapted LO-SDA baseline, using all six observation sources over April–June 2025 and ERA5 as the reference. For Z, R, T, U, and V, values are listed in the order 300/500/850 hPa. T2M, U10, V10, and MSL are surface variables.

MethodZRTUVT2MU10V10MSL
LO-SDA291.26/213.88/162.3821.13/18.65/18.481.59/1.50/1.776.46/4.63/3.276.01/4.29/3.021.252.072.02212.66
Appa290.68/213.45/162.2221.14/18.66/18.491.59/1.50/1.776.46/4.63/3.276.01/4.29/3.021.252.072.02212.64
AGF194.27/163.40/140.7320.61/20.44/14.371.27/0.99/1.293.91/3.05/2.313.79/2.91/2.201.261.361.38184.90
PPGF170.41/145.85/133.7717.67/15.53/13.491.15/0.93/1.243.75/2.94/2.253.63/2.81/2.151.201.351.37178.48

The adapted LO-SDA baseline performs comparably to Appa under this observation setting, while PPGF+VCLG achieves lower RMSE than both methods across all 19 reported variables. This additional comparison supports the effectiveness of RISDA for joint generative assimilation of real-world heterogeneous observations.

Missing Recent Literature (Weakness 11):

We thank the reviewer for pointing out these relevant recent works. We have added citations of both ForcingDAS and FlowDAS in the Introduction of the revised manuscript.

Summary Response

We sincerely thank the reviewer for the detailed and constructive feedback. We are encouraged by the reviewer’s recognition of the importance of addressing real-world multi-source data assimilation, the demonstrated improvements in analysis accuracy and inference efficiency, and the ablation studies showing the contributions of the individual components of our framework.

The reviewer’s main concerns relate to the clarity of the multi-source formulation and observation-guidance design, the scope of baseline comparisons, missing related work, and manuscript presentation. We have clarified the methodological choices underlying our guidance framework, added references for prior multi-source score-based assimilation, and explained our motivation and experimental scope. We have also supplemented the comparison with an adapted LO-SDA baseline, incorporated the suggested recent literature, and revised the acronym, terminology, and mathematical notation. We hope these clarifications, revisions, and additional results address the reviewer’s concerns and improve the clarity and completeness of the manuscript.

Reviewer_Vh3TOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces RISDA (Reliability-Informed Score-based Data Assimilation), a framework designed for the assimilation of real-world multi-source observations. Source-specific observation-to-latent (O2L) encoders map heterogeneous observations into a shared latent space, enabling direct guidance for posterior sampling. Variance-Calibrated Latent Guidance (VCLG) parameterizes a diagonal observation-error covariance in the latent likelihood, using the centered variances of observation-ERA5 latent residuals calibrated for each source and latent channel. Priority-Preserving Gradient Fusion (PPGF) orders source-wise gradient components by estimated reliability at each latent location and channel. Experiments with five microwave instruments (ATMS, AMSU-A, MHS, MWTS, MWHS) and conventional surface observations (CONVI) over April-June 2025 report lower latitude-weighted RMSE against ERA5 in most of the reported variables, reduces RMSE relative to Appa across all 69 channels, and reduces inference time by nearly an order of magnitude relative to Appa.

Critical strengths

S1: The paper is cleanly written and easy to follow, with a well-motivated idea that targets a practically important and challenging setting of assimilating real-world heterogeneous multi-source observations into a global score-based DA framework.

S2: The proposed components are supported by targeted comparisons, and the fusion ablation is well designed to separate the effect of reliability-based ordering from that of the stopping rule.

Critical weaknesses

W1: The reproducibility of this work would be impaired if no code is available. Uploading the anonymous code including any preprocessing and hyperparameter & seed documentation could help ensure reproducibility.

W2: The evaluation is limited to a single season and reports pooled RMSE without specifying random seeds, providing error bars, or conducting significance tests; and hyperparameter sensitivity, which helps inform model robustness, is not evaluated. Some claims also warrant more careful qualification; for example, although RISDA avoids full retraining, it still requires source-specific encoder training (100,000 updates; Appendix C) and calibration, a pipeline that should be more explicitly acknowledged. The manuscript would also benefit from proofreading to address typographical errors and inconsistencies in notation.

Questions for authors

Q1: Could the authors provide how many posterior samples and random seeds were used to compute each reported RMSE, as well as variability estimates across seeds (e.g., standard deviations or confidence intervals)? Such information would help establish the robustness of the smaller reported improvements.

Q2: How sensitive is the proposed method to the calibration data, given that the spatially resolved statistics restrict the calibration errors to each latent position (Appendix E)? For example, how does the advantage of PPGF over AGF change with a shorter calibration window, or with a longer gap between calibration and evaluation than the January-March versus April-June split used here?

Q3: In the fusion ablation, the RMSE of the spatially calibrated AGF+VCLG baseline is not provided. Could the authors report it to help clarify the comparison and align the interpretation of Figures 3 and 8 with Table 1?

Q4: There seems to be typos in the text, for example, in Section 3: "Priority- Preserving" contains an additional space; Section 3.2.2: RiO2LR_i^{O2L} is inconsistent with RO2LiR^i_{O2L} used elsewhere; Section 4.2: the superscript SD in PPGFSDPPGF^{SD} is not defined. Please fix them wherever needed.

3 comments

We appreciate the critical questions and suggestions regarding reproducibility, evaluation robustness, calibration sensitivity, missing ablation baselines, and minor typographical corrections. We provide additional experiments and clarifications below:

Reproducibility and Open Source (Weakness 1):

We have uploaded our implemented code in the paper to an anonymous GitHub repository:https://anonymous.4open.science/r/RISDA-FE63/

Evaluation Rigor and Robustness (Weakness 2 and Question 1):

The main experiments evaluate 363 analysis times, using one generated sample per analysis time with the random seed fixed at 0. Unless otherwise specified, all experiments use seed 0. In addition, we have extended the evaluation to cover all seasons of 2025 over 1300 analysis times. For this extended experiment, observation-error statistics were calibrated using October–December 2024, with no temporal overlap between calibration and evaluation. Given observation availability over the extended period, we use four observation sources (ATMS, AMSU-A, MHS, and CONVI) for all compared methods, as shown in Table 1. Besides, we also evaluate the resulting analyses against independent observations ****that were not assimilated by our system. This provides an observational verification beyond agreement with the ERA5 reference, shown in Table 2.

Table 1. One-year RMSE (↓) comparison among different methods against ERA5. For the pressure-level variables Z, R, T, U, and V, values are listed in the order 300/500/850 hPa. T2M, U10M, V10M, and MSL are surface variables. The lowest RMSE for each variable and pressure level is shown in bold.

MethodZRTUVT2MU10MV10MMSL
Appa437.62/303.98/208.5524.05/21.56/20.301.93/1.89/2.028.14/5.60/3.777.63/5.19/3.501.312.462.41268.07
AGF+VCLG277.29/216.68/167.3824.38/25.35/16.371.59/1.20/1.484.74/3.53/2.574.52/3.35/2.421.371.521.53215.27
PPGF+VCLG229.64/186.65/156.9019.61/18.32/14.751.30/1.06/1.374.27/3.27/2.454.09/3.10/2.331.311.481.49207.04

Table 2. RMSE (↓) comparison against radiosonde observations. For each variable, values are listed in the order 300/500/850 hPa. The lowest RMSE for each variable is shown in bold.

MethodZRTUV
Appa396.36/234.64/106.2825.14/22.92/18.222.05/1.93/1.797.75/5.14/3.267.24/4.80/3.06
AGF+VCLG329.13/240.35/188.7328.53/34.18/18.182.02/1.48/1.775.30/4.03/3.165.08/3.88/3.04
PPGF+VCLG268.43/210.97/182.0122.86/22.49/16.201.62/1.29/1.624.69/3.71/3.104.51/3.52/2.99

As shown in Table 2, PPGF+VCLG achieves the lowest RMSE among the compared methods in 14 of the 15 reported variables, except Z850. Together with the ERA5-based evaluation in Table 1, these results support the effectiveness of the proposed method both in reconstructing the reanalysis reference and in improving agreement with radiosonde observations that were not assimilated by our system.

We also evaluated sensitivity to the sampling steps, using 32, 64, and 128 steps under the same four-source setting over the full year of 2025, with all other settings fixed. The RMSE values remain close across these configurations, suggesting limited sensitivity to the sampling-step count within the tested range.

Sample stepsZRTUVT2MU10V10MSL
32231.95/188.78/157.1819.76/18.45/14.811.31/1.07/1.384.28/3.28/2.464.10/3.10/2.331.321.481.50207.57
64229.64/186.65/156.9019.61/18.32/14.751.30/1.06/1.374.27/3.27/2.454.09/3.10/2.331.311.481.49207.04
128229.49/186.19/158.1119.58/18.31/14.741.31/1.06/1.364.28/3.28/2.464.10/3.11/2.321.301.481.49208.30

To assess robustness, we report 95% confidence intervals under two complementary evaluation settings: (i) evaluation across all analysis times in the April–June 2025 test period with the random seed fixed at 0; and (ii) repeated sampling at eight selected analysis times in May, using ten different random seeds at each time. The results are presented in Figures 5 and 6 in the Appendix, respectively.

Qualification of Method Claims (Weakness 2):

We clarify that avoiding full retraining refers to reusing the pretrained atmospheric autoencoder and diffusion prior. Incorporating a new observation source still requires training its source-specific observation encoder and estimating the error statistics used for guidance calibration and source-priority ordering. In our experiments, each observation encoder is trained for 100,000 updates, as specified in Appendix C, and the calibration procedure is described in Appendix E.

Sensitivity to Calibration Data (Question 2): We have extended our evaluation to cover the full year of 2025, which includes evaluations with longer time gaps from the calibration data. With the calibration statistics estimated from October–December 2024 and kept fixed throughout evaluation, PPGF+VCLG achieves lower RMSE than AGF+VCLG across all 19 reported variable–level combinations in both Q1 and Q4 of 2025. These results support the persistence of its relative benefit over an extended evaluation period.

ZRTUVT2MU10V10MSL
AGF Q1245.89/195.48/161.6023.16/24.00/16.231.67/1.13/1.384.52/3.46/2.474.28/3.23/2.321.361.441.44212.71
AGF Q4292.17/227.85/174.6125.56/26.74/17.141.64/1.26/1.524.91/3.61/2.654.73/3.47/2.491.421.571.58222.87
PPGF Q1204.10/171.08/155.4818.65/17.15/14.581.30/1.01/1.294.10/3.22/2.373.90/3.00/2.241.321.411.41209.29
PPGF Q4229.27/189.69/159.9020.28/19.02/15.301.34/1.09/1.404.37/3.32/2.534.21/3.16/2.381.351.531.54210.52

Clarification of the AGF+VCLG (Question 3):

The table below reports the absolute RMSE of the spatially calibrated AGF+VCLG baseline used in Figure 3 of the manuscript, evaluated with all six observation sources over April–June 2025.

MethodZRTUVT2MU10MV10MMSL
AGF+VCLG307.20/226.19/161.6029.78/34.48/17.771.84/1.34/1.585.07/3.68/2.604.74/3.56/2.431.471.461.48205.55

The spatially calibrated AGF baseline has higher RMSE than the shared-variance AGF baseline in this setting, showing that finer-grained variance calibration does not automatically improve additive fusion. One possible explanation we suppose is that local variance estimates can amplify imperfect source guidance, while spatial pooling may yield more stable weighting. Figure 3 evaluates the fusion strategies using identical spatial variance estimates and source-specific guidance, thereby isolating the effect of fusion. Importantly, PPGF also outperforms the shared-variance AGF baseline in Table 1 of the manuscript. Together, these comparisons support the practical benefit of reliability-informed multi-source fusion within RISDA.

Typographical and Notation Corrections (Question 4):

We have carefully proofread the manuscript, fixed spacing errors (e.g., "Priority-Preserving"), resolved notation inconsistencies, and explicitly defined all previously omitted symbols (e.g., superscript SD).

Summary Response

We sincerely thank the reviewer for the positive, comprehensive, and constructive feedback. We are greatly encouraged by the reviewer's recognition of our practical motivation and clear presentation ("...cleanly written and easy to follow, with a well-motivated idea...") as well as the experimental design ("...supported by targeted comparisons, and the fusion ablation is well designed...").

The reviewer’s main concerns relate to reproducibility, evaluation robustness, calibration sensitivity, the completeness of the fusion comparisons, and the scope of our claims. We have provided an anonymous code repository, extended the robustness evaluations and verification against unassimilated radiosonde observations, examined performance over an extended period with fixed calibration statistics, and supplemented the spatially calibrated AGF baseline results. We have also clarified the training and calibration required for new observation sources and corrected notation and typographical issues. We hope these additions and clarifications address the reviewer’s concerns and strengthen the empirical support for RISDA.

Dear Reviewer Vh3T,

We appreciate your valuable input and the time you have devoted to assessing our work. We are writing to briefly follow up on our responses, as the discussion phase ends on November 18 at 11:59 pm AoE.

If your schedule allows, we would be grateful to hear whether our explanations and additional results address the issues you raised. We remain available for further discussion and would be happy to clarify any remaining questions before the deadline.

Best regards,

Authors of Submission 41346

Reviewer_G4u9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes RISDA, a framework for reconstructing short atmospheric state trajectories conditioned on heterogeneous observations. Source-specific encoders are trained against ERA5 through a frozen atmospheric decoder, mapping satellite and surface observations into a shared atmospheric latent space. Observation–ERA5 latent residuals are then used to calibrate Gaussian guidance uncertainties and establish source priorities. During diffusion sampling, Priority-Preserving Gradient Fusion combines source gradients in reliability order and stops accumulating them when a conflicting contribution is encountered. Experiments evaluate global atmospheric reconstruction using five microwave instrument types and conventional surface observations, comparing RMSE against ERA5 and inference time with an adapted Appa baseline and several ablated variants.

Critical strengths
  1. A practically motivated approach to heterogeneous observation integration. The paper addresses the mismatch between satellite brightness temperatures and atmospheric state variables using source-specific observation encoders. Evaluation with multiple real observation sources at global 0.25° resolution makes the setting substantially more relevant to practical applications than simplified synthetic-observation experiments.

  2. Demonstrated improvements in reconstruction accuracy and inference efficiency. RISDA improves agreement with ERA5 relative to the adapted Appa baseline. Controlled ablations also support the contribution of the proposed fusion strategy. Under the reported setup, inference time decreases from 30.19 to 3.51 minutes per sample on one H800 GPU, an approximately 8.6-fold speedup. This provides concrete evidence of a useful accuracy–runtime improvement within the evaluated framework.

Critical weaknesses
  1. The probabilistic interpretation of the generated samples remains insufficiently established. Observation encoders are trained to reconstruct ERA5, but this does not by itself establish that their outputs follow the assumed conditionally unbiased Gaussian observation model. Marginal residual diagnostics provide partial support but do not establish the required conditional error properties. Furthermore, PPGF selectively discards conflicting guidance components, and the resulting vector field is not shown to correspond to a coherent posterior score. The RMSE experiments therefore support reconstruction accuracy, but do not establish posterior fidelity or uncertainty calibration.

  2. The evaluation does not establish the broader assimilation advantage. ERA5 is used for prior learning, observation-encoder supervision, reliability calibration, and evaluation. Although the periods are temporally separated, improved agreement with ERA5 does not establish improved accuracy relative to the true atmospheric state or an ability to correct ERA5 errors. The approach also depends on paired observations and a comprehensive reanalysis reference for training and calibration. Comparisons are primarily against adapted Appa and internal variants; widely used assimilation methods such as LETKF are absent. Consequently, the advantage over established assimilation approaches remains unclear.

Questions for authors
  1. Posterior interpretation and validation. What distribution are the final RISDA samples intended to approximate after PPGF modifies the guidance? Have you evaluated posterior accuracy or uncertainty calibration, for example through coverage diagnostics or a controlled problem with a reference posterior? If posterior fidelity is not an intended claim, please clarify the scope of the posterior-sampling interpretation.

  2. Validation beyond ERA5 agreement. Do you have results against independent observations excluded from the method’s inputs and supervision, comparing both RISDA and ERA5? Such evidence would help distinguish improved ERA5 reconstruction from improved atmospheric state estimation.

  3. Established assimilation baselines. Is a comparison with LETKF or another established assimilation method available under a reasonably comparable observation setting? If not, please clarify the obstacles to such a comparison and the scope of the claimed assimilation advantage.

6 comments

We appreciate the critical questions regarding the probabilistic interpretation of PPGF, the reference truth for evaluation, and comparisons with established assimilation baselines. We have performed additional clarifications and evaluations to address each point:

Posterior Sampling and Probabilistic Interpretation (Weakness 1 and Question 1):

RISDA addresses the practical challenge of joint generative assimilation of real-world heterogeneous multi-source observations. Our contributions and conclusions are grounded in empirical evidence on atmospheric analysis accuracy, inference efficiency, and controlled comparisons of guidance-fusion strategies. Within this framework, the Gaussian latent-error model is an empirically motivated working approximation for constructing guidance, while PPGF provides an empirically motivated, reliability-informed rule for combining source-specific guidance. Its source-wise guidance follows the prior-plus-likelihood formulation of score-based data assimilation, with latent likelihood approximations motivated by empirical residual diagnostics. As discussed in Section 3.4, learned observation encoders and calibrated error statistics introduce approximation errors that can vary across sources, locations, and latent channels. PPGF addresses this practical issue through reliability-informed fusion, preserving higher-priority guidance when source gradients disagree. Our claims about PPGF are mostly based on fusion ablations and atmospheric analysis results, which evaluate the practical effectiveness of this design. We have complemented the scope of PPGF in the main text after Eq.11 .

Independent Observation Validation (Weakness 2 and Question 2):

To address your concern, we have extended the evaluation to cover all seasons of 2025. For this extended experiment, observation-error statistics were calibrated using October–December 2024, with no temporal overlap between calibration and evaluation. Given observation availability over the extended period, we use four observation sources (ATMS, AMSU-A, MHS, and CONVI) for all compared methods, as shown in Table 1. Besides, we also evaluate the resulting analyses against independent observations that were not assimilated by our system. This provides an observational verification beyond agreement with the ERA5 reference, shown in Table 2.

Table 1. One-year RMSE (↓) comparison among different methods against ERA5. For the pressure-level variables Z, R, T, U, and V, values are listed in the order 300/500/850 hPa. T2M, U10M, V10M, and MSL are surface variables. The lowest RMSE for each variable and pressure level is shown in bold.

MethodZRTUVT2MU10MV10MMSL
Appa437.62/303.98/208.5524.05/21.56/20.301.93/1.89/2.028.14/5.60/3.777.63/5.19/3.501.312.462.41268.07
AGF+VCLG277.29/216.68/167.3824.38/25.35/16.371.59/1.20/1.484.74/3.53/2.574.52/3.35/2.421.371.521.53215.27
PPGF+VCLG229.64/186.65/156.9019.61/18.32/14.751.30/1.06/1.374.27/3.27/2.454.09/3.10/2.331.311.481.49207.04

Table 2. RMSE (↓) comparison against radiosonde observations. For each variable, values are listed in the order 300/500/850 hPa. The lowest RMSE for each variable is shown in bold.

MethodZRTUV
Appa396.36/234.64/106.2825.14/22.92/18.222.05/1.93/1.797.75/5.14/3.267.24/4.80/3.06
AGF+VCLG329.13/240.35/188.7328.53/34.18/18.182.02/1.48/1.775.30/4.03/3.165.08/3.88/3.04
PPGF+VCLG268.43/210.97/182.0122.86/22.49/16.201.62/1.29/1.624.69/3.71/3.104.51/3.52/2.99

As shown in Table 2, PPGF+VCLG achieves the lowest RMSE among the compared methods in 14 of the 15 reported variables, except Z850. Together with the ERA5-based evaluation in Table 1, these results support the effectiveness of the proposed method both in reconstructing the reanalysis reference and in improving agreement with radiosonde observations that were not assimilated by our system.

We would like to explicitly clarify that we do not present direct accuracy comparisons between RISDA and ERA5 in the manuscript. ERA5 serves as the training reference and evaluation benchmark in widespread ML-based method[1,2,3]. ERA5 is a top-tier global reanalysis dataset that integrates a substantially broader range of global observations (including polar-orbiting and geostationary satellites, radiosondes, ships, radar, and surface stations), whereas RISDA currently assimilates only a restricted subset of heterogeneous satellite and surface observations, with far fewer observations than ERA5. The additional radiosonde evaluation examines whether the improvements over the compared methods also extend to observations that were not assimilated by RISDA. This demonstrates RISDA's effectiveness and practical value in handling real heterogeneous observations and improving atmospheric state estimation.

[1]Price, Ilan, et al. "Probabilistic weather forecasting with machine learning." Nature 637.8044 (2025): 84-90.

[2]Bi, Kaifeng, et al. "Accurate medium-range global weather forecasting with 3D neural networks." Nature 619.7970 (2023): 533-538.

[3]Sun J A, Fan H, Gong J, et al. Align-DA: Align Score-based Atmospheric Data Assimilation with Multiple Preferences[J]. Advances in Neural Information Processing Systems, 2026, 38: 63539-63568.

Incorporation of Baseline Comparisons (Weakness 2 and Question 3):

Regarding the reviewer's suggestion of traditional data assimilation baselines such as LETKF, we wish to clarify why direct comparison is unfeasible under this experimental setup: a matched LETKF comparison would require a suitable background ensemble and complex physical observation operators (such as RTTOV) for our heterogeneous satellite observations. Our current implementation uses a 13-pressure-level atmospheric representation and does not include a validated radiative-transfer-based LETKF pipeline. Establishing such a comparison would require additional adaptation and validation of the observation interface and ensemble assimilation configuration, which we have not completed in the present revision.

To broaden the comparison within generative assimilation, we implement an adapted LO-SDA baseline by augmenting our existing Appa implementation with a latent optimization stage at every diffusion sampling timestep. The adapted baseline uses the same L2O observation operators and observation inputs as Appa, with all other experimental settings kept unchanged. The results are reported in Table 3.

Table 3. RMSE (↓) comparison including the adapted LO-SDA baseline, using all six observation sources over April–June 2025 and ERA5 as the reference. For Z, R, T, U, and V, values are listed in the order 300/500/850 hPa. T2M, U10, V10, and MSL are surface variables.

MethodZRTUVT2MU10V10MSL
LO-SDA291.26/213.88/162.3821.13/18.65/18.481.59/1.50/1.776.46/4.63/3.276.01/4.29/3.021.252.072.02212.66
Appa290.68/213.45/162.2221.14/18.66/18.491.59/1.50/1.776.46/4.63/3.276.01/4.29/3.021.252.072.02212.64
AGF194.27/163.40/140.7320.61/20.44/14.371.27/0.99/1.293.91/3.05/2.313.79/2.91/2.201.261.361.38184.90
PPGF170.41/145.85/133.7717.67/15.53/13.491.15/0.93/1.243.75/2.94/2.253.63/2.81/2.151.201.351.37178.48

The adapted LO-SDA baseline performs comparably to Appa under this observation setting, while PPGF+VCLG achieves lower RMSE than both methods across all 19 reported variables. This additional comparison supports the effectiveness of RISDA for joint generative assimilation of real-world heterogeneous observations.

Summary Response

We sincerely thank the reviewer for the comprehensive, insightful, and constructive feedback. We are greatly encouraged by the reviewer's positive recognition of our practical motivation ("addresses the mismatch between satellite brightness temperatures and atmospheric state variables... evaluation with multiple real observation sources makes the setting substantially more relevant to practical applications"), demonstrated accuracy and efficiency gains ("improves agreement with ERA5... inference time decreases from 30.19 to 3.51 minutes per sample, an approximately 8.6-fold speedup"), and clarity of presentation ("clear and easy to follow").

The review’s main concerns relate to the probabilistic interpretation of PPGF, validation beyond ERA5 agreement, and comparisons with established assimilation approaches. We have clarified that RISDA is a practical, empirically motivated framework for joint generative assimilation of real-world heterogeneous multi-source observations, with its contributions supported by demonstrated analysis accuracy and inference efficiency. We have also expanded the evaluation, added verification against radiosonde observations not assimilated by RISDA and an adapted LO-SDA baseline, and clarified the scope of the comparisons. We hope these clarifications and additional results address the main concerns and support a reassessment of our work.

Reviewer_Z3YGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RISDA, a Reliability-Informed Score-based Data Assimilation framework for heterogeneous multi-source atmospheric observations. The proposed method employs observation-to-latent (O2L) encoders to map satellite and surface observations from different sources into a unified atmospheric latent space, addressing the differences in variable dimensions and spatial distributions among heterogeneous observations. Based on this shared latent representation, the authors introduce Variance-Calibrated Latent Guidance (VCLG), which estimates observation-error variances using the statistical characteristics of latent-space residuals. Furthermore, Priority-Preserving Gradient Fusion (PPGF) is proposed to determine the fusion priority of source-specific gradients based on their residual MSE, enabling the integration of multiple observation sources during diffusion sampling. Experiments are conducted using five types of microwave satellite observations and one type of conventional surface observation, covering 69 atmospheric variables in global atmospheric data assimilation. The proposed framework is evaluated in terms of both assimilation accuracy and inference efficiency.

Critical strengths
  1. This paper proposes the RISDA framework to address the differences in spatial distribution, variable dimensions, and observation reliability among heterogeneous multi-source atmospheric observations. Three modules, namely O2L, VCLG, and PPGF, are designed to achieve observation guidance and fusion from different sources in a unified latent space.
  2. The authors propose a combination of the O2L encoder and VCLG, which calibrates observation guidance using channel-wise error variances in the latent space. Compared with the conventional L2O pathway, this approach avoids repeatedly invoking the observation operator during sampling and improves inference efficiency.
  3. The experiments cover five types of microwave satellite observations and one type of conventional ground-based observation, with evaluations conducted on 69 atmospheric variables. The proposed method achieves certain performance improvements and demonstrates research value.
Critical weaknesses
  1. The authors conduct model training, error calibration, and final performance evaluation exclusively on ERA5 data, which is insufficient to demonstrate that the proposed method improves the accuracy of estimating the true atmospheric state. Furthermore, the evaluation is limited to the period from April to June 2025, without sufficient cross-seasonal or cross-year generalization experiments.
  2. The authors only select Appa as the external baseline, lacking comparisons with other recent mainstream deep-learning data assimilation approaches, including end-to-end satellite data assimilation methods (e.g., FuXi-DA and FuXiWeather2) and Latent Ensemble Score Filter / Latent Optimization-based methods (e.g., Latent-EnSF, LO-SDA, LD-EnSF, and Align-DA).
  3. The comparison of different gradient fusion strategies in Table 1 is not entirely fair. AGF+VCLG and PCGrad+VCLG use spatially shared channel-wise variances (v_c^i), whereas PPGF+VCLG uses spatially varying variances (v_{p,c}^i). Therefore, the performance improvement may result from both spatial variance calibration and changes in the gradient fusion strategy, making it difficult to isolate the contribution of PPGF itself.
  4. The authors provide only module-level (e.g., VCLG and PPGF) and pathway-level (O2L vs. L2O) ablation studies, without more fine-grained ablations or sensitivity analyses. For example, the effects of important settings, such as the latent channel dimension (\tilde{C}=128), the number of sampling steps in PCSample, and the gradient truncation threshold in Equation (10) (currently fixed at (<0)), have not been independently investigated under different diffusion noise levels (t).
  5. The authors approximate the observation-error covariance using a diagonal covariance matrix based on the approximately Gaussian characteristics of latent-space residuals and their relatively weak inter-channel correlations. However, the statistical validation remains insufficient. Figure 2 and Appendix D only present error distributions and Q–Q plots for a small number of latent channels, which is insufficient to establish that all 128 channels approximately follow Gaussian distributions.
Questions for authors
  1. This paper uses Appa as the only external baseline. Could the authors explain why no comparisons were conducted with other recent atmospheric data assimilation methods, such as FuXi-DA, FuXiWeather2, LO-SDA, LD-EnSF, and Align-DA?
  2. The ablation studies mainly focus on comparisons between different modules and observation pathways. Could the authors further investigate the sensitivity of key parameters, including the latent channel dimension (128), the number of PCSample sampling steps (64), and the gradient truncation threshold in Equation (10), to verify the stability of the proposed method under different parameter settings?
  3. The model training, error calibration, and final performance evaluation in this study all rely on ERA5 data. Could the authors further evaluate the generalization capability of the proposed method in real-world atmospheric state estimation and under different meteorological conditions?
4 comments

Thank you for your detailed and constructive feedback. We have expanded the evaluation period, added verification against unassimilated radiosonde observations, supplemented an additional baseline, investigated sampling-step sensitivity, and extended the latent-error diagnostics. Below, we address your concerns point by point.

Extension of Evaluation Period and Reference Truth Validation (Weakness 1 and Question 3):

We have extended the evaluation to cover all seasons of 2025. For this extended experiment, observation-error statistics were calibrated using October–December 2024, with no temporal overlap between calibration and evaluation. Given observation availability over the extended period, we use four observation sources (ATMS, AMSU-A, MHS, and CONVI) for all compared methods, as shown in Table 1. Besides, we also evaluate the resulting analyses against radiosonde observations that were not assimilated by our system. This provides an observational verification beyond agreement with the ERA5 reference, shown in Table 2.

Table 1. One-year RMSE (↓) comparison among different methods against ERA5. For the pressure-level variables Z, R, T, U, and V, values are listed in the order 300/500/850 hPa. T2M, U10M, V10M, and MSL are surface variables. The lowest RMSE for each variable and pressure level is shown in bold.

MethodZRTUVT2MU10MV10MMSL
Appa437.62/303.98/208.5524.05/21.56/20.301.93/1.89/2.028.14/5.60/3.777.63/5.19/3.501.312.462.41268.07
AGF+VCLG277.29/216.68/167.3824.38/25.35/16.371.59/1.20/1.484.74/3.53/2.574.52/3.35/2.421.371.521.53215.27
PPGF+VCLG229.64/186.65/156.9019.61/18.32/14.751.30/1.06/1.374.27/3.27/2.454.09/3.10/2.331.311.481.49207.04

Table 2. RMSE (↓) comparison against radiosonde observations. For each variable, values are listed in the order 300/500/850 hPa. The lowest RMSE for each variable is shown in bold.

MethodZRTUV
Appa396.36/234.64/106.2825.14/22.92/18.222.05/1.93/1.797.75/5.14/3.267.24/4.80/3.06
AGF+VCLG329.13/240.35/188.7328.53/34.18/18.182.02/1.48/1.775.30/4.03/3.165.08/3.88/3.04
PPGF+VCLG268.43/210.97/182.0122.86/22.49/16.201.62/1.29/1.624.69/3.71/3.104.51/3.52/2.99

In this year-long evaluation, PPGF+VCLG matches or improves upon the compared baselines across all 19 reported variables, supporting the overall accuracy advantage observed in the main experiments. In the radiosonde evaluation in Table 2, it achieves the lowest RMSE on 14 of the 15 reported combinations. Z850 is the exception. These results support the effectiveness of the proposed approach over a full annual cycle and against a third-party observations withheld from our assimilation system.

Scope of the baseline comparisons (Weakness 2 and Question 1):

Our original manuscript did not include direct comparisons with all the suggested methods for two main reasons.

First, end-to-end deterministic approaches such as FuXi-DA and FuXiWeather2 have been discussed in Section 2.1. These methods learn supervised mappings from observations and background information to deterministic atmospheric analysis fields. Our work focuses on a different research direction: joint generative assimilation of real-world heterogeneous multi-source observations, where observations guide sampling from a learned atmospheric prior. We therefore prioritize generative assimilation baselines that directly inform this research question. Moreover, adapting the end-to-end systems to our particular observation sources, input configuration, and evaluation protocol would require substantial implementation and training efforts as disscussed in Section 2.1. Establishing such matched comparisons is beyond what we can reliably complete within the rebuttal period.

Second, existing latent generative assimilation methods, including Latent-EnSF, LO-SDA, LD-EnSF, and Align-DA, were developed and evaluated under observation settings that differ from our heterogeneous multi-instrument setting, often incorporating simulated observations or observations directly expressed in atmospheric state variables. Applying these methods to real satellite brightness temperatures requires additional observation operator. Bridging this gap, while enabling reliable joint guidance from multiple observation sources, is exactly a central motivation for RISDA.

Our Appa baseline already incorporates this adaptation: we design and train source-specific latent-to-observation (L2O) operators to enable observation-space guidance from real satellite observations. To broaden the comparison, we implement an adapted LO-SDA baseline by augmenting our existing Appa implementation with a latent optimization stage at every diffusion sampling timestep, following the latent optimization strategy of LO-SDA. The adapted baseline uses the same L2O observation operators and observation inputs as Appa, with all other experimental settings kept unchanged. This allows us to evaluate an additional generative assimilation strategy under the same real-world multi-source observation setting. The results are reported in Table 3.

Table 3. RMSE (↓) comparison including the adapted LO-SDA baseline, using all six observation sources over April–June 2025 and ERA5 as the reference. For Z, R, T, U, and V, values are listed in the order 300/500/850 hPa. T2M, U10, V10, and MSL are surface variables.

MethodZRTUVT2MU10V10MSL
LO-SDA291.26/213.88/162.3821.13/18.65/18.481.59/1.50/1.776.46/4.63/3.276.01/4.29/3.021.252.072.02212.66
Appa290.68/213.45/162.2221.14/18.66/18.491.59/1.50/1.776.46/4.63/3.276.01/4.29/3.021.252.072.02212.64
AGF194.27/163.40/140.7320.61/20.44/14.371.27/0.99/1.293.91/3.05/2.313.79/2.91/2.201.261.361.38184.90
PPGF170.41/145.85/133.7717.67/15.53/13.491.15/0.93/1.243.75/2.94/2.253.63/2.81/2.151.201.351.37178.48

The adapted LO-SDA baseline performs comparably to Appa under this observation setting, while PPGF+VCLG achieves lower RMSE than both methods across all 19 reported variables. This additional comparison supports the effectiveness of RISDA for joint generative assimilation of real-world heterogeneous observations.

Disentangling PPGF from spatial variance calibration (Weakness 3):

We apologize for the inaccurate description of the variance-calibration settings in the original manuscript. PCGrad+VCLG and both PPGF variants reported in Table 1 of the manuscript use spatially resolved variances. The original description of PCGrad+VCLG as using spatially shared variances was incorrect and has been corrected. We have corrected this description in the manuscript to reflect the actual experimental settings. To isolate the contribution of gradient fusion, Figure 3 provides a controlled comparison between PPGF+VCLG and a spatially calibrated AGF+VCLG baseline. Both methods use the same spatially resolved variance estimates and source-specific guidance, differing only in their fusion strategy. This comparison demonstrates the benefit of PPGF relative to additive fusion while holding variance fixed.

Sensitivity to the number of sampling steps (Weakness 4 and Question 2):

We have added a sampling step sensitivity experiment using the same four observation sources over the full year of 2025. We evaluate 32, 64, and 128 sampling steps while keeping the remaining settings fixed, including the random seed at 0.

Table 4. One-year RMSE (↓) comparison among different sample steps against ERA5. For the pressure-level variables Z, R, T, U, and V, values are listed in the order 300/500/850 hPa. T2M, U10M, V10M, and MSL are surface variables.

Sample stepsZRTUVT2MU10V10MSL
32231.95/188.78/157.1819.76/18.45/14.811.31/1.07/1.384.28/3.28/2.464.10/3.10/2.331.321.481.50207.57
64229.64/186.65/156.9019.61/18.32/14.751.30/1.06/1.374.27/3.27/2.454.09/3.10/2.331.311.481.49207.04
128229.49/186.19/158.1119.58/18.31/14.741.31/1.06/1.364.28/3.28/2.464.10/3.11/2.321.301.481.49208.30

As shown in Table 4, RMSE values remain close across 32, 64, and 128 sampling steps, suggesting limited sensitivity to the sampling-step count within the tested range. All compared methods use the same fixed 128-channel latent representation. Changing this dimensionality would require retraining the atmospheric autoencoder, diffusion prior, and observation encoders, so a dimensionality sweep is beyond the scope of the current revision. The zero threshold in Eq. (10) identifies opposing gradient directions and is applied to the current gradients at each diffusion timestep.

Expanded diagnostics of latent observation errors (Weakness 5):

To address the concern, we have added latent error histograms with Gaussian fits and Gaussian Q–Q plots for all 128 MHS latent channels in Appendix D (Figures 12 and 13). The results provide empirical support for using Gaussian marginal approximations and a diagonal covariance as a practical likelihood model. We consequently describe the diagonal Gaussian likelihood as an empirically motivated approximation.

Summary Response

We sincerely thank the reviewer for the constructive, detailed, and insightful comments, which help us substantially improve our manuscript. We are encouraged that the reviewer acknowledges the novelty and technical design of RISDA ("addresses differences in spatial distribution, variable dimensions, and observation reliability among heterogeneous observations"), its high inference efficiency ("avoids repeatedly invoking the observation operator during sampling and improves inference efficiency"), and the effectiveness of our method ("achieves performance improvements and demonstrates research value").

The review’s main concerns centered on the evaluation period and reliance on ERA5, the coverage and fairness of baseline comparisons, hyperparameter sensitivity, and the statistical support for latent-error modeling. In response, we have extended the evaluation to a full annual cycle, added verification against unassimilated radiosonde observations and an adapted LO-SDA baseline, clarified the variance settings and controlled fusion comparisons, investigated sampling-step sensitivity, and expanded the error diagnostics to all 128 MHS latent channels. We believe these additional experiments and clarifications substantially address your main concerns and strengthen the evidence for RISDA’s effectiveness in joint generative assimilation of real-world heterogeneous multi-source observations. We would appreciate your reconsideration of our work in light of these additional results and clarifications.

Dear Reviewer Z3YG,

Thank you again for your valuable and constructive suggestions, which have helped us strengthen our work.

Following your feedback, we have added a full-year evaluation covering all seasons of 2025, validation against radiosonde observations withheld from our assimilation system, an additional comparison with an adapted LO-SDA baseline, and a sampling-step sensitivity study using 32, 64, and 128 steps.

Most of your concerns overlap with those raised by other reviewers Vh3Tand fZci. We are encouraged that they both have acknowledged that our responses and additional evidence address their main concerns and have subsequently raised their scores. We hope these additions also help address the issues you identified.

As the discussion period draws to a close, we would greatly appreciate it if you could review our responses and additional results when your schedule permits. We would welcome your reassessment and any remaining questions, and would be glad to provide further clarification.

Thank you for your time and consideration. We look forward to hearing your thoughts.

Best regards,
Authors of Submission 41346

Comments on the paper

2 comments

Dear authors and reviewers,

The author-reviewer discussion window will close tomorrow on November 18.

@authors: If not done already, please respond to the reviewers' comments before the discussion window closes. I would advise short and concise responses over lengthy explanations.

@reviewers: Please check the authors' responses and provide any additional comments before the discussion window closes. Consider updating your initial review if necessary, and in particular your score if your opinion has changed. You may increase your score if you feel the authors' responses warrant it, maintain your score if your opinion remains the same, or even decrease your score if you feel the other reviews or the authors' responses reveal new shortcomings. At the very least, please acknowledge the authors' responses.

Thank you for your attention and cooperation.

The AC

AuthorsOpenReview ↗

Rebuttal Summary and Revisions

Dear Area Chairs, Senior Area Chairs, and Reviewers,

We sincerely thank all four reviewers for their thoughtful evaluations and constructive feedback. We appreciate their recognition of the practical importance of assimilating real-world heterogeneous atmospheric observations, the shared latent representation, and the demonstrated improvements in analysis accuracy and inference efficiency. Their suggestions have helped us strengthen the experimental evidence, clarify the methodological choices, and improve the manuscript.

Rebuttal Highlights.

  1. Full-Year Evaluation and Independent Observational Validation. We extended the evaluation to 1,300+ analysis times across all seasons of 2025 using four observation sources, with calibration statistics estimated from October–December 2024 and held fixed throughout evaluation. PPGF+VCLG matches or improves upon the compared baselines across all 19 reported variables. We also added validation against radiosonde observations withheld from our assimilation inputs, achieving the lowest RMSE on 14 of 15 reported variables.

  2. Expanded Baseline Comparison. We added an adapted LO-SDA baseline under the same six-source observation setting as the main experiments. The adaptation augments our Appa implementation with latent optimization during sampling while retaining the same L2O observation operators and observation inputs. RISDA-M achieves lower RMSE across all 19 reported variables. In our responses, we also discussed the relationship to supervised and conventional assimilation methods and clarified the scope of the comparison.

  3. Sampling Sensitivity and Evaluation Variability. We conducted full-year experiments with 32, 64, and 128 sampling steps, keeping other settings fixed, including seed 0. The closely matched RMSE values suggest limited sensitivity within the tested range. We also reported 95% confidence intervals under two complementary settings: evaluation across April–June analysis times with seed 0, and repeated sampling with ten random seeds at each of eight selected analysis times in May.

  4. Latent-Error Diagnostics and Probabilistic Interpretation. We expanded the diagnostics to include error histograms with Gaussian fits and Q–Q plots for all 128 MHS latent channels. We clarified the Gaussian likelihood as an empirically motivated working approximation and PPGF as a practical, reliability-informed guidance-fusion strategy. Our empirical conclusions concern atmospheric analysis accuracy and inference efficiency.

  5. Methodological Clarity and Reproducibility. We corrected the description of the variance settings to reflect that PCGrad and both PPGF variants use spatially resolved variances, and clarified the controlled fusion comparisons. Our responses further explained the frozen decoder, observation masking, and source-specific uncertainty. We also clarified that incorporating a new source requires encoder training and calibration while reusing the atmospheric autoencoder and diffusion prior, and provided an anonymous code repository during the discussion.

Summary of Manuscript Revisions.

  • Main Text: Added the adapted LO-SDA comparison (Secs. 4.1–4.2, Table 1), full-year ERA5 evaluation and independent radiosonde validation (Sec. 4.2, Tables 2–3), and clarified the practical interpretation of PPGF and the evaluation conclusions (Secs. 3.4 and 5).
  • Appendix A.1: Added the full-year sampling-step sensitivity study using 32, 64, and 128 steps (Table 6).
  • Appendix A.2: Added evaluations across analysis times and sampling seeds with 95% confidence intervals (Figures 5–6).
  • Appendix D: Added histograms with Gaussian fits and Q–Q plots for all 128 MHS latent channels (Figures 12–13).
  • Variance Settings: Corrected and clarified the shared versus spatially resolved variance settings in Sec. 4.2 and Appendix E.
  • Presentation and Related Work: Updated the acronym to RISDA-M, defined MMPS and instrument acronyms, corrected notation and formatting, and incorporated the suggested ForcingDAS and FlowDAS references.

Overall, these discussions and revisions have strengthened the empirical support for RISDA-M and clarified the scope and interpretation of its contributions. We hope that our responses and revised manuscript provide a more comprehensive basis for evaluating this work.

We sincerely appreciate your time, constructive feedback, and careful consideration.

Best regards,
Authors of Submission 41346

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.