acceptodds
Under review as a conference paper at ICLR 2027

WeatherBridge: Forecast-Aware Multi-Source Input Adaptation for Pretrained Weather Forecasting Models

Abstract

Data-driven weather forecasting models are commonly trained on ERA5, yet operational deployment must use timely initial states from physics-based numerical weather prediction (NWP) systems whose distributions differ substantially from ERA5. Direct initialization with these states can degrade autoregressive forecasts, motivating input adaptation that keeps the pretrained forecasting model unchanged. Here, we introduce WeatherBridge, a unified, source-conditioned adapter that maps heterogeneous operational NWP initial states to ERA5 reference states without updating the pretrained forecasting model. WeatherBridge predicts the structured residual corrections with a Swin-UNet and injects source embeddings across multiple spatial scales, allowing the model to capture shared atmospheric structure and source-specific discrepancies. Its training objective combines latitude-weighted ERA5 reconstruction with One-step Rollout Alignment (ORA), which backpropagates one-step forecast error through a frozen FuXi model to optimize the adapted states for both initial state fidelity and downstream forecast compatibility. We evaluate WeatherBridge using daily GFS, HRES, and CMA initializations from 1 January to 22 November 2025 across six variables and lead times up to 240 hours. Relative to raw source initialization of FuXi, WeatherBridge reduces the mean RMSE of FuXi over the 6- 240-hour rollout for 17 of 18 source-variable combinations, with a median reduction of 10.2% and a maximum reduction of 46.3%. The adapter trained with frozen FuXi also improves FengWu forecasts on GFS inputs without retraining either the adapter or FengWu. These results demonstrate that forecast-aware input adaptation can improve forecasts from heterogeneous operational NWP inputs and transfer across pretrained ERA5-based forecasting models.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Domain Adaption, Weather Forecasting, Multi-source Input, Deep Learning

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

This paper proposes WeatherBridge, a pre-processing module to transform operational NWP initial states into ERA5 reference states. Experiments show that WeatherBridge can significantly improve the RMSE and ACC of FuXi forecasts initialized with GFS and CMA data, and can be transferred to the FengWu model without further training, bringing improvements.

Critical strengths

WeatherBridge shows strong performance in improving the FuXi model’s forecasts with GFS and CMA initial states, which demonstrates that this work has practical value.

Critical weaknesses

My main concern is the novelty. Even within AI-based weather forecasting, a very similar approach has already been used in FuXi-DA [1]. In that work, the baseline EXP_CORR is described as follows: ``a DL model is employed to correct the error between the background field and ERA5, yielding the corrected background field.'' The only major difference from WeatherBridge appears to be whether the background field is a previous forecast result or an initial state from a different NWP source. In addition, FuXi-DA uses the training loss L10+1T∑t=1TL1tL_1^0 + \frac{1}{T}\sum_{t=1}^{T}L_1^t, and the FuXi model remains frozen during training. The ORA loss in this paper appears to correspond to replacing L1L_1 with RMSE and setting T=1T=1. Even though I acknowledge the value of the source conditioning mechanism, I do not think the contributions of this paper meet ICLR's requirements.

[1] Xu, X., Sun, X., Han, W. et al. FuXi-DA: a generalized deep learning data assimilation framework for assimilating satellite observations. npj Clim Atmos Sci 8, 156 (2025).

Questions for authors
  1. Swin-UNet is a common CV backbone. Have authors made any specific design changes to adapt it from CV to AI-based weather forecasting?
  2. Why do the validation and test sets not cover a full year of data? Weather patterns should differ to some extent at different times of the year, and not covering a full year may make the evaluation insufficient.
  3. In the FuXi forecasting experiments, using raw GFS initial states gives worse results than using raw CMA data. However, why does GFS WeatherBridge perform slightly better than CMA WeatherBridge after preprocessing?
  4. In the second row of Figure 3, there seems to be a light blue line in plots (a relatively clear example is at (40,0.75)(40, 0.75) in the T850 RMSE plot). What data does this line represent?

I acknowledge that this work has practical value, but frankly, my current position is between clear reject and weak reject. I may adjust my score based on the authors' response.

Reviewer_mpi1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes WeatherBridge, a shared, source-conditioned input adapter that maps operational NWP fields from GFS, HRES, and CMA into an ERA5-aligned representation while keeping the pretrained forecasting model frozen. It uses a residual Swin-UNet with multiscale source and temporal conditioning, trained through a combination of latitude-weighted reconstruction loss and one-step forecast alignment through frozen FuXi. Evaluation covers 326 daily initializations in 2025, six meteorological variables, and forecast horizons up to 240 hours. The authors report reduced mean rollout RMSE in 17 of 18 source–variable combinations, with the largest gains on GFS and limited changes on HRES. Additional experiments examine component ablations, source imbalance, and scaling. The FuXi-trained adapter also improves FengWu forecasts on GFS inputs without retraining, providing evidence of transfer across forecasting backbones.

Critical strengths

1.An effective solution to a practical deployment challenge, supported by substantial evaluation. The distribution mismatch between ERA5-trained models and NWP initial fields is practically relevant. WeatherBridge addresses this mismatch using a shared adapter for three data sources while keeping the forecasting backbone frozen. Experiments covering 326 test initializations, six variables, and forecast horizons up to 240 hours demonstrate substantial reductions in forecast errors for GFS and CMA compared with raw inputs. Statistical significance tests and source-imbalance experiments in the appendix further strengthen confidence in the results. 2.The transfer results across forecasting models are valuable. The adapter trained with FuXi improves FengWu forecasts initialized with GFS inputs without retraining. This suggests that the learned corrections have some transferability across forecasting backbones, extending beyond the FuXi model used during training. However, the current evidence is limited to one additional model and one input source.

Critical weaknesses

1.The relationship to FuXi-DA and the additional contributions require further clarification.FuXi-DA already uses a front-end network to adjust atmospheric states, combining ERA5 state reconstruction with forecast-error supervision through a frozen FuXi model to train the front end. This shares some conceptual overlap with the forecast-aware input adaptation proposed here. Although the two approaches differ in their tasks and input settings, the paper does not discuss this relevant precedent. 2.Quantitative comparisons with other data adaptation methods are missing.The experiments primarily compare raw inputs with WeatherBridge-adapted inputs and include ablations of the proposed method. These results support its effectiveness relative to using inputs without adaptation. However, the paper lacks comparisons with alternative methods for adapting heterogeneous data to forecasting models, making its relative advantages difficult to assess. Including representative adaptation baselines would help clarify the method’s practical value.

Questions for authors

1.How are WeatherBridge’s FuXi-compatible outputs converted into FengWu-compatible inputs, particularly regarding humidity variables, channel ordering, and normalization? Do the raw and adapted inputs undergo the same conversion procedure? 2.WeatherBridge adjusts all 70 channels, while the forecast evaluation focuses on six variables and does not report results for humidity or precipitation. Could the authors clarify whether improvements extend to other output variables, and whether any variables exhibit degradation? 3.The authors characterize joint training on data from GFS, HRES, and CMA as showing “no negative interference between sources.” However, some metrics worsen compared with single-source training; for example, CMA’s 6-hour Z500 RMSE increases from 51.99 to 53.85. Could the authors clarify how they balance performance across different data sources and variables when selecting the final adapter jointly trained on all three data sources? 4.FuXi-DA uses a front-end network to adjust atmospheric states, combining ERA5 state reconstruction with forecast-error supervision through a frozen FuXi model. Given the differences in tasks and input settings, what are the key methodological distinctions between WeatherBridge and FuXi-DA, and how do the existing experiments support the additional contributions of this work?

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

The paper addresses a real operational problem: ERA5-trained AI forecast models (FuXi, GraphCast, Pangu) degrade when initialized from operational NWP analyses (GFS, HRES, CMA) whose distributions differ from ERA5. The authors propose WeatherBridge, a 138M-parameter source-conditioned residual Swin-UNet adapter that maps operational fields into an “ERA5-like” state while keeping the 1.56B-parameter forecasting backbone frozen. Training combines latitude-weighted ERA5 reconstruction with a One-step Rollout Alignment (ORA) loss that backpropagates one-step forecast error through the frozen FuXi. On daily 00Z initializations over Jan–Nov 2025, the adapter reduces mean rollout RMSE for 17/18 source-variable pairs (max ~46% for GFS), and the FuXi-trained adapter transfers to FengWu without retraining.

Critical strengths

The paper addresses a practical mismatch between ERA5-trained forecasters and operational NWP initial states. Its frozen-backbone adapter shows substantial FuXi improvements for GFS and CMA, with experiments across three sources and a useful cross-backbone test on FengWu.

Critical weaknesses

Missing baselines make the gains hard to attribute. No simple bias-correction baseline. A per-grid-point/per-channel climatological correction or linear regression from source to ERA5 is nearly free and may capture much of the discrepancy. Without it, it is unclear how much of the 46% GFS gain requires a 138M-parameter Swin-UNet. Train/test mismatch in ORA. ORA trains with the preceding frame from ERA5, [xEt−1,x^Et][x^{E}{t-1}, \hat{x}^{E}{t}], but at inference both frames are adapted source fields, [x^Et−1,x^Et][\hat{x}^{E}{t-1}, \hat{x}^{E}{t}]. The adapted state is thus optimized for a conditioning context that never occurs operationally. Relatedly, the ablation (Fig. 5) shows ORA’s gains are concentrated at short lead times and “generally diminish at longer lead times” — so the conclusion’s claim that the improvement “holds over the whole 6h to 240h rollout” appears to rest mostly on the reconstruction term, not ORA. This should be stated more precisely.

Questions for authors

1.Clarification and evidence: Was ORA ever trained with two adapted source frames, matching inference? If so, how does its full-rollout performance compare with the reported setup? If not, can the authors assess whether this train–inference mismatch affects the claimed ORA benefit? 2.Evidence affecting the assessment: How much of WeatherBridge's full-rollout improvement remains over a simple source-to-ERA5 bias-correction baseline and over reconstruction-only training, evaluated on the same sources and test period?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.