acceptodds
Under review as a conference paper at ICLR 2027

Learning Kilometer-Scale Weather Prediction with Global-Regional Alignment

Abstract

Kilometer-scale regional weather forecasting is essential for local weather warnings and weather-sensitive decisions. Existing data-driven approaches often rely on numerical forecasts for large-scale guidance or require additional training of global forecasting components. Pretrained global weather models offer an efficient source of large-scale forecasts, motivating their reuse to guide high-resolution regional prediction. However, this coupling requires aligning global and regional representations across different grids and integrating global guidance with local interactions to advance regional states. We propose ScaleCast, a regional forecasting framework that addresses these challenges through Global–Regional Alignment. Its Global–Regional Conversion module aligns joint global and regional representations with regional locations, while the Global–Regional Alignment and Dynamics block combines aligned guidance with regional neighborhood interactions. Experiments using ERA5 global analyses on a 0.25° grid and CERRA regional reanalysis at 5.5 km spacing demonstrate improved regional forecasts across surface and upper-air variables, with a single trained model supporting multiple global forecast drivers (i.e., Pangu-Weather, GraphCast, and HRES) without specific retraining. Fine-tuning on HRRR at 3 km spacing further demonstrates the framework’s adaptability to a different regional domain and spatial resolution. Windstorm case studies show improved cyclone positioning and core-pressure estimates, while comparisons with HadISD station observations show closer agreement with local temperature and humidity changes. Code is available at https://anonymous.4open.science/r/ScaleCast-7C2D

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Regional forecasting, Global-Regional Alignment

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Vm6fOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ScaleCast, a regional weather forecasting framework that improves prediction accuracy through global-regional alignment and learnable positional projections. The framework integrates large-scale atmospheric information from pretrained global weather forecasting models with fine-scale regional dynamics. Its flexible design allows different global forecasting models to serve as drivers, with experiments demonstrating the effectiveness of different global-regional configurations. The authors further evaluate ScaleCast against numerical and data-driven forecasting baselines, as well as station observations from HadISD, demonstrating its competitive performance in regional weather prediction.

Critical strengths
  • The key strengths of this paper is its extensive evaluation across different global forecasting models and validation datasets. By demonstrating consistent improvements with multiple global drivers and evaluating predictions against both gridded reference data and station observations, the authors provide evidence of the framework's flexibility and practical applicability.
Critical weaknesses
  1. Insufficient analysis of autoregressive forecasting behavior. Although the model supports autoregressive forecasting up to 72 hours, the main quantitative evaluation focuses on forecasts up to 30 hours. Furthermore, the relatively stable error trajectories are not sufficiently explained, particularly regarding the respective contributions of regional autoregressive dynamics and continuously updated global forecast guidance. This makes it difficult to assess whether the reported improvements primarily reflect learned regional dynamics or reliance on the global forecasts.

  2. Insufficient differentiation from closely related work. The proposed global-regional alignment framework shares substantial conceptual similarities with recent global-regional coupling approaches, particularly ScaleMixer (Chen et al., 2026). Without a clearer methodological comparison, it is difficult to assess the novelty and distinct contribution of the proposed alignment mechanism.

Questions for authors
  1. Autoregressive forecasting and error accumulation Despite its autoregressive design, ScaleCast shows relatively stable errors over increasing lead times. Could the authors clarify whether this stability primarily results from the model architecture or the continuous guidance from global forecasts? Since new global predictions are provided at each step, it is unclear to what extent the model independently learns regional atmospheric dynamics. A quantitative evaluation over the full 72-hour horizon would also strengthen this claim.

  2. Evaluation of locally varying weather variables High-resolution regional forecasting is particularly important for variables strongly influenced by local conditions, such as wind and precipitation. However, these variables are not included in the reported evaluation. Is this due to data availability or model limitations? Can ScaleCast effectively predict these variables?

  3. Why is Graph-EFM evaluated using a four-member ensemble mean, while StormCast-Reg is evaluated with only a single member, potentially limiting the benefits of its generative approach? Was the diffusion component included in StormCast-Reg?

  4. What criteria are used to spatially crop the global forecast fields to the regional domain?

  5. Could the authors clarify the methodological differences between ScaleCast and ScaleMixer, particularly regarding their global-regional alignment mechanisms?

5 comments
AuthorsOpenReview ↗

Response for Reviewer Vm6f (1/2)

We thank the reviewer for the constructive comments and suggestions. We have extended the evaluation to 72 hours and added wind, ensemble-size, and global-input coverage comparisons. We will incorporate these results into the revised manuscript, clarify the autoregressive forecasting and baseline configurations, and expand the discussion of ScaleMixer.

W1 & Q1. Regional State Evolution under Continued Global Guidance from 3.

Thank you for this helpful suggestion. We have extended the quantitative evaluation to +72 h. The table below reports the mean RMSE across the 14 three-hour leads from +33 to +72 h. Publicly available CERRA forecasts extend to +30 h, so CERRA NWP results are unavailable for this interval.

Table R1. Mean RMSE over +33–72 h. Lower is better.

MethodMSLPT2mRH2mT500T850Z500Z850RH500RH850
ECMWF-HRES183.21.465N/A1.0750.9971.5571.330N/AN/A
CERRA NWPN/AN/AN/AN/AN/AN/AN/AN/AN/A
Pangu190.61.801N/A1.0841.0721.4711.378N/AN/A
GraphCast140.11.683N/A0.9961.0731.4451.016N/AN/A
ML-LAM137.21.5569.7500.9501.0521.4441.04220.31013.674
Hi-LAM134.91.4939.8400.9671.0481.4051.03720.17513.517
Graph-EFM135.81.6819.8230.9671.1281.4041.04020.17313.841
OneForecast158.32.37710.6470.9571.1611.5741.17720.44014.163
StormCast133.61.4499.9230.9631.0561.4321.02320.15713.884
SC + Pangu134.91.4789.4110.9710.9501.3381.00820.24313.946
SC + GraphCast115.01.3808.7610.8650.9001.1990.84818.32012.679
SC + HRES113.31.4499.1770.8390.9371.1850.85719.63612.305

The model rolls out stably through +72 h with regional autoregressive updates and global forecast guidance from the same initialization cycle.

W2 & Q5. Relationship to ScaleMixer.

ScaleMixer is closely related and will be discussed in the revised paper. It uses internal global-model tokens and bidirectional global–regional interaction, including adaptive key-position sampling and two-stage cross-attention. ScaleCast consumes external forecast fields and updates the regional state, supporting the evaluated neural and numerical forecast products.

Table R2. Interface and coupling comparison with ScaleMixer.

AspectScaleCastScaleMixer
Global inputExternal forecast fieldsInternal global-model tokens
CouplingField-guided regional state updateBidirectional global-regional token interaction
Global parameters during regional trainingFrozen global forecast providerFrozen pretrained global model

The advantage of our method is that it can be integrated into various models. We will add the related-work discussion.

Q2. Wind and Precipitation Outputs.

Thank you for this suggestion. We have extended wind speed on HRRR under actual GraphCast forecast conditioning.

Table R2. Mean 10-m wind RMSE (m/s) over +3–72 h. Lower is better.

RegionMethodU10V10Wind speed
EastSC + GraphCast1.3671.4631.398
EastBilinear GraphCast1.8712.1031.861
WestSC + GraphCast2.2222.1492.349
WestBilinear GraphCast2.8022.7942.717

Precipitation is available in CERRA as a forecast accumulation, while the CERRA analysis fields do not include precipitation. So, we are unable to conduct a precipitation evaluation.

Q3. Ensemble Size and the StormCast Diffusion Component.

Thank you for your suggestion. The original StormCast-Reg was used with residual diffusion. We now report both single-member and four-member ensemble-mean results for StormCast and will use the four-member results in Table 1.

StormCast single-member and four-member ensemble-mean RMSE. Lower is better.

Lead rangeMethodMSLPT2mRH2mT500T850Z500Z850RH500RH850
+3–30 hStormCast, 1 member90.51.0046.6910.7130.7600.7350.66816.6229.905
+3–30 hStormCast, 4 members87.30.9896.6710.7020.7490.7210.64916.4389.795
+33–72 hStormCast, 1 member134.51.4729.9430.9741.0681.4391.03720.29314.004
+33–72 hStormCast, 4 members133.61.4499.9230.9631.0561.4321.02320.15713.884

The improvements are modest but consistent across all nine variables in both lead ranges. For example, T2m RMSE decreases from 1.004 to 0.989 K over +3–30 h and from 1.472 to 1.449 K over +33–72 h. We will also clarify the StormCast configuration and ensemble protocol.

AuthorsOpenReview ↗

Response for Reviewer Vm6f (2/2)

Q4. Coarse-Input Coverage and the Larger-Crop Comparison.

We have included comparative experiments regarding this crop in Figure 17 of Appendix C3.2; here, we provide further explanation and additional comparisons.

Table R4a. Coarse-grid geographic coverage.

Global cropLatitude (degrees N)Longitude (degrees E)
128 x 12833.50 to 65.25-8.00 to 23.75
128 x 20833.50 to 65.25-17.75 to 34.00

Coverage uses the inclusive rectangle bounded by coarse-grid centers.

Table R4b. Crop sensitivity experiment, RMSE, +3-72 h.

ForcingVariable / unit128×208128×128
PanguT2m / K1.3411.370
PanguZ500 / dam1.2201.412
PanguRH2m / pp8.9828.861
ERA5 diagnosticT2m / K1.0011.077
ERA5 diagnosticZ500 / dam0.4860.904
ERA5 diagnosticRH2m / pp7.4657.598

In this table, the larger-crop model reduces Pangu-conditioned T2m RMSE from 1.370 to 1.341 K and Z500 from 1.412 to 1.220 dam, while RH2m increases by 0.121 percentage points.

Dear Reviewer Vm6f,

As the discussion period ends tomorrow, we would appreciate it if you could take a look at our additional experiments and analyses. Please let us know whether these additional details and explanations have addressed your questions.

Thank you for your time and consideration.

Best regards,

The Authors

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

This paper proposes ScaleCast, a kilometer-scale regional weather forecasting framework that integrates pretrained global forecasts with regional atmospheric states through Global–Regional Alignment. The method combines learned token alignment, regional attention, and residual state updates. Experiments on CERRA and HRRR show improvements over several numerical and AI-based forecasting baselines. However, some aspects of the experimental setup, generalization claims, and methodological validation require further clarification.

Critical strengths
  1. The proposed framework provides an approach to regional forecasting using pretrained global weather models without driver-specific retraining.

  2. The paper includes extensive comparisons with numerical and AI-based models, module ablations, and evaluations against station observations and windstorm tracks.

Critical weaknesses
  1. Section 3 trains ScaleCast using ERA5 analyses at the target time, whereas inference relies on global model forecasts. This introduces a training–inference distribution shift whose impact is not evaluated. The temporal alignment of global forecasts with the 3-hour regional prediction intervals also needs clarification.

  2. The HRRR experiment in Figure 6 uses ERA5 analyses at the forecast valid time rather than GraphCast forecasts, despite being labeled "SC+GraphCast." Its comparison with operational HRRR NWP is therefore not information-matched and does not establish operational superiority.

  3. The claimed flexibility across global forecast drivers is not fully supported. Appendix Table 4 shows RMSE increases of 26.7–98.6% when switching from Pangu-Weather to GFS. The authors should clarify the effects of driver distribution shift and temporal interpolation.

  4. GRConv relies on a learned position-specific token mapping rather than explicit geographic coordinates. Its generalization to different grids and domains remains unclear, particularly since the HRRR transfer experiment requires full-model fine-tuning.

  5. The main evaluation excludes wind and precipitation, which are important for kilometer-scale forecasting. Moreover, quantitative comparisons in Table 1 cover only 30 hours, although the framework supports 72-hour forecasts. Longer-lead and more comprehensive variable evaluations are needed.

  6. Table 2 evaluates the proposed modules only under single-step prediction. Their contributions to multi-step forecasting remain unclear. Ablations of GRConv, regional attention, and residual decoding over longer autoregressive rollouts would strengthen the results.

Questions for authors
  1. How sensitive is ScaleCast to the difference between ERA5 analysis forcing during training and global forecast forcing during inference?

  2. Can the authors provide an information-matched HRRR evaluation using actual global forecasts rather than future ERA5 analyses?

  3. How well does GRConv generalize to different geographic domains and grid configurations without full-model fine-tuning?

  4. Do the improvements from the proposed modules persist over 30–72-hour autoregressive forecasts?

  5. How are the different temporal resolutions of global forecast drivers handled in the 3-hour regional prediction framework?

4 comments
AuthorsOpenReview ↗

Response for Reviewer scZT (1/2)

We thank the reviewer for the constructive comments and suggestions. We have added analyses of global forecast conditioning and geographic transfer, extended the evaluation to 72 hours, and supplemented the wind and multi-step component comparisons. We will incorporate these results, clarify the experimental setup, and discuss the identified limitations in the revised manuscript.

W1 & Q1. Quantifying Analysis-to-Forecast Forcing Shift.

We quantified the forcing shift while fixing the regional checkpoint, initial states and 361 evaluation starts.

Table R3. Forcing sensitivity and same-source bilinear references, T2m RMSE (K), +3-72 h.

ForcingScaleCastBilinear
Pangu forecast1.3411.756
GraphCast forecast1.4191.685
ERA5 analysis1.0011.446

Over +3–72 h, T2m RMSE rises from 1.001 K with valid-time ERA5 analyses to 1.341 and 1.419K with Pangu, GraphCast. Thus, regional refinement improves T2m for each tested forecast driver, while a substantial analysis-to-forecast gap remains.

W1 & Q5. How are the different temporal resolutions of global forecast drivers handled in the 3-hour regional prediction framework?

HRES provides hourly forecast outputs, and Pangu supports one-hour inference. GraphCast produces forecasts at six-hour intervals, with intermediate forecast fields obtained through linear temporal interpolation between adjacent outputs from the same initialization cycle.

G(t0+6k+3)=12[G(t0+6k)+G(t0+6k+6)].G(t_0+6k+3)=\tfrac{1}{2}\left[G(t_0+6k)+G(t_0+6k+6)\right].

W2 & Q2. Clarifying the HRRR Setup and Evaluating Forecast-Based Conditioning.

Thank you for highlighting this ambiguity. The reference to valid-time ERA5 analyses in the appendix setup was a writing error, which we will correct.

The tables below compare both conditioning sources with HRRR NWP. We additionally report SC+ERA5 as an analysis-conditioned diagnostic. SC+ERA5 achieves lower T2m and MSLMA RMSE than SC+GraphCast in all reported settings.

**T2m RMSE (K) **

Region / leadSC+GraphCastSC+ERA5 diagnosticHRRR NWP
East +3 h0.8010.7860.813
East +3–48 h1.8131.6141.926
West +3 h1.6491.5851.726
West +3–48 h2.2232.0712.317

**MSLMA RMSE (hPa) **

Region / leadSC+GraphCastSC+ERA5 diagnosticHRRR NWP
East +3 h0.7010.6940.739
East +3–48 h1.1271.0961.157
West +3 h0.8320.8190.829
West +3–48 h1.2451.1791.284

W3. Driver Distribution Shift and Limitations of GFS Transfer.

Thank you for pointing this out. We agree that the 26.7–98.6% RMSE increases in Appendix Table 4 reveal a limitation of directly switching a Pangu-trained model to GFS. The model does not fully adapt to the change in driver distribution. Temporal interpolation aligns the forecast times, but differences in forecast errors and field statistics remain.

GFS-driven refinement by variable, +3–72 h for CERRA.

Variable / unitScaleCast + GFSBilinear GFS
T2m / K2.1022.309
Z500 / dam1.3591.387

These improvements demonstrate useful refinement of GFS forecasts, while the degradation relative to Pangu shows that performance depends on the driver. We will qualify the flexibility claim and explicitly acknowledge direct transfer to an unseen driver as a limitation.

W4 & Q3. Geographic Generalization of GRConv.

Thank you for raising this concern. To evaluate explicit geographic alternatives, we added coordinate-aware cross-attention and a fixed four-neighbor geographic map on HRRR. We compare GRConv with coordinate attention and a fixed geographic mapping on CERRA and after transfer to HRRR over +3–72 h.

Table R2. Mean RMSE over +3–72 h under GraphCast conditioning on CERRA and after transfer to HRRR. Lower is better.

Evaluation regionMethodMSLMA (hPa)T2m (K)SH2m (g/kg)T500 (K)T850 (K)Z500 (dam)Z850 (dam)
CERRAGRConv1.2362.0960.9531.1321.4272.2571.392
CERRACoordinate attention1.3412.1571.1711.2531.5182.5091.520
CERRAFixed geographic1.3882.2361.2121.4241.7942.7831.649
HRRR (transfer)GRConv1.3542.2970.9621.2201.4572.3691.964
HRRR (transfer)Coordinate attention1.4952.4161.1711.3131.6412.4362.182
HRRR (transfer)Fixed geographic1.6912.7421.0521.6981.9372.6392.211

These results support the effectiveness of the learned alignment in both the source region and the transfer setting.

AuthorsOpenReview ↗

Response for Reviewer scZT (2/2)

W5 & Q4. Additional Variables and Evaluation to 72 Hours.

Thank you for this suggestion. We have extended the evaluation to +72 h and added 10-m wind components and wind speed on HRRR under actual GraphCast forecast conditioning. Publicly available CERRA forecasts extend to +30 h, so CERRA NWP results are unavailable for this interval.

Table R1. Mean RMSE over +33–72 h . Lower is better.

MethodMSLPT2mRH2mT500T850Z500Z850RH500RH850
ECMWF-HRES183.21.465N/A1.0750.9971.5571.330N/AN/A
CERRA NWPN/AN/AN/AN/AN/AN/AN/AN/AN/A
Pangu190.61.801N/A1.0841.0721.4711.378N/AN/A
GraphCast140.11.683N/A0.9961.0731.4451.016N/AN/A
ML-LAM137.21.5569.7500.9501.0521.4441.04220.31013.674
Hi-LAM134.91.4939.8400.9671.0481.4051.03720.17513.517
Graph-EFM135.81.6819.8230.9671.1281.4041.04020.17313.841
OneForecast158.32.37710.6470.9571.1611.5741.17720.44014.163
StormCast133.61.4499.9230.9631.0561.4321.02320.15713.884
SC + Pangu134.91.4789.4110.9710.9501.3381.00820.24313.946
SC + GraphCast115.01.3808.7610.8650.9001.1990.84818.32012.679
SC + HRES113.31.4499.1770.8390.9371.1850.85719.63612.305

Table R2. Mean 10-m wind RMSE (m/s) over +3–72 h. Lower is better.

RegionMethodU10V10Wind speed
EastSC + GraphCast1.3671.4631.398
EastBilinear GraphCast1.8712.1031.861
WestSC + GraphCast2.2222.1492.349
WestBilinear GraphCast2.8022.7942.717

ScaleCast retains improvements over its corresponding global driver across all available variables.

Precipitation is available in CERRA as a forecast accumulation, while the CERRA analysis fields used as verification targets in this study do not include precipitation. So, we are unable to conduct a precipitation evaluation.

W6. Contributions of the Proposed Components to Multi-Step Forecasting.

Thank you for this suggestion. We extended the component comparisons to autoregressive forecasts through +72 h.

Table R9. Archived module comparisons over +3-72 h.

(a)T2m RMSE (K)

Variant+3 h+6 h+12 h+18 h+24 h+30 h+36 h+48 h+60 h+72 hMean
Ours1.1870.8581.1341.1121.1091.1601.3591.2911.5691.5031.2282
Gated attention1.1740.8531.1391.1321.1221.1831.3861.3131.5981.5311.2431
Differential attention1.1760.8451.1381.1141.1121.1701.3881.3071.6161.5321.2398
Standard attention1.0800.8761.1801.1441.1991.2281.4061.3811.6091.5731.2676
Fixed geographic alignment1.3220.8451.1051.1011.1061.1531.3491.3051.5801.5521.2418

(b)Z500 RMSE (dam)

Variant+3 h+6 h+12 h+18 h+24 h+30 h+36 h+48 h+60 h+72 hMean
Ours0.5780.3740.4310.5510.6090.7430.8551.2101.6202.1570.9128
Gated attention0.5810.4560.5120.6150.6900.8040.9001.2501.6392.1700.9617
Differential attention0.5850.4290.4790.5960.6750.8020.8971.2551.6502.1860.9554
Standard attention0.5560.5730.6980.7810.8460.9551.0711.3801.7862.2661.0912
Fixed geographic alignment0.5790.4130.4820.5910.6700.7680.8871.2471.6412.1680.9446

Dear Reviewer scZT,

As the discussion period ends tomorrow, we would appreciate it if you could take a look at our additional experiments and analyses. Please let us know whether these additional details and explanations have addressed your questions.

Thank you for your time and consideration.

Best regards,

The Authors

As the discussion period ends in 10 hours, we would appreciate your feedback on whether our responses have addressed your concerns. If so, we hope you will reconsider your rating.

Thank you for your time and valuable feedback, which have helped improve our paper.

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

This paper proposes ScaleCast, a kilometer-scale regional forecasting framework Global–Regional Alignment. Its Global–Regional Conversion module aligns joint global and regional representations with regional locations. The Global–Regional Alignment and Dynamics block combines aligned guidance with regional neighborhood interactions. A core component is the residual prediction, which predicts the weather after 3 hours using the persistence forecast as the baseline, and predict the residual using the forecast generated by global models.

Experiments on ERA5 global dataset and CERRA regional dataset demonstrate the good performance of ScaleCast in the next 30 hours.

Critical strengths
  1. General and practically useful framework. ScaleCast is a general framework which integrates global model for accurate kilometer-scale regional weather forecasting. Also, it can work with different global forecasting models such as HRES, Pangu and GraphCast.
  2. The forecasting performance of ScaleCast is competitive in comparison with current NWP and AI-based global models. In particular, as show in Table 1, ScaleCast with GraphCast outperforms other baselines in most variables.
  3. The alation study in Table 2 shows clearly which components are useful.
Critical weaknesses
  1. Limited evaluation for only 30 hours. Please extend the empirical studies from 30 hours to 72 hours. Although ScaleCast can predict the weather in the next 72 hours, in most settings only empirical results in the next 30 hours are provided, e.g., Table 1. I would like to the full comparison in the next 72 hours.
  2. Limited novelty and potential generalization issue with GRConv. GRConv is fundamentally a learned linear projection across spatial tokens. In addition, GRConv learns the weights for fixed tokens, which may limit its generality and applicability across different geographic regions. It is highly recommended to compare GRConv with coordinate-aware corss-attention, sparse geographic mapping to test its geographic transfer to different regions, as the HRRR experiment requires fine-tuning and therefore does not establish zero-shot geographic transfer.
Questions for authors
  1. Forecasting horizon: please provide the detailed empirical comparison in the next 72 hours in Table 1.
4 comments
AuthorsOpenReview ↗

Response for Reviewer HQvX

We thank the reviewer for the constructive comments and suggestions. We have extended the evaluation to 72 hours and added comparisons with geographic alignment alternatives. We will incorporate these results and the corresponding clarifications into the revised manuscript.

W1 & Q1. Extending Evaluation to 72 Hours.

Thank you for this helpful suggestion. We have extended the evaluation to 72 hours. The table below reports the mean RMSE across the 14 three-hour forecast leads from +33 to +72 h. Publicly available CERRA forecasts extend to +30 h, so CERRA NWP results are unavailable for this interval.

Table R1. Mean RMSE over +33–72 h. Lower is better.

MethodMSLPT2mRH2mT500T850Z500Z850RH500RH850
ECMWF-HRES183.21.465N/A1.0750.9971.5571.330N/AN/A
CERRA NWPN/AN/AN/AN/AN/AN/AN/AN/AN/A
Pangu190.61.801N/A1.0841.0721.4711.378N/AN/A
GraphCast140.11.683N/A0.9961.0731.4451.016N/AN/A
ML-LAM137.21.5569.7500.9501.0521.4441.04220.31013.674
Hi-LAM134.91.4939.8400.9671.0481.4051.03720.17513.517
Graph-EFM135.81.6819.8230.9671.1281.4041.04020.17313.841
OneForecast158.32.37710.6470.9571.1611.5741.17720.44014.163
StormCast133.61.4499.9230.9631.0561.4321.02320.15713.884
SC + Pangu134.91.4789.4110.9710.9501.3381.00820.24313.946
SC + GraphCast115.01.3808.7610.8650.9001.1990.84818.32012.679
SC + HRES113.31.4499.1770.8390.9371.1850.85719.63612.305

W2. GRConv and Geographic Alignment Alternatives.

To test geographic alternatives, we added coordinate-aware cross-attention and a fixed four-neighbor geographic map on HRRR. We compare GRConv with coordinate attention and a fixed geographic mapping on CERRA and after transfer to HRRR over +3–72 h.

Table R2. Mean RMSE over +3–72 h under GraphCast conditioning on CERRA and after transfer to HRRR. Lower is better.

Evaluation regionMethodMSLMA (hPa)T2m (K)SH2m (g/kg)T500 (K)T850 (K)Z500 (dam)Z850 (dam)
CERRAGRConv1.2362.0960.9531.1321.4272.2571.392
CERRACoordinate attention1.3412.1571.1711.2531.5182.5091.520
CERRAFixed geographic1.3882.2361.2121.4241.7942.7831.649
HRRR (transfer)GRConv1.3542.2970.9621.2201.4572.3691.964
HRRR (transfer)Coordinate attention1.4952.4161.1711.3131.6412.4362.182
HRRR (transfer)Fixed geographic1.6912.7421.0521.6981.9372.6392.211

These results support the effectiveness of the learned alignment in both the source region and the transfer setting.

Dear Reviewer HQvX,

As the discussion period ends tomorrow, we would appreciate it if you could take a look at our additional experiments and analyses. Please let us know whether these additional details and explanations have addressed your questions.

Thank you for your time and consideration.

Best regards,

The Authors

Thanks for providing additional experiments. The comparison from 33 to 72 hours looks SC can achieve competitive performance. Since I already give a positive score, I would like to keep it.

Thank you for your follow-up and for your positive assessment of our work. We’re glad that our response addressed your concerns. We have included the additional experiments from the rebuttal in the final version of the paper. We appreciate the time and care you put into reviewing our work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of paper revision

Dear Reviewers,

We sincerely thank all reviewers (HQvX, scZT, and Vm6f) for their valuable feedback. Following your suggestions, we have added experiments and clarified the evaluation settings and methodological distinctions in the revised manuscript.

  • Extended forecast evaluation. Added quantitative comparisons over +33–72 h across nine variables in Table 1, Section 4.2, extending the main evaluation through 72 h. (HQvX, scZT, Vm6f)

  • Geographic transfer and multi-step comparisons. Added comparisons of GRConv with coordinate-aware cross-attention and fixed geographic mapping on CERRA and in zero-shot transfer to HRRR. Added T2m and Z500 comparisons at ten autoregressive lead times through 72 h in Section 4.4. (HQvX, scZT, Vm6f)

  • Forcing sensitivity and HRRR evaluation. Quantified the analysis-to-forecast forcing gap with the regional checkpoint, initial states, and evaluation starts fixed in Appendix C.1.2. Added GraphCast-conditioned HRRR comparisons and ERA5-conditioned diagnostics in Table 7, Appendix C.3.3. (scZT)

  • Input alignment and coverage. Clarified temporal alignment and precipitation verification-data availability in Appendix A.1. Added geographic coverage and larger-crop comparisons in Figure 17 and Table 6, Appendix C.3.2. (scZT, Vm6f)

  • Related-work discussion. Added a comparison with ScaleMixer in Section 2, clarifying the distinction between its internal global–regional token interaction and ScaleCast’s use of external forecast fields. (Vm6f)

The revised material is highlighted in blue for ease of reference. We hope these additions and clarifications address your concerns and would appreciate your reconsideration of the evaluation in light of the additional evidence.

Thank you again for your time and constructive suggestions.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.