acceptodds
Under review as a conference paper at ICLR 2027

Xaurora: Generative Weather Forecasting with Denoising Stochastic Interpolants from a Foundation Model Prior

Abstract

Deep learning has revolutionised weather forecasting in recent years, especially through atmospheric foundation models, which offer competitive skill for a fraction of the computational costs of classic physics-based models. However, most existing foundation models are deterministic, limiting the generation of large ensembles for accurate uncertainty quantification, extreme weather risk assessment, and long-range weather forecasting. Furthermore, these models incur a large, often prohibitive, computational overhead to train from scratch. To address these shortcomings, we turn a pretrained deterministic prior model, namely the Aurora foundation model, into a generative ensemble-prediction model. To that end, we introduce a novel generative method, _Denoising Stochastic Interpolants_, combined with a _replay buffer_ for Stochastic Differential Equation (SDE) rollout, enabling probabilistic training of SDE trajectories. Our stochastic foundation model, _Xaurora_, is finetuned from the small Aurora version, yet it approaches the state-of-the-art on global ensemble metrics and is competitive with the large version of Aurora. Our method is parameter and sample efficient, and generates skilful 15-day forecasts in 13 minutes. Our results demonstrate that deterministic foundation models can be efficiently extended into even stronger stochastic models.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Weather Forecasting, Foundation Models, Probabilistic Forecasting, Generative Modelling, Stochastic Interpolants

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6EHXOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

Xaurora converts the deterministic Aurora-Small model into a generative weather forecaster using denoising stochastic interpolants. The network defines an SDE drift conditioned on historical states, intermediate sampling states, and pseudo-time, trained using a replay buffer on fair-CRPS loss (requiring 42 hours on 16 A100s). Generating one ensemble member's 15-day forecast at 6-hour intervals takes approximately 13 minutes on a single A100 GPU. On 56 test initializations from 2022 at 0.25° resolution, Xaurora yields lower ensemble-mean errors than both Aurora-Small and Aurora-Large across most evaluated variables and lead times. The spectral analysis also identifies a persistent noise floor and insufficient ensemble spread at large spatial scales.

Critical strengths
  1. A coherent and empirically useful extension of a pretrained forecaster. The endpoint-prediction formulation provides a natural way to initialize from Aurora, while the extension of replay-buffer training keeps the training cost under control. The reported improvements over the deterministic prior and the optimization ablations support the practical value of the complete approach. The evaluation provides evidence that the proposed adaptation can support useful ensemble forecasts.

  2. Informative analysis of the method's limitations. The spectral diagnostics show that large-scale underdispersion can coexist with excess small-scale variability, which aggregate spread-skill ratios can conceal. The noise-scaling and spatial-correlation interventions further investigate this behavior and report trade-offs rather than claiming that the calibration problem is solved. This analysis adds scientific value beyond reporting aggregate forecast scores.

Critical weaknesses
  1. The unit gaussian noise use in paper does not match up with weather forcasting task. Per-variable normalization addresses numerical scales but does not establish that the same normalized noise amplitude is appropriate relative to each variable's short-range conditional uncertainty. Overall variability can differ substantially from uncertainty about the next six hours. The independent grid-based noise also does not directly encode spatial or cross-variable relationships. The reported noise floor and scale-dependent miscalibration show that the resulting uncertainty is not uniformly satisfactory. Appendix B.1 also shows that noise interventions have variable-dependent effects and do not resolve the problem across lead times.

  2. The parameterization conclusion exceeds what Figure 2 establishes. Figure 2 inserts the same endpoint predictor directly into endpoint, drift, and noise output roles. Its result supports the convenience of direct endpoint initialization, but does not exclude drift or noise predictors initialized using the corresponding algebraic transformations. The claim that only the proposed parameterization can preserve the prior should therefore be restricted to the tested initialization. This distinction can be addressed by more precise claims without negating the empirical contribution.

Questions for authors
  1. Calibration interpretation: Based on the existing spectral diagnostics and noise interventions, which variables, spatial scales, and lead times retain the most consequential miscalibration? Please clarify the scope of the calibration claims and how the reported aggregate scores should be interpreted under these limitations.

  2. Parameterization claim: Is Figure 2 intended only to compare direct reuse of an endpoint predictor without output transformations? Please distinguish this claim from an advantage over alternatives initialized to preserve the same dynamics through the transformations in Equation (6) and Appendix A.5.

2 comments
AuthorsOpenReview ↗

Official response [1/2]

Dear reviewer,

Thank you for your comments. We are glad to hear that you found the spectral diagnostic valuable and share your belief that the approach supports useful ensemble forecasts. Your comments on the noise structure brought forward new perspectives for future work. A new version of the manuscript was uploaded, changes relative to the submission are highlighted in blue.

The unit gaussian noise used in paper does not match up with weather forecasting task.

We agree with your point of view, and this is the motivation for the experiments in Appendix B.8. Theoretically, pure white noise should be sufficient to generate any distribution, including atmospheric states. However, with finite integration steps, imposing different noise structures could make the denoising more efficient by homogenising the signal to noise ratio across spatial scales. Unfortunately, the research on the topic is very limited. Hence, we treat the idea as promising future work.

Nevertheless, we appreciate your idea about short-term conditional uncertainty. We ran an experiment where the white noise is scaled by ERA5 climatological residual uncertainty (STD[xt+Δt−xt]\mathrm{STD}[\mathbf{x}^{t+\Delta t} - \mathbf{x}^t]) corresponding to the forecast time (i.e. we compute one variance map per 6h per calendar day). The results were generally negative: Z500 got closer to SSR=1 but had a slight decrease in skill, Q850 became more underdispersive. In general, the large-scale underdispersion was not addressed. Our hypothesis is that this intervention only scales the variance but leaves the spectrum essentially white, which cannot fix the spectral calibration. The current B.8 experiments highlight the importance of spectral properties of the noise, more specifically increasing large scale power and decreasing small scale power (i.e. generally “redder” noise).

A simpler approach to capture short-term uncertainty would be to use ERA5 EDA uncertainty, which quantifies it exactly. We mention this as future work in the paper. Adding a perturbation with variance given by this uncertainty to the initial condition would likely bootstrap large-scale spread and improve the ensemble dispersion on short lead times and beyond.

Calibration interpretation: Based on the existing spectral diagnostics and noise interventions, which variables, spatial scales, and lead times retain the most consequential miscalibration? Please clarify the scope of the calibration claims and how the reported aggregate scores should be interpreted under these limitations.

The manuscript revision of November 11 updates the sampler and subtracts the residual noise at inference. This fixes the noise floor and leaves the variables homogeneously underdispersive. It also includes a spectral postprocessing section that fixes the ensemble calibration across spatial scales.

Figure 15 in Appendix B.1 shows the aggregate SSR for different variables. The lead time evolution is consistent: the ensemble starts underdispersed and gets closer to calibration over the course of a 15-day rollout. The increase is not monotonic; some variables display a small drop in SSR on short lead times. In general, the specific humidity variable (Q) has the best SSR.

The spectral post-processing functions G(ℓ)\mathcal{G}(\ell) and H(s)\mathcal{H}(s) (Figure 29, Appendix B.5) show that the most consequential underdispersion happens in short-term large-scale structures. This is also confirmed by Figure 30 in Appendix B.5 that shows the spectral SSR at different lead times before and after post-processing. Short-term large-scale spread could be increased by including initial condition uncertainty (from ERA5 EDA as in GenCast).

AuthorsOpenReview ↗

Official response [2/2]

The parameterization conclusion exceeds what Figure 2 establishes. Figure 2 inserts the same endpoint predictor directly into endpoint, drift, and noise output roles. Its result supports the convenience of direct endpoint initialization, but does not exclude drift or noise predictors initialized using the corresponding algebraic transformations.

Correct, the model output could be transformed into a drift term (Eq. 6) and supervised with drift matching (Eq. 3). Appendix A.5.1 establishes that output transformation is not necessary to train in a different output space. It can be achieved implicitly by weighting the denoising loss function with a τ\tau-dependent coefficient. Our analysis focuses on DSI trained in x1\mathbf{x}_1-space. The benefits of training in a different output space are left as future work. Section 4.2 now claims that DSI is the only parameterisation in which the prior can be used without transformation. The ablations also acknowledge that output transformation is not covered.

Is Figure 2 intended only to compare direct reuse of an endpoint predictor without output transformations? Please distinguish this claim from an advantage over alternatives initialized to preserve the same dynamics through the transformations in Equation (6) and Appendix A.5.

Yes, this is exactly the intent of Figure 2. Appendix A.5.1 shows that output transformation is equivalent to training a DSI with a τ\tau-weighted loss function. The space in which the loss is calculated is expected to have an effect on the model performance. For instance, in DDPM, Li & He 2026 reports a 0.20 improvement in FID score on ImageNet 64x64 when swapping x\mathbf{x}-loss for v\mathbf{v}-loss in x\mathbf{x}-prediction. However, our experiments and ablations focus on direct SI parameterisations without output transformation. The benefits of implicit learning in different output spaces is left for future work. The manuscript was updated for transparency.

Reviewer_d5scOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Xaurora, a generative weather forecasting framework obtained by fine-tuning a pretrained Aurora model. It introduces Denoising Stochastic Interpolants to preserve the deterministic prior at initialization and combines replay-buffer rollout training with a single-backpropagation CRPS objective. Experimental results demonstrate the effectiveness of the proposed methods.

Critical strengths
  1. The problem and architectural motivation are practically relevant.

  2. The datasets and metrics used for evaluation are appropriate and reasonable.

Critical weaknesses

Missing comparisons with recent probabilistic baselines. The main evaluation only compares Xaurora with IFS ENS, GenCast-O, and the Aurora family. Several recent and directly relevant probabilistic models (e.g., AIFS-CRPS, ArchesWeatherGen, ATLAS, U-Cast, and NeuralGCM) are discussed in the related work but never compared experimentally. In addition, Aurora 1.5 is the most direct competitor (a CRPS-trained probabilistic fine-tune of Aurora), yet it is excluded from the skill comparison for being "only available as an unpublished manuscript", while its training-cost numbers are still used in Table 7 to support the efficiency claims. If Aurora 1.5 is reliable enough to cite for compute comparisons, its forecast skill should be compared as well.

The motivation and the evaluation are mismatched. Extreme weather risk assessment is a key stated motivation (As proposed in the Abstract and Introduction), yet the paper contains no evaluation on extreme events or risk-oriented metrics (e.g., threshold-exceedance skill or case studies of high-impact events).

The reference list contains formatting issues. Many published works are still cited in their arXiv versions, e.g., DDPM (NeurIPS 2020), LoRA (ICLR 2022), and Attention Is All You Need (NeurIPS 2017). The authors should update these entries to the official venues and unify the citation format according to the conference template.

Questions for authors
  1. Could you include a quantitative comparison with recent probabilistic DLWP models (e.g., AIFS-CRPS, ATLAS, U-Cast, and especially Aurora 1.5) on the same WeatherBench protocol? If Aurora 1.5 cannot be evaluated, please justify why its compute figures in Table 7 are considered reliable while its skill scores are not.
  2. The evaluation uses only 56 initialisation dates in a single year (2022), and operational baselines are scored against HRES-T0 while Xaurora is scored against ERA5. How sensitive are the conclusions in Figure 4 to the choice of verification dataset, and could you quantify the uncertainty of the main comparisons (e.g., via paired bootstrap as in Appendix A.4.2)?
  3. Since extreme weather risk assessment is a stated motivation, could you provide an evaluation focused on extreme events (e.g., threshold-exceedance metrics or case studies), and discuss how the observed underdispersion and the noise floor affect reliability for such events?
  4. The paper contains several spelling and grammatical errors. The authors are encouraged to carefully proofread and revise the entire manuscript before the next submission:
    1. On page 8, "The increase in RSME" should be "The increase in RMSE".
    2. On page 22, "discontuity" should be "discontinuity".
    3. On page 25, "the learning rate and the the stationary points" contains a duplicated "the".
    4. On page 26, "every rollouts" should be "every rollout".
3 comments
AuthorsOpenReview ↗

Official response [1/2]

Dear reviewer,

We thank you for your feedback. Following your suggestion, we included additional analyses demonstrating the good performance of Xaurora on extreme events, making the study more complete and compelling. A new version of the manuscript was uploaded, changes relative to the submission are highlighted in blue.

Missing comparisons with recent probabilistic baselines. [...] Could you include a quantitative comparison with recent probabilistic DLWP models (e.g., AIFS-CRPS, ATLAS, U-Cast, and especially Aurora 1.5) on the same WeatherBench protocol? If Aurora 1.5 cannot be evaluated, please justify why its compute figures in Table 7 are considered reliable while its skill scores are not.

Running and evaluating all recent probabilistic DLWP models with the same WeatherBench protocol is computationally prohibitive. Instead, we rely on already available evaluations, many of which, however, use different years as test sets and operate at lower resolution. The table below shows an overview of the available models: AIFS-CRPS, ATLAS, U-Cast, and Aurora 1.5 are not available. Only FGN-O, GenCast-O and IFS ENS are directly comparable to Xaurora.

0.25°, 2022

Model2T, 3 d2T, 5 d2T, 10 dZ500, 3 dZ500, 5 dZ500, 10 dU500, 3 dU500, 5 dU500, 10 dQ850, 3 dQ850, 5 dQ850, 10 d
Xaurora-11.6-16.1-12.4-27.3-21.4-11.5-15.4-14.0-9.6-16.4-15.8-11.9
Xaurora (PP)-0.5-7.0-9.6-12.4-10.1-7.9-8.4-8.0-7.5-9.9-11.2-10.5
FGN-O23.018.38.321.917.67.113.612.75.713.011.86.4
GenCast-O17.412.25.66.05.02.63.64.32.37.46.84.1
0.25°, 2020
Model2T, 3 d2T, 5 d2T, 10 dZ500, 3 dZ500, 5 dZ500, 10 dU500, 3 dU500, 5 dU500, 10 dQ850, 3 dQ850, 5 dQ850, 10 d
---------------------------------------
GenCast18.612.14.16.85.73.35.15.43.39.77.02.6
Astra19.312.34.08.46.03.35.15.43.59.37.22.9
1.5°, 2020
Model2T, 3 d2T, 5 d2T, 10 dZ500, 3 dZ500, 5 dZ500, 10 dU500, 3 dU500, 5 dU500, 10 dQ850, 3 dQ850, 5 dQ850, 10 d
---------------------------------------
NeuralGCM ENS–––5.75.73.43.75.33.31.41.7-0.8
ArchesWeatherGen8.43.9-0.64.23.63.22.14.13.35.94.40.5
Sonny ENS5.33.0-0.45.04.43.03.35.03.55.44.51.0

While Xaurora underperforms compared to FGN-O, GenCast-O and IFS ENS, we note that Xaurora is not presented as the best state-of-the-art weather forecasting model. Instead, the strength of our work is the ability to create a generative model from a good-performing deterministic prior, with minimal additional parameter count and finetuning. Xaurora manages to outperform the strong Aurora prior, with performance approaching state-of-the-art operational NWP/DLWP systems.

Regarding Aurora 1.5 (Weyn et al. 2026), we estimate the results manually from the error curves in Figure 1, as the actual numbers are not explicitly written in the paper (we reached out to the authors, who could not share the exact results). We estimate the RMSE gain of Aurora 1.5 over Aurora operational for 2T at 17% at 3 days and 11% at 5 days. In comparison, Xaurora’s relative improvement over Aurora-S is 15.6% and 19.4% for the same lead times and variable. We note that this is not a direct quantitative comparison of Xaurora and Aurora 1.5, since Aurora 1.5 is using 2024 as a test set, while Xaurora (alongside other methods) is using 2022. Furthermore, the prior from which the two methods bootstrap is different. Nevertheless, the comparison shows that both methods improve over their prior by comparable amounts. A new section comparing Xaurora with a matching non-SDE baseline is available in Appendix B.6.2. As expected from the literature (Kossaifi et al., 2026), the results are comparable, with a small but significant advantage for Xaurora on longer lead times.

The compute numbers in Table 7 are taken from the Aurora 1.5 paper (Table 2).

AuthorsOpenReview ↗

Official response [2/2]

The evaluation uses only 56 initialisation dates in a single year (2022), and operational baselines are scored against HRES-T0 while Xaurora is scored against ERA5. How sensitive are the conclusions in Figure 4 to the choice of verification dataset, and could you quantify the uncertainty of the main comparisons (e.g., via paired bootstrap as in Appendix A.4.2)?

WeatherBench recommends evaluating models on the analysis used to initialise them. It is imperfect, but is the fairest approach. The effect on the metrics is well known (see official benchmark: https://sites.research.google/gr/weatherbench/probabilistic-scores/): operational models perform worse on ERA5 than on HRES-T0, especially at shorter lead times. Therefore, the evaluation protocol used in the paper does not favour Xaurora.

We update the metrics figures (Figure 4 and Figures in B.1) with confidence intervals bootstrapped over initialisations. All models are now evaluated on the same initialisations and with 10 members, except GenCast-O because the forecasts are not publicly available, only the metrics without uncertainties.The re-evaluation barely changes the performance of IFS ENS (Appendix A.4.4, Figure 9). The conclusions of Section 5.1 are unaffected.

The table below shows the uRMSE improvement (%) of the Xaurora models over the Aurora baselines with confidence intervals. Significant results are marked with an asterisk. The improvements of the raw and post-processed Xaurora forecasts over Aurora-S are significant for all variables and lead times. The same conclusions hold for Aurora-L, except for Z500 (raw) at 3 days (non-significant improvement) and U500 (raw) at 3 days (non-significant worsening).

VariableAurora-S vs rawAurora-S vs post-processedAurora-L vs rawAurora-L vs post-processed
Day 3
2T15.62 ± 0.93*18.79 ± 0.93*4.66 ± 1.05*8.24 ± 1.08*
Z50016.83 ± 1.74*19.93 ± 1.77*1.89 ± 2.405.55 ± 2.39*
U5008.96 ± 0.71*11.51 ± 0.72*-0.38 ± 0.932.43 ± 0.93*
Q8509.35 ± 0.60*11.74 ± 0.62*2.96 ± 0.57*5.53 ± 0.59*
Day 5
2T19.39 ± 1.30*22.00 ± 1.31*12.44 ± 1.72*15.27 ± 1.72*
Z50014.46 ± 1.91*17.44 ± 1.93*6.71 ± 3.07*9.96 ± 3.04*
U50012.24 ± 1.02*14.56 ± 1.00*8.13 ± 1.52*10.56 ± 1.50*
Q85014.59 ± 0.61*16.58 ± 0.61*11.72 ± 0.65*13.77 ± 0.65*
Day 10
2T25.80 ± 1.85*27.14 ± 1.84*25.47 ± 1.67*26.82 ± 1.68*
Z50020.34 ± 2.20*21.69 ± 2.20*22.67 ± 2.23*23.98 ± 2.23*
U50019.50 ± 1.45*20.59 ± 1.44*21.45 ± 1.46*22.52 ± 1.46*
Q85019.85 ± 0.78*20.90 ± 0.79*21.16 ± 0.85*22.20 ± 0.86*

The motivation and the evaluation are mismatched. [...]Since extreme weather risk assessment is a stated motivation, could you provide an evaluation focused on extreme events (e.g., threshold-exceedance metrics or case studies), and discuss how the observed underdispersion and the noise floor affect reliability for such events?

A new section (5.3) on extreme events was added to the paper. It highlights that Xaurora substantially improves the representation of extreme hot and cold events over both Aurora-S and Aurora-L. On this task, Xaurora is on par with IFS ENS and FGN (used because GenCast forecasts for 2022 are not available on WeatherBench), and could have a slight advantage on cold extremes. We validate these results with two heatwave case-studies (one in the US, one in Europe).

The Brier score decomposition shows that underdispersion costs Xaurora reliability, not resolution: its resolution matches IFS ENS and FGN, and post-processing improves the reliability term at short lead times, especially for cold extremes.

The noise floor was removed by the November 11 manuscript update. Thus, it does not affect the reliability.

The paper contains several spelling and grammatical errors. The authors are encouraged to carefully proofread and revise the entire manuscript before the next submission. [...] The reference list contains formatting issues.

The listed typos were corrected, references updated to their published venues, and the uploaded manuscript was carefully proofread.

Thank you for the response. In view of the authors' efforts, I will raise my score from 1 to 2. However, given the paper's limited novelty and overall quality (e.g., the lack of comparisons with key baselines), I still lean toward rejection.

Reviewer_vtV8OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Xaurora, an ensemble weather forecasting model obtained by fine-tuning Aurora-S. The method combines a terminal-state parameterization of stochastic interpolants, replay-buffer rollout training, and a fair CRPS objective. The authors evaluate the model on global medium-range forecasting and provide additional analyses of ensemble calibration and spatial spectra.

Critical strengths
  1. The method requires only limited architectural modification to Aurora-S. The added parameters account for about 0.64% of the original model, and the initialization is designed to preserve the behavior of the pretrained deterministic model.

  2. The replay-buffer training procedure is clearly described and avoids backpropagating through complete nested SDE rollouts. This is a practical way to reduce the training cost of stochastic autoregressive forecasting.

Critical weaknesses
  1. Limited methodological novelty. Extending deterministic weather models to probabilistic forecasting is not new, as demonstrated by ArchesWeatherGen, probabilistic Aurora extensions, and AIFS-CRPS. The paper also describes DSI as analogous to clean-state prediction in diffusion models. Therefore, the main contribution seems to be the warm-start parameterization combined with replay-buffer training, but its significance is not sufficiently established.

  2. Unclear necessity and practical advantage of SDE sampling. The paper motivates SDE sampling partly by the excessive smoothing of direct CRPS-trained models, but provides no matched comparison. Xaurora itself exhibits excessive small-scale energy and a high-frequency noise floor. The MC-DropPath baseline is not trained with the same CRPS and rollout objectives, while the MAE/CRPS ablation does not isolate the benefit of SDE sampling. Moreover, improvements over Aurora-S partly come from ensemble averaging, operational comparisons use different analysis datasets, and some variables remain underdispersive. Given the 20 network evaluations per member per 6-hour step, the performance-cost advantage over simpler probabilistic alternatives remains unclear.

Questions for authors
  1. Could the authors provide a matched non-SDE probabilistic baseline using the same trainable parameters, replay-buffer rollout strategy, two-member fair CRPS objective, and training budget, but without iterative SDE sampling? Would Xaurora still show an advantage?

  2. What evidence shows that SDE sampling alleviates the excessive smoothing of direct CRPS-trained models? Could the authors compare member-wise power spectra and scale-dependent spread against a directly CRPS-trained baseline to distinguish realistic small-scale variability from additional high-frequency noise?

  3. Does Xaurora retain an advantage at matched inference cost? The 20 network evaluations per member per 6-hour step could alternatively support more ensemble members or a larger model. Could the authors compare CRPS, calibration, and forecast skill under the same inference budget?

5 comments
AuthorsOpenReview ↗

Official response [1/3]

Dear reviewer,

Thank you for your comments. We appreciate your positive feedback on the appropriateness of the architecture and practicality of the replay buffer. Thanks to your review, we have a better understanding of the practical benefits of CRPS-training and generative modelling in the context of weather forecasting. We conducted additional experiments and believe the results address your concerns. A new version of the manuscript was uploaded, changes relative to the submission are highlighted in blue.

Limited methodological novelty

We respectfully disagree. Our approach has several methodological novelties and advantages that stem from them: (1) the DSI enables parameter- and sample-efficient finetuning, (2) sampling is faster than existing generative DLWP models: 7 minutes for a 6-hourly high-resolution 15-day forecast on a single A100, GenCast takes 8 minutes with 12h timesteps on one Cloud TPUv5, and (3) Xaurora genuinely surpasses its prior and the method scales to the large version of Aurora.

Extending deterministic weather models to probabilistic forecasting is not new, as demonstrated by ArchesWeatherGen, probabilistic Aurora extensions, and AIFS-CRPS.

While other groups have proposed probabilistic DLWP models from deterministic ones, to our knowledge, we are the first to directly finetune a pretrained deterministic model into a generative DLWP. While being conceptually close, Xaurora differs from ArchesWeatherGen substantially in scale and execution. To generate a single prediction, the latter estimates the ensemble mean using four independent frozen backbones, on top of which a separate flow matching model captures the residuals. This residual-based approach has clear limitations, which are addressed by Xaurora: (1) poor generalisation due to overfitting and biases of the frozen backbone – our backbone is not frozen, we directly finetune it, (2) non-Gaussian residuals – we predict full states and use stochastic interpolants that can bridge any distribution, (3) infeasibility of rollout training for generative models – we implement rollout training for SDE-based models. Xaurora also operates at a substantially finer resolution: 0.25º and 6h timesteps instead of 1.5º and 24h. There is no evidence that the ArchesWeatherGen setup would generalise to such an increase in complexity.

CRPS-finetuning of DLWP models is a concurrent and distinct line of work. AIFS-CRPS (Lang et al. 2026) was not finetuned from AIFS (Lang et al. 2024); it is a distinct model, similar to the distinction between GenCast (Price et al. 2025) and GraphCast (Lam et al. 2023). Only recently was the approach shown to be effective to finetune deterministic models (ESFM (Ozdemir et al. 2026), Aurora 1.5 (Weyn et al 2026)). Even though the backbone and application are the same, our approach is methodologically orthogonal to these works. These papers do not undermine the generative direction in DLWP, as illustrated by recent works (Kossaifi et al. 2026, Stock et al. 2026, Aich et al. 2026). On the contrary, we believe that both research directions have their strengths and are worthy of scientific exploration.

The paper also describes DSI as analogous to clean-state prediction in diffusion models

Yes, our DSI is analogous in rationale to clean-state prediction in diffusion models, but we show that DSI outperforms DDPM (see ablations). Our rollout training implementation uses 10 Euler-Maruyama steps, which is sufficient for DSI to bridge xt\mathbf{x}^{t} to xt+Δt\mathbf{x}^{t+\Delta t}, but not for diffusion, which starts from pure noise. We also hypothesise that the new encoder heads used to embed the SDE state could be more difficult to train with diffusion because of high noise levels around τ=0\tau=0. Furthermore, the DSI is novel. We derive the SI drift parameterisations for x1\mathbf{x}_1- and W\mathbf{W}-prediciton (Appendix A.5). We also establish that DSI can be implicitly trained in drift and noise space by reweighting the loss function (Appendix A.5.1).

The main contribution seems to be the warm-start parameterization combined with replay-buffer training, but its significance is not sufficiently established.

Rollout training is an important component to achieve skilful and stable autoregressive forecasts, but is prohibitively expensive for large-scale problems. In this context, our SDE rollout training algorithm is an important technical novelty: it shows that the benefits of multistep adaptation are accessible to generative models with negligible overhead. This was one of the main limitations identified by ArchesWeatherGen and it has implications for generative autoregressive forecasting beyond weather prediction.

AuthorsOpenReview ↗

Official response [2/3]

Unclear necessity and practical advantage of SDE sampling. The paper motivates SDE sampling partly by the excessive smoothing of direct CRPS-trained models, but provides no matched comparison. [...] Could the authors provide a matched non-SDE probabilistic baseline using the same trainable parameters, replay-buffer rollout strategy, two-member fair CRPS objective, and training budget, but without iterative SDE sampling? Would Xaurora still show an advantage?

We have conducted a new ablation study to assess the performance of a non-SDE sampling pipeline. We trained a model with the same hyperparameters, loss function, replay buffer, but without SDE sampling. This new baseline was added to the ablations, Section 5.4 in the updated manuscript. The results and figures are in Appendix B.6.2 and summarised below.

As expected from the literature (Kossaifi et al., 2026), the results are comparable. However, Xaurora still shows a small yet significant advantage, especially on longer lead times. It is significantly better than the non-SDE baseline on uRMSE and CRPS across lead times for Z500 and 2T, and both models are comparable on U500 and Q850. The non-SDE version is less underdispersive on short lead times (even over-dispersive for Z500 and U500). Xaurora surpasses the non-SDE version in terms of calibration at approximately 6 days for 2T, 15 days for Z500 and U500, and 10 days for Q850. At 15 days, the generative version of Xaurora is significantly better across all metrics and variables.

In general, the added value of generative models over single-step approaches is flexibility at inference:

  • Spread/skill control at inference. In general, increasing the number of SDE steps improves the quality of the generation. Furthermore, the spread term is explicit in SI (Wiener process). As illustrated by the experiments in Appendix B.8, the amplitude and structure of the noise can be changed at inference to modify the spread or include physical knowledge.
  • Conditioning and posterior sampling. A score or denoiser enables guidance, inpainting and data assimilation by adding a likelihood term at sampling time, without retraining. Andrae et al. 2026, Manshausen et al. 2025, Chen et al. 2025 treated generative data assimilation. Aich et al. 2026 explored posterior sampling for weather and climate tasks.
  • Physical constraints can be enforced over the SDE integration at inference. Utkarsh et al. 2025, Cheng et al. 2025. In contrast, once a single-step model is trained, little can be changed to influence the performance.

What evidence shows that SDE sampling alleviates the excessive smoothing of direct CRPS-trained models? Could the authors compare member-wise power spectra and scale-dependent spread against a directly CRPS-trained baseline to distinguish realistic small-scale variability from additional high-frequency noise?

The small scales of Z500 and U500 are overdispersed by both models, but most strongly for the non-SDE approach. Both approaches suffer from large-scale underdispersion. This is most pronounced with the non-SDE model on 2T, which could be related to the poor skill observed on this variable. The non-SDE spectra have too much high-frequency power for Z500 and too little large-scale power for Q850. 2T is slightly smoother in the high-frequencies with the SDE. The related motivation statement in the related work section was removed.

Does Xaurora retain an advantage at matched inference cost? The 20 network evaluations per member per 6-hour step could alternatively support more ensemble members or a larger model. Could the authors compare CRPS, calibration, and forecast skill under the same inference budget?

The manuscript revision of November 11 updates the sampler to use 10 network function evaluations (NFEs) per member per 6-hour step. Hence, we compare the finite-ensemble metrics of 10 SDE members, 10 non-SDE members (matched ensemble), and 100 non-SDE members (matched number of evaluations, extrapolated from 10-member statistics). The results are similar to those of the first question: Xaurora retains an advantage. Even at matched NFEs, the SDE performs substantially better on 2T and on Z500 CRPS. RMSE values are comparable and the non-SDE version is better on Q850 CRPS.

AuthorsOpenReview ↗

Official response [3/3]

Improvement over Aurora-S partially comes from ensemble averaging.

Correct, and this is one of the reasons why we use ensemble forecasts in practice. Section 5.1 uses unbiased metrics such that ensembles of different sizes (IFS ENS, GenCast, Xaurora) can be fairly compared. The text acknowledges that ensembling provides a statistical advantage over deterministic models. Appendix B.6.3. of the revised manuscript shows that a two-member Xaurora ensemble (or even a single member for some variable and lead times) beats the Aurora-S prior.

operational comparisons use different analysis datasets.

WeatherBench recommends evaluating models on the analysis used to initialise them. It is imperfect, but is the fairest approach. The effect on the metrics is well known (see official benchmark: https://sites.research.google/gr/weatherbench/probabilistic-scores/): operational models perform worse on ERA5 than on HRES-T0, especially at shorter lead times. Therefore, the evaluation protocol used in the paper does not favour Xaurora.

Some variables remain underdispersive.

The manuscript revision of November 11 updates the sampler and subtracts the residual noise at inference. This fixes the noise floor and leaves the variables homogeneously underdispersive, especially on the short lead times and large spatial scales. We fix the ensemble dispersion with a simple spectral post-processing step. Sampling initial condition uncertainty (using ERA5 EDA as in GenCast) can also improve underdispersiveness. We leave this for future work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final communication

Dear Area Chair, reviewers,

We thoroughly addressed the concerns and questions raised in the original review, resulting in a substantially improved manuscript. The main limitation of the submission (noise floor) has been resolved, and the paper now includes an extreme heat forecasting analysis and a paired comparison with a non-SDE baseline.

We would like to thank Reviewer 2 (d5sc) for raising their score. We believe that our latest responses to Reviewer 1 (vtV8) and Reviewer 3 (6EHX) address their remaining questions and concerns. If that is the case, we kindly ask the reviewers to consider raising their score. Below we re-iterate some key points from our latest responses:

Reviewer 1 (vtV8) and Reviewer 2 (d5sc): We would like to highlight that Aurora 1.5 has not yet been published at a peer-reviewed venue. The authors were not willing to share the official metrics, as the model is still under development and not peer-reviewed. The ICLR 2027 reviewer guidelines clearly stipulate that comparison with unpublished work is not expected and that lack thereof “cannot be a basis for rejection”.

Reviewer 1 (vtV8): The conceptual and methodological novelty of Xaurora was characterised in our last comment. Furthermore, we extensively covered the comparison with non-SDE sampling. We demonstrate that Xaurora has a small but significant advantage especially on large-scale variables and longer lead times. This finding is consistent with Kossaifi et al. (2026).

Reviewer 2 (d5sc): We would like to reiterate that comparing with all latest models is not only computationally prohibitive, but also would not change the conclusions. GenCast and IFS ENS are reference points, not competitors that we are trying to outperform. Instead, we introduce a novel approach for turning a deterministic atmospheric foundation model into a generative one. For transparency, we included the comparison with all the WeatherBench models in Appendix B.1.

Reviewer 3 (6EHX): We would like to thank reviewer 3 for their thoughtful review. We would greatly appreciate their feedback on these revisions and we would be grateful if they could support our work in the next stage of the review process, given their positive initial feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.