acceptodds
Under review as a conference paper at ICLR 2027

From Global to Local: Efficient Regional Weather Downscaling with Global Weather Foundation Model

Abstract

Accurate regional weather prediction requires resolving fine-scale structure while remaining consistent with global dynamics. Traditional limited-area models rely on computationally expensive simulations, while many learning-based approaches frame the problem as super-resolution, overlooking statistical and physical mismatches across scales. We propose a foundation-model-driven downscaling framework that learns regional refinements of global forecasts by augmenting a pretrained weather model backbone with lightweight, multi-scale prediction heads operating directly in its latent space. Despite being pretrained on substantially coarser data, the backbone supports regional adaptation at a tenfold finer grid spacing, without retraining it. The proposed approach uses operational global analyses as inputs and high-resolution regional NWP fields as training targets. It is evaluated not only against gridded datasets but also against ground-based weather station observations, enabling analysis of systematic biases between global analyses, regional WRF forecasts, and real world mesurements. Our model outperforms or matches operational WRF-ARW forecasts on all surface variables and achieves lower RMSE on upper-air variables, at a fraction of the computational cost. It also shows markedly better rollout stability.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Weather Forecasting, Foundation Models, Downscaling, Scientific Machine Learning, Climate & Weather

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_YfZfOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript proposed a simple but effective method to modify the backend of foundation model Aurora to achieve weather downscaling.

Critical strengths

The paper is written reasonably well. The method appears to be simple. Comprehensive evaluations are included in the manuscription.

Critical weaknesses

There could be a persistent distribution shift between Aurora (which is mainly trained by ERA5) and HRRR. The proposed method appears to be using the fine-tuned head to correct the bias. A question is if the distribution shift can't be corrected by the fine-tuned head.

Questions for authors
  1. How is HadISD data used for fine-tuning since they are not on the grid? The paper discussed the regridding of HRRR and DANR data, but not HadISD
  2. Table 1, > $ 1 ^\circ C$ RMSE for T2 is still substantial. Any reason why?
  3. Again in Table 1, the table only listed the RMS of u10u_{10} and v10v_{10}, what are the percentage RMSE?
  4. Do numerical results for HRRR cover the whole CONUS?
2 comments

We thank the reviewer for the time and effort devoted to reviewing and evaluating our work.

W1

Our decoder heads adapt Aurora’s frozen representations to HRRR targets, which can include correcting systematic differences between datasets. To test whether the accuracy gains also constitute useful fine-scale spatial variation, we performed two complementary comparisons.

  • Pooled RMSE, replacing each 40×40 prediction patch with its 4×4 mean -- preserving coarse corrections while removing within-block fine-scale structure
  • HRRR-biased MLP-4-2, a complementary strong bias-correction baseline: a two-layer MLP following our design but predicting coarse 0.25° outputs.

Both coarse representations have similar RMSEs and are outperformed by our full model across all surface variables. RMSE reductions vs. HRRR-biased MLP-4-2 are 17.2% (2 m temperature), 10.1% (10 m u-wind), 8.3% (10 m v-wind), and 3.0% (MSLP). Thus fine-scale prediction adds non-negligible skill beyond coarse bias correction, especially for temperature/wind; the smaller MSLP gain is consistent with its smoother spatial structure.

VariableFull model RMSEPooled RMSEHRRR-biased MLP-4-2
2 m temperature (°C)1.18201.43041.4269
10 m u-wind component (m/s)1.26301.40091.4043
10 m v-wind component (m/s)1.29421.40911.4118
Mean sea level pressure (hPa)0.56410.57570.5815

These results support the predictive value of sub-0.25° spatial variation beyond the tested coarse correction.

Q1

HadISD is used exclusively for evaluation, not for fine-tuning. This is explicitly stated in Section 3 (Data) of the paper.

Since HadISD consists of station-based observations rather than gridded fields, no regridding is performed. For evaluation, station observations are compared with model predictions at the nearest 0.025° grid point. When multiple stations map to the same grid cell, their observations are averaged for intensive variables, as described in Appendix B.4.

Q2

At +6 h, Table 2 reports 1.1743°C for our nearest-neighbor configuration and 1.1749°C for WRF-ARW, both evaluated against the same HRRR analyses under the same protocol. The approximately 1.18°C error therefore represents temperature accuracy comparable to the operational NWP baseline in this evaluation.

The 2022 per-gridpoint RMSE map, rebuttal-outputs/average-rmse-t2m-2022.png (available in our anonymous repository), shows higher errors over mountainous terrain and around the Great Lakes. In mountainous regions, steep elevation gradients can produce sharp local temperature variations over short distances. Accurately predicting these variations remains challenging even at our 0.025° output grid spacing, nominally about 2.7 km. Around the Great Lakes, strong land–water thermal contrasts provide another plausible source of locally complex temperature patterns.

Q3

Below we present percentage RMSE errors calculated in two different ways:

  • Mean wind speed percentage RMSE (MWS RMSE) where both u10 and v10 RMSEs are divided by the mean wind speed calculated over the whole 2022 year (e.g. MWS RMSE (u10) = RMSE (u10) / mean_wind_speed * 100). Mean wind speed was choosen as the non-negative, physically meaningful scale shared by both components.
  • Std normalized percentage RMSE (STD RMSE) where both u10 and v10 RMSEs are divided by the standard deviation over the whole 2022 year (e.g. STD RMSE (u10) = RMSE (u10) / std (u10) * 100).

As can be seen below, for both metrics, our model still outperforms WRF-ARW NWP system.

MWS RMSE u10MWS RMSE v10STD RMSE u10STD RMSE v10
MLPLH-40-2 | MLP-8-2 ↑NN 40 | AUR-MLP-4-0 ↑Bicubic 4030%30.8%37.5%34.6%
WRF-ARW32.7%33.9%40.9%38.1%

The reference mean wind speed and component standard deviations are calculated from HRRR analyses over all timestamps in 2022. The same reference statistics are used for both models.

Q4

No, the main HRRR-grid results do not cover the whole geographical CONUS. Since HRRR is natively defined on a Lambert Conformal Conic (LCC) projection, we regrid it onto a regular latitude–longitude grid aligned with the IFS HRES coordinate system (Section 3.2). The exact extents of CONUS evaluation domain are:

DatasetSouthNorthWestEastGrid (lat × lon)
HRRR25.025°N50°N123°W72.025°W1000 × 2040

No US-boundary or land-only mask is applied.

Thanks for the additional comments. I will take them into consideration when I revise my reviews.

Reviewer_1oYPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a regional weather downscaling framework built on the pretrained Aurora foundation model, using lightweight decoder heads to produce high-resolution forecasts without retraining the backbone. The approach is evaluated against regional NWP forecasts and other baselines. However, the paper needs substantial improvements in presentation, experimental reproducibility, and clarification of the contributions of the proposed decoder.

Critical strengths
  1. The work addresses regional weather downscaling using a pretrained global weather foundation model, with the aim of reducing the dependence on computationally expensive regional modeling workflows.

  2. The proposed framework investigates regional adaptation through lightweight decoder modules rather than retraining the entire foundation model.

Critical weaknesses
  1. The introduction is unnecessarily long and spends too much space reviewing previous work without clearly identifying the problem and contributions. It would be sufficient to explain that existing ML-based regional models offer advantages in accuracy and efficiency over traditional NWP but often rely on NWP inputs and lateral boundary conditions (LBCs).

  2. The related work section lacks a clear structure. Each paragraph should first introduce a research direction and then discuss its limitations, rather than simply listing previous studies.

  3. Section 3.1 presents the global and regional datasets in a confusing manner. A table should summarize their roles as inputs, training targets, validation data, or test data, together with the temporal coverage, spatial and temporal resolutions, and data splits.

  4. The paper does not specify the exact periods covered by Aurora's pretraining and fine-tuning. These details are necessary to assess possible overlap with the training and evaluation periods and the risk of data leakage.

  5. The regridding procedure lacks essential implementation details. The source projection parameters, including standard parallels, central meridian, and projection origin, are not specified. The target grid coordinates, origin, spacing, dimensions, cropping extent, valid-point mask, and interpolation implementation are also unclear. These choices can affect both the training targets and evaluation results.

  6. Section 4 should first explain Aurora's architecture and how its existing design limits regional downscaling before introducing the proposed modifications. The current description mixes Aurora's original components with the newly introduced modules, making it difficult to distinguish the contributions of this work.

  7. The mathematical notation in Section 4 is inconsistent and difficult to follow. Several symbols are introduced without clear definitions immediately following the corresponding equations.

  8. Section 4.4 does not adequately specify the model configurations used in the experiments, including the number of decoder heads, whether variables are predicted jointly or separately, and whether orography is included. The training details are also incomplete, particularly the number of training steps and epochs, stopping criteria, and spatial patch size associated with each batch.

  9. The extensive ablation experiments in Section 5 are difficult to follow because of the complicated notation and insufficient explanations of the model variants. Each experiment should clearly identify the configurations being compared and explain whether it evaluates the contribution of a new module or isolates another source of error.

  10. The description of the single-head decoder in Section 5 appears inconsistent with the earlier model definition. A configuration with subscript 0 is defined as linear, but the corresponding model is later described as nonlinear.

  11. The reported advantage over WRF-ARW reflects the combined performance of Aurora and the regional decoder. It does not establish that the improvement over WRF-ARW is attributable solely to the newly introduced decoder.

  12. There is an incorrect table reference in Appendix D that should be corrected.

  13. The numerical claim in Appendix J concerning Table 5 appears incorrect. For the two-hidden-layer configuration, the (u_{10}) RMSE changes from 1.2681 to 1.2731 (approximately 0.39%), while MSLP changes from 0.5700 to 0.5653 (approximately 0.82%). These differences are not both below the stated 0.3%.

  14. Table 7 reports results from ten random seeds using the best configuration, but the exact configuration is not specified. Its numerical results also differ slightly from the best results in Table 1, without explanation.

Questions for authors
  1. Which Aurora checkpoint was used, and what periods were included in its pretraining and fine-tuning? Is there any overlap with the training or evaluation periods?

  2. Can the authors provide a complete dataset summary and the exact source projection, target grid, cropping, masking, and interpolation configurations needed to reproduce the results?

  3. What are the exact model and training configurations, including decoder heads, variable-specific settings, orography inputs, patch sizes, training duration, and stopping criteria?

  4. How much of the reported improvement over WRF-ARW can be attributed to the proposed regional decoder rather than the combined Aurora–decoder forecasting system?

  5. Can the authors clarify the model definitions used in the ablation experiments, correct the numerical inconsistency in Table 5, and explain the differences between Tables 1 and 7?

1 comment

We thank the reviewer for the time and effort devoted to reviewing and evaluating our work.

We thank the reviewer for identifying the presentation issues and small inconsistencies (Weaknesses 1, 2, 6, 7, 10, 12, and 14). These mainly reflect inadvertent mistakes and editorial refinements that we will address in the camera-ready version.

W5

HRRR and DANRA use tangent Lambert conformal conic source projections with the following parameters and regular latitude–longitude target grids:

ParameterHRRRDANRA
Standard parallels (φ₁ = φ₂)38.5°N56.7°N
Latitude of projection origin38.5°N56.7°N
Central meridian97.5°W25°E
False easting / northing0 / 0 m0 / 0 m
Spherical Earth radius6,371,229 m6,367,470 m
Native array (rows × columns); spacing1059 × 1799; 3 km589 × 789; 2.5 km
Target grid (lat × lon)1000 × 2040360 × 1120
Target latitude centres50 − 0.025i, i = 0,…,99961 − 0.025i, i = 0,…,359
Target longitude centres (degrees east)237 + 0.025j, j = 0,…,2039−4 + 0.025j, j = 0,…,1119

Prediction variables use bilinear interpolation (xESMF 0.9.2; ESMF/ESMPy 8.9.1). Native grid-relative winds are rotated to geographic U/V before interpolation. Preserved bilinear operators support all 2,040,000 HRRR and 403,200 DANRA target centres, all lie within interpolation support, so no extrapolation is required.

W9

The naming convention (MLP-S-L; H/LH; ↑NN/Bicubic_40; | and +) is fully defined in Section 4.1, and Tables 1–2 enumerate every configuration. However, the prose interleaves results and ablation logic without stating which axis each paragraph isolates. In camera-ready, Section 5 will be restructured around explicit ablation axes, each introduced by a sentence stating what is isolated:

  • Upsampling method (NN vs Bicubic): interpolation scheme.
  • Learned vs non-learned (IFS HRES↑ vs AUR-MLP-4-0↑): value of learned representations.
  • AUR-MLP-4-0 non-biased vs biased: linear bias correction
  • Latent vs patch space (AUR-MLP-4-0+MLP-40-1 vs AUR-MLP-4-0+MLPP-40-1): value of accessing full embeddings.
  • Depth (L=0 vs 1 vs 2): effect of decoder capacity.
  • Multi-head (single vs |MLP-8-L): benefit of multi-scale heads.
  • Orography (none vs H vs LH): orography-injection mechanism.
  • Pretrained decoder inclusion (with/without AUR-MLP-4-0): contribution of Aurora's native decoder.

W10

MLPP-40-0 is a linear decoder operating on Aurora's 4×4 output patches; Section 5 will be corrected to "linear" in camera-ready.

W13

In camera-ready we will loosen the claim from "below 0.3%" to "below 1%", accurately reflecting all measured differences across single- and two-hidden-layer configurations.

Q1 & W4

Aurora was pretrained on 10 datasets with varying temporal coverage, all ending in 2020. We use the version fine-tuned on IFS HRES-025 T0 analysis fields from 2016 to 2021. There is no temporal overlap between training and evaluation periods.

Q2 & W3

DatasetRoleDomain/res.Grid (lat×lon)TimeSurface varsAtmospheric vars (levels)Periods
IFS HRES (WeatherBench2)Input (global)global 0.25° regular lat–longlobal6-hourly analysis2 m temp, 10 m U/V, MSLPgeopotential, temp, q, U/V, w at 13 levels 50–1000 hPainput only
HRRRTarget+NWP baselineCONUS 25.025–50°N, 123–72.025°W, ~3 km1000×20406-hourly2 m temp, 10 m U/V, MSLPtemp, U/V, geopotential at 10 levels 200–1000 hPatrain 2017–2021; eval 2022
DANRATarget52.025–61°N, 4°W–23.975°E, 2.5 km360×11206-hourly2 m temp, 10 m U/V, MSLPtemp, U/V, geopotential at 11 levels 100–1000 hPatrain 2017–2019; eval 2020 excl. Nov–Dec
HadISDEval only (stations)CONUS ~25–50°N, ~123–72°W, station-level~1,800 pts6-hourly2 m temp, 10 m U/V—eval 2022 (CONUS) / 2020 (Danish domain)

Q3 & W8

All of these details are already provided in the paper. The decoder heads in the MLP-S-L family and orography injection methods are described in Section 4.1. Aurora's backbone produces patch embeddings corresponding to 1°x1° spatial patches, and the native decoded output is a 4x4 grid per patch. Training duration is approximately 5 hours on 4 H100 GPUs, or 20 hours on a single GPU (Section 4.4). No early stopping or convergence criteria were used.

Q4 & W11

The regional decoder cannot be evaluated in isolation from Aurora's backbone, as it operates directly on Aurora's latent embeddings. Table 1 shows that both components contribute. Aurora's backbone provides rich latent representations, which is then refined by decoders.

Q5 & W12 & W14

Table 7 reports AUR-MLP-4-0↑NN40 | MLP-40-2 | MLP-8-2↑NN40, which we mistakenly identified as the best-performing configuration. We will correct this by explicitly naming the model. The seeding experiment uses a high-performing configuration to quantify random-seed variation.

Reviewer_Rej2OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a regional weather forecasting framework that freezes the Aurora backbone and trains lightweight, multi-scale decoders on its latent representations to produce 0.025° outputs from 0.25° global inputs. It evaluates the approach over the contiguous United States and the Danish domain using high-resolution gridded data, weather station observations, and physical consistency diagnostics, reporting competitive accuracy against regional forecasting baselines at low computational cost.

Critical strengths
  1. The frozen-backbone approach offers a practical and computationally efficient route to kilometer-scale regional forecasts, with competitive accuracy against operational baselines.

  2. Evaluation against both gridded targets and station observations strengthens the evidence that improvements extend beyond matching the training data source.

Critical weaknesses
  1. The novelty relative to prior frozen-backbone adaptation methods is insufficiently established. In particular, Prithvi WxC (Schmude et al., 2024, Section 3.1.3) already demonstrates downscaling with frozen core layers, contradicting its characterization as requiring full backbone retraining. The baseline comparison is also too limited: a direct comparison with the Prithvi WxC adaptation approach under a consistent evaluation protocol is needed to establish the proposed method’s added value.

  2. The experiments do not sufficiently separate fine-scale predictive gains from systematic bias correction between IFS inputs and HRRR targets. Although an “HRRR-biased” baseline is included, its training procedure and capacity to correct biases are insufficiently described, leaving this attribution unresolved. Stronger bias-correction baselines and scale-separated evaluation are needed to substantiate the claimed downscaling capability.

Reference: Schmude, J., et al. (2024). Prithvi WxC: Foundation Model for Weather and Climate. arXiv:2409.13598. https://doi.org/10.48550/arXiv.2409.13598

Questions for authors
  1. How much of the improvement remains after aggregating predictions and targets to 0.25°? Do the gains also hold for fine-scale residuals and extreme temperature or wind conditions?
  2. For the comparisons with WRF-ARW and StormCast, are initialization times, valid times, evaluation samples, and spatial coverage identical? Which StormCast ensemble statistic is used to compute RMSE?
3 comments

We thank the reviewer for the time and effort devoted to reviewing and evaluating our work.

W1

We thank the reviewer for highlighting this similarity. We acknowledge our original description was imprecise, both Prithvi WxC and our method use a frozen foundation model for regional downscaling. However, two clear distinctions remain:

  • Training computation: Prithvi WxC places trainable downscaling layers before the frozen backbone, therefore the gradients must propagate through the frozen 2.3B-parameter backbone, incurring activation-storage and backward-pass costs. Our trainable decoder operates after the backbone on precomputed cached latent embeddings. Backbone forward pass is computed once, then training avoids backbone forward/backward passes and activation storage. This lets a 16-sample batch fit on one GPU rather than requiring one GPU per sample. Reusing the frozen global model with lightweight heads also supports low- and high-resolution rollouts without further training.
  • Experimental setup: Prithvi WxC’s uses 2096-2100 CORDEX simulation and downscales daily means from 1.32° to 0.11° in a perfect-model framework, forming coarse inputs by degrading the high-resolution target. We downscale 0.25° to 0.025° (where fine-scale variability matters more) using inputs from a different dataset, available operationally at real-time. These differences in target scale, temporal resolution, and input construction limit direct comparison. Training and evaluation settings of Prithvi differs significantly from any real-world pipeline, whereas we aim at operational applicability.

We will clarify our description of Prithvi WxC in the camera-ready version.

Q1 & W2

First, we note that, in addition to the HRRR experiments, we provide a zero-shot HadISD stations evaluation in which our model favors WRF-ARW even more strongly than on HRRR. If our gains were mainly attributable to bias correction, and if WRF-ARW better represented real-world fine-scale structure, in-situ weather-station results should confirm this.

The HRRR-biased baseline is Aurora’s pretrained linear decoder (AUR-MLP-4-0), fine-tuned on HRRR fields with the backbone frozen. It uses the same MAE objective and training protocol as all the other variants, it captures only the linear bias correction.

To separate bias-correction, we use:

  • Pooled RMSE, replacing each 40×40 prediction patch with its 4×4 mean -- preserving coarse corrections while removing within-block fine-scale structure
  • HRRR-biased MLP-4-2, a complementary strong bias-correction baseline: a two-layer MLP following our design but predicting coarse 0.25° outputs.

Both coarse representations have similar RMSEs and are outperformed by our full model across all surface variables. RMSE reductions vs. HRRR-biased MLP-4-2 are 17.2% (2 m temperature), 10.1% (10 m u-wind), 8.3% (10 m v-wind), and 3.0% (MSLP). Thus fine-scale prediction adds non-negligible skill beyond coarse bias correction, especially for temperature/wind; the smaller MSLP gain is consistent with its smoother spatial structure.

VariableRMSE (IFS2HRRR)Pooled RMSE (IFS2HRRR)HRRR-biased MLP-4-2
2 m temperature1.18201.43041.4269
10 m u-wind1.26301.40091.4043
10 m v-wind1.29421.40911.4118
MSLP0.56410.57570.5815

For scale-separated evaluation, our anonymous repository includes scale-dependent skill vs. baselines, including a strong U-Net post-processing baseline and WRF-ARW (rebuttal-outputs/scale-skill-surf.png). We evaluate structural agreement with HRRR analysis over full-year 2022: fields are mean-removed, Hann-windowed, Fourier-transformed, and cross-spectral correlation is computed across 32 spatial wavelength bins (~6–384 km; signed correlation −1 to 1).

Q2

WRF-ARW, StormCast, and our model are evaluated over CONUS at the same +6h lead using 00/06/12/18 UTC initializations during the same 2022 evaluation period, with the corresponding reference grid and HadISD stations. StormCast RMSE uses the five-member ensemble mean.

Spatial coverage requires one qualification: WRF-ARW covers the full regridded HRRR domain, while StormCast covers only a subset.

DatasetSouthNorthWestEastGrid (lat × lon)
HRRR25.025°N50°N123°W72.025°W1000 × 2040
StormCast32.025°N44.75°N107.5°W88.025°W510 × 780

Table 2 evaluates WRF-ARW, Aurora, and our model over the full regridded HRRR domain. Appendix J Table 8 compares StormCast and our best model over the same restricted domain. We will clarify this coverage difference in the Table 2 caption and main text, and refer readers to Table 8. Our response to Reviewer Xe6N (Weakness 2) further compares initialization, data assimilation, boundary conditioning, and rollout procedures.

Thank you for addressing my other concerns. However, without a quantitative comparison with Prithvi WxC under comparable settings, the performance advantages of the proposed approach remain insufficiently established. I therefore maintain my original score.

We agree that the value of our adaptation strategy should be demonstrated experimentally. However, direct comparison would conflate adaptation effects with differences in resolution and dataset (Prithvi WxC MERRA-2, 0.5° × 0.625°; Aurora: ERA5-IFS HRES T0, 0.25° × 0.25°). We therefore compare both approaches using a shared, frozen Aurora backbone and a reimplementation of the Prithvi/Granite-style adapter, isolating the effect of the adaptation strategy from pre-trained model quality.

We evaluate six configurations to isolate the effects of the decoder, a learned convolutional adapter, and input resolution.

Common components:

  • All pipelines use Aurora’s encoder (Perceiver), backbone, and original linear decoder, which remain frozen.
  • The learned convolutional adapter is inserted before the backbone.
  • For coarser input convolutional adapter incresase embeddings' dimensionality. Subsequently pixel shuffle is used to restore the required backbone dimension.
  • The Prithvi decoder/upscaling stage is applied after Aurora’s original decoder where indicated.
  • We do not implement additional skip connection as it is implicitely included in the Aurora architecture.
Exp.InputConv. adapterOutput decoder
1Full-resolution IFSpreserves dim.Aurora linear + Prithvi upscale
2Full-resolution IFSpreserves dim.Our MLP decoder
3Full-resolution IFSNoneAurora linear + Prithvi upscale
4Full-resolution IFSNoneOur MLP decoder
5IFS coarsened 2×2\timesincrease dim. + pixel shuffleAurora linear + Prithvi upscale
6IFS coarsened 2×2\timesincrease dim. + pixel shuffleOur MLP decoder

Our MLP decoder is AUR-MLP-40-2|8-2 + AUR-MLP-4-0, the best configuration without orography.

Experimental comparisons:

  • 1 vs. 2 and 3 vs. 4: isolate the decoder's influence, with and without the learned adapter.
  • 1 vs. 3 and 2 vs. 4: isolate the influence of the learned adapter.
  • 5 and 6: evaluate coarsened-input variants of experiments 1 and 2, using 2× input coarsening and pixel-shuffle upsampling to reproduce as faithfully as possible the corresponding Prithvi configuration.

Learned adapter addition effect The key difference is computational: upstream adapters in the Prithvi approach require gradients to propagate through the frozen backbone (Exp. 1, 2, 5, 6), whereas our downstream readout (Exp. 3, 4) trains on cached embeddings, eliminating repeated backbone execution and activation storage. This yields a 27.6× higher optimizer-step throughput. Within the same training budget, the identity-initialized adapter achieves lower RMSE than the random-initialized run (reductions of 8.7% for T2, 17.3% for U10, 16.3% for V10, and 19.5% for MSLP). The learned adapter addition (Exp. 1, 3) resulted in worse metrics across all variables.

Decoder architecture effect Comparing the two pairs of experiments (1, 2) and (3, 4) suggests that our MLP decoder achieves lower RMSE than the Aurora linear decoder followed by Prithvi-style upscaling. Directly decoding from embeddings can be more effective than relying grid-based outputs, which is consistent with our previous conclusions.

Coarsening effect As expected, coarsened input variants show worse RMSE.

These experiments do not rank the original foundation models. They isolate the computational and adaptation trade-offs between upstream preprocessing and downstream readout training on a common backbone.

Prithvi-inspired adaptation with full model forward pass

Time budget extended by 4x wrt. original experiments**

Exp.T2U10V10MSLP
14.96821.80221.93523.0838
23.09241.58181.57741.8104
55.13271.91752.03713.2980
63.64881.80561.69432.0501

Our downscaling strategy -- only decoders training

Exp.T2U10V10MSLP
31.65801.48551.50001.3547
4 (Our method)1.18321.27311.30240.5653

We hope these additional experiments directly address the reviewers’ concerns by demonstrating the performance and computational advantages of our proposed approach under a controlled, comparable setting. We hope these results help clarify the contribution of our method and address the remaining concerns regarding its novelty and practical value.

Reviewer_Xe6NOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The frozen-backbone transfer result and independent station evaluation provide useful empirical evidence. However, uncertainty about event-specific fine-scale recovery, equivalent operational inputs, and unseen-region generalization limits the strength of the conclusions. Clarifying these issues could materially affect my final rating.

Critical strengths

na

Critical weaknesses
  1. Fine grid spacing is not sufficient evidence of recovered mesoscale forecast information. The heads independently decode each 1° patch, and the backbone never receives the evolving fine-scale regional state. Improved RMSE can therefore reflect bias correction, terrain-conditioned climatological structure, and smoothing as well as event-specific downscaling. The spectral and physical analyses are useful, but do not establish localization and skill of newly resolved features. Verification should include scale-dependent skill against strong spatial post-processing baselines, patch-boundary diagnostics, and event-level tests of fronts, coastal flows, strong winds, and temperature extremes.

  2. The operational comparisons require clearer control of available information and initialization. The global model is driven by IFS analyses, while WRF-ARW, StormCast, and ML-LAM use different forecasting systems, input states, and boundary protocols. The exact initialization timestamps, forecast cycles, assimilation availability, target valid times, and StormCast ensemble statistic should be documented. In particular, rollout results must clarify whether Aurora is advanced autoregressively without later analyses. Without this audit, lower RMSE does not isolate the decoder contribution or establish superiority under equivalent operational information.

  3. Generalization and rollout claims are broader than the adaptation experiments establish. Joint training on CONUS and Denmark demonstrates sharing across two observed regions, rather than transfer to an unseen region. Similarly, stable regional outputs largely inherit the global backbone's evolution, since fine-scale predictions are not fed back into its dynamics. Cross-region holdouts, multi-year tests, and event-stratified rollout verification are needed to assess whether the heads learn transferable regional dynamics rather than region-specific mappings. The paper should distinguish stability inherited from Aurora from any stability contributed by the regional decoder.

Questions for authors

na

1 comment

We thank the reviewer for the time and effort devoted to reviewing and evaluating our work.

W1

We agree that fine grid spacing alone is not sufficient evidence of recovered mesoscale information, so we have added requested checks and evaluated agains strong U-Net (descibed below) baseline.

Patch-boundary diagnostics: We compare spatial gradient magnitudes at patch seams vs. interior pixels over the evaluation period. The seam ratio is geometric-mean of seam gradient divided by interior gradient (ratio near 1 indicates no boundary artifacts). Patch-free ground truth is the control. We evaluate at decoder-specific patches: 40 px (MLP-40), 10 px (original decoder: 4 px/token, ×10 nearest-neighbour upsampling), and 5 px (MLP-8: 8×8 px/token, ×5 tiling). Figures: rebuttal-outputs/seam-gradient-histogram-surf.png and seam-interior-gradient-ratio-surf.png.

Scale-skill scores: Anonymous repository includes scale-dependent skill vs. baselines, including a strong U-Net post-processing baseline and WRF-ARW forecasts (rebuttal-outputs/scale-skill-surf.png). Over full-year 2022 HRRR analysis, fields are mean-removed, Hann-windowed, Fourier-transformed, and cross-spectral correlation is computed across 32 wavelength bins (~6–384 km). Correlation ranges from −1 to 1. Our model has substantially higher short-wavelength coherence than strong learned post-processing baselines, including an attention-based U-Net.

Event-level evaluation: Event visualizations (rebuttal-outputs/event*) compare our best model, Aurora, and HRRR.

U-Net baseline details: U-Net (ADM-style attention at the two deepest levels) baseline is build on the frozen Aurora's model and performs pixel-level superresolution of Aurora's output rather than decoding from its latent representation. It is trained on the same 2017–2021 period (2022 held out for evaluation) using both surface and atmospheric variables.

W2

We provide below a table with all the requested information.

Despite differing initialization sources and boundary conditioning, we ensure fair evaluation by giving each method its required inputs and evaluating all forecasts at the same valid times and period. Aurora is rolled out autoregressively, predictions are fed back as inputs, with no IFS analyses injected and no post-processing. For spatial coverage, Table 2 evaluates WRF-ARW, Aurora, and our model over the full regridded HRRR domain but StormCast on a restricted domain. Appendix J Table 8 compares StormCast and our best model over that same restricted domain. We will clarify this difference in the Table 2 caption and main text, and refer readers to Table 8.

AttributeOur modelsWRF-ARW (HRRR)StormCast 5-memberML-LAM (DANRA)Aurora 0.25° base
Init. sourceIFS HRES t=0 analysis from WeatherBench2HRRR analysis (V2/V3 RAP-derived; V4 HRRRDAS ensemble mean)HRRR analysis states (conditioned on NCEP GFS 0.25° global forecasts)DANRA analysis (+ SYNOP station info)IFS HRES t=0 analysis from WeatherBench2
Init. times00/06/12/18 UTC (6-hourly) or 00/12 (12-hourly)00/06/12/18 UTC00/06/12/18 UTC00/12 UTC00/06/12/18 UTC
AssimilationNone (IFS analyses as-is)GSI hybrid 3D EnVar hourly; radar reflectivity every 15 min; lightning; cloud analysis; 36-member HRRRDAS covariances (v4)None (HRRR analysis includes assimilation)None (DANRA includes 3D-Var + optimal interpolation)None (IFS analyses as-is)
LBC sourceN/A (global)RAP (Rapid Refresh)N/A (global conditioning)IFS HRES forecasts during eval (ERA5 during training)N/A (global)
Target valid timesAll: +6 to +48 h every 6 hAll: +6 to +48 h every 6 hAll: +6 to +48 h every 6 hAll: +6 to +48 h every 6 hAll: +6 to +48 h every 6 h
Spatial res.0.025°3 km native LCC3 km native LCC2.5 km native LCC0.025° (NN-upsampled from 0.25°)
Temporal res.6-hourlyhourly to +18 h; 6-hourly to +48 hhourly3-hourly6-hourly
Ensemble stat.N/A (deterministic)N/A (deterministic)5-member meanN/A (deterministic)N/A (deterministic)
Eval. period2022 CONUS; 2020 DANRA Jan–Oct202220222020 Jan–Oct2022

W3

We respectfully clarify that the paper does not claim:

  • Transfer to unseen regions. Table 6 (Appendix J) shows that one decoder trained jointly on CONUS and Denmark loses little accuracy vs. region-specific models, demonstrating multi-region training stability and scalability, while generalization beyond the training domain is explicitly future work (Section 6).
  • Rollout stability superior to Aurora or attribute it primarily to the regional decoder. Instead, the frozen Aurora backbone provides stable global evolution that the regional decoder inherits, avoiding separate rollout fine-tuning, a deliberate design choice, not a limitation, which we will make explicit in revision.

Comments on the paper

1 comment

Dear Reviewers and Area Chair,

We thank the reviewers for their careful reading and constructive feedback. The discussion has helped us sharpen the contribution, strengthen the experimental evidence, and identify aspects of the presentation that require clarification. We summarize the main additions and conclusions below.

1. Additional evidence of fine-scale predictive skill beyond coarse bias correction.

We added pooled version of our predictions that removes sub-0.25° spatial variation, and a stronger nonlinear coarse-resolution (U-Net) baseline and HRRR-biased MLP-4-2. The full-resolution model outperforms both. These results support the predictive value of fine-scale spatial variation beyond the bias correction.

We also provide scale-dependent verification, including comparisons with an attention-based U-Net trained to post-process Aurora outputs using the same training period and variables. The cross-spectral analysis shows stronger short-wavelength agreement with HRRR for our latent-space decoder than for the learned post-processing baselines. Additional patch-boundary diagnostics and event visualizations examine spatial continuity and local structure.

These analyses complement the HadISD station evaluation.

2. Clarified relationship to Prithvi WxC and the scope of our contribution.

We acknowledge that Prithvi WxC also demonstrates downscaling with frozen core layers. A direct comparison between the original systems is confounded by differences in backbone architecture, pretraining data, resolution, and input distributions. We therefore added controlled experiments using a shared, frozen Aurora backbone to compare upstream adapters with downstream readout training. Cached-latent training achieves 27.6× higher optimizer-step throughput and therefore much better results in comparable time (4x longer for Prithvi) by avoiding repeated backbone execution and activation storage.

Our evaluation also addresses operational short-range forecasting from IFS HRES initial conditions to HRRR forecasts at a six-hour lead, distinct from Prithvi WxC's published downscaling settings.

3. Explicit evaluation protocols and spatial coverage.

Our responses document initialisation sources and cycles, data assimilation, boundary conditioning, forecast lead times, evaluation periods, and ensemble statistics.

4. Reproducibility, presentation, and scope.

We supplied dataset and grid specifications, interpolation and Aurora training-data information. We will clarify the ablation structure and correct the identified notation, numerical statements, and configuration descriptions in camera-ready version.

We also clarify that our paper explicitly stated joint CONUS–Denmark training demonstrates adaptation across two observed regions; unseen-region generalisation remains future work. Regional rollout stability inherits Aurora’s global evolution, and the operational results concern the combined Aurora–decoder system.

We believe that all the reviewers' concerns were answered by our rebuttal.

We thank the reviewers once again for the time devoted to through analysis of our work. This will undoubtedly improve the quality of the final manuscript.

Best regards, The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.