acceptodds
Under review as a conference paper at ICLR 2027

TIDE: Tendency-Informed Dynamic Encoding for Detail-Preserving Global Weather Forecasting

Abstract

Skillful global weather forecasting is essential for agricultural production, energy management, and disaster mitigation. Despite the rapid progress of data-driven forecasting models, two challenges remain insufficiently addressed. First, historical atmospheric states are often compressed through static or weakly conditioned temporal aggregation, limiting the representation of persistent evolution, first-order tendencies, and accelerating changes. Second, repeated spatial downsampling and autoregressive rollout tend to suppress fine-scale structures, resulting in progressively over-smoothed forecasts. To address these challenges, we propose TIDE, a tendency-informed and detail-preserving framework for global weather forecasting. Its State-Tendency-Acceleration Fusion module constructs first- and second-order temporal differences and adaptively integrates them with the latest state across the bottleneck and pathways, therefore better capturing persistent evolution and accelerating atmospheric changes. To preserve spatial structures, TIDE employs a multiresolution analysis and reconstruction architecture that learns temporally evolving detail memories and selectively restores them through gated decoding. Experiments on ERA5 dataset demonstrate that TIDE achieves state-of-the-art performance in global weather forecasting, establishing a new paradigm in this field.

Primary area:
learning on time series and dynamical systems
Keywords:
Weather Forecasting, Spatiotemporal Modeling, Deep Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_CcwWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

Tendency-Informed Dynamic Encoding (TIDE) constructs state, first-order difference, and second-order difference features from three consecutive historical states and combines them at multiple scales. Detail-memory branches and gated decoding help reconstruct spatial information. On ERA5 at 0.5° resolution, TIDE outperforms the retrained baselines on several forecast metrics and frequency-band errors.

Critical strengths

Explicit temporal differences are straightforward to implement and integrate into a multiscale network. Using them in both the main pathway and the detail branches provides a practical way to incorporate historical states. The experiments include overall forecast metrics, frequency-band analysis, and module ablations.

Critical weaknesses

The reproduced GraphCast, Pangu-Weather, and FengWu results raise concerns that these baselines were not sufficiently trained or optimized. I do not find the change from 0.25° to 0.5° resolution a sufficient explanation on its own for the apparent performance degradation.

So, the authors should provide training configurations, convergence results, and checks against public reference results under comparable settings. Otherwise, it is difficult to tell whether TIDE’s advantage comes from the method or from weak baseline implementations.

In addition, the evaluation is missing an important reference: climatology. Following WeatherBench2[1], they should also specify how RMSE and ACC are computed and aggregated.

[1] Rasp S, Hoyer S, Merose A, et al. WeatherBench 2: A benchmark for the next generation of data‐driven global weather models[J]. Journal of Advances in Modeling Earth Systems, 2024, 16(6): e2023MS004019.

Questions for authors
  1. Can the authors provide the full training configurations and convergence curves for GraphCast, Pangu-Weather, and FengWu, together with reference comparisons under compatible settings, to rule out insufficient training or implementation differences? Can the effect of resolution be tested separately?

  2. Can the authors add a climatology baseline on the current 0.5° grid and recheck RMSE and ACC using the WeatherBench2 evaluation procedure, clarifying which variables and lead times retain skill over climatology?

2 comments

W1-2, Q1: Unfair Baseline Comparisons

Due to space limitations, the initial version of our manuscript did not provide complete reproduction configurations for the three baselines. We sincerely apologize for this omission. We have now included the relevant details in the appendix of the revised manuscript and added corresponding explanations in the main text. These details demonstrate that all reproduced baselines were adequately trained and had essentially converged.

Furthermore, the original baseline models were trained on high-resolution datasets with a spatial resolution of 0.25°, allowing them to learn from four times as much spatial information as the dataset used in our study. Consequently, their originally reported forecasting accuracy is naturally higher than that achieved by our reproduced implementations. Unfortunately, due to GPU memory limitations, we are unable to fully train these models at 0.25° resolution and therefore cannot conduct comparative evaluations at this resolution.

W3, Q2: Comparison with Climatology

We have additionally evaluated climatology as a baseline using the 2021 test set and updated the corresponding result figures in the revised manuscript. Given the definition of climatology, the ACC metric is not meaningful for this baseline. Therefore, we report only RMSE values in the figures. The RMSE is calculated as follows.

For variable vv, forecast initialization time ss, and forecast lead time τ\tau, the latitude-weighted spatial RMSE for an individual sample is defined as:

RMSE⁡s,v(τ)=1NlatNlon∑i=1Nlat∑j=1Nlonwi(x^s,τ,v,i,j−xs+τ,v,i,j)2\operatorname{RMSE}_{s,v}(\tau) = \sqrt{ \frac{1}{N_{\mathrm{lat}}N_{\mathrm{lon}}} \sum_{i=1}^{N_{\mathrm{lat}}} \sum_{j=1}^{N_{\mathrm{lon}}} w_i \left( \hat{x}_{s,\tau,v,i,j} - x_{s+\tau,v,i,j} \right)^2 }

where x^\hat{x} denotes the predicted value, xx represents the corresponding ERA5 ground truth at the forecast valid time, and ii and jj denote the latitude and longitude grid indices, respectively. The latitude weights are normalized as follows:

wi=cos⁡ϕi1Nlat∑k=1Nlatcos⁡ϕkw_i = \frac{\cos \phi_i} {\displaystyle \frac{1}{N_{\mathrm{lat}}} \sum_{k=1}^{N_{\mathrm{lat}}}\cos \phi_k}

where ϕi\phi_i represents the ii-th latitude. This normalization ensures that the mean latitude weight is equal to 1.

The RMSE at each forecast lead time is then obtained by taking the arithmetic mean over NτN_\tau forecast initialization samples:

RMSE⁡v(τ)=1Nτ∑s=1NτRMSE⁡s,v(τ)\operatorname{RMSE}_{v}(\tau) = \frac{1}{N_\tau} \sum_{s=1}^{N_\tau} \operatorname{RMSE}_{s,v}(\tau)

Taking Z500 as an example, the results show that TIDE outperforms climatology for forecast lead times of up to 9 days, whereas Pangu-Weather, GraphCast, and FengWu outperform climatology for lead times of up to 6, 7, and 5 days, respectively.

I appreciate the authors' effort. Nevertheless, the remaining limitations continue to affect my overall assessment. I maintain my original score.

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

TIDE predicts global weather from three six-hourly states using state–tendency–acceleration features, multiscale detail branches, gated reconstruction and auxiliary losses. It is trained on ERA5 at 0.5° with 83 channels and compared against retrained Pangu-Weather, FengWu and GraphCast. The manuscript reports improved global RMSE/ACC and frequency-band errors, two typhoon cases and 72-hour component ablations. The intended contribution is explicit temporal evolution modeling together with better retention of spatial detail.

Critical strengths
  1. Oversmoothing and temporal conditioning matter in weather prediction. The temporal split is explicit, and the paper reports substantial gains over three retrained architectures. Table 1 improves 72-hour Z500 RMSE from 201.91 for raw three-frame concatenation to 180.81 for the full model; gains are also reported for temperature variables.
  2. Table 1 tests one/two/three-frame inputs, raw concatenation, removal of detail memory, ungated addition and full-resolution-only supervision. Frequency-resolved errors and held-out-year case studies extend evaluation beyond one global aggregate. These results justify further investigation of the implementation.
Critical weaknesses
  1. STA does not establish the claimed new temporal modeling mechanism. The state, first difference and second difference in Eq. (7) form an invertible linear change of basis of the same three features. Because Eq. (8) begins with an unconstrained linear layer, that transform can be absorbed exactly into its weights. At matched feature locations and MLP structure, STA and raw-feature fusion have the same function class. The observed gap may reflect initialization, conditioning or where fusion occurs; it does not establish additional information or expressivity. Table 1 does not specify that the concatenation control matches every fusion location. Similarly, the detail branch is a learned strided-convolution skip pathway, without an explicit complementary high-pass decomposition or reconstruction constraint. A useful inductive bias is possible, but the present 'new paradigm' interpretation is not earned.
  2. The experiments do not yet validate the main performance and detail-preservation claims. Section 4.1 states that baselines are retrained, but omits their capacities, optimization budgets, normalization, multistep training, convergence and seed variability. Pangu also uses a different prediction interval. Matching data alone cannot establish a fair state-of-the-art comparison, and superiority of published baselines over NWP does not transfer automatically to these retrained versions. Moreover, lower high-frequency error power does not by itself establish retained signal power: attenuation can reduce squared error when phase is uncertain. Figure 3 gives error spectra without forecast/target power or coherence, and two track examples do not establish cyclone-core or intensity fidelity. The empirical evidence therefore supports a promising recipe more strongly than the stated mechanism and broad conclusion.
Questions for authors
  1. Does the raw-three-frame control use the same per-frame encoder, MLP width and fusion locations at every detail level and the bottleneck? Given the exact linear reparameterization, what controlled evidence supports an optimization or inductive-bias advantage across seeds, and what claim remains beyond that advantage?
  2. Please provide the complete training and adaptation configurations for every model, including channel scaling, capacity, update budget, rollout training, validation selection and convergence. Are the large gains preserved against converged, appropriately tuned baselines under the same information and compute regime?
  3. Can forecast/target power spectra and cross-spectral agreement show that the L4–L5 improvement preserves useful signal rather than suppressing it? A comparison to an ordinary gated multiscale skip branch, with the same three-frame input and auxiliary losses, would directly test the proposed detail mechanism.
1 comment

W1, Q1: Validity of the STA Module and Ablation Experiments

First, we acknowledge an error in the description of our method in the original manuscript. Specifically, the temporal differences are actually processed by an activation function (GELU) before being passed to the MLP. We sincerely apologize for the misunderstanding this may have caused the reviewers and have corrected this description in the revised manuscript.

Second, the number of input frames in our ablation experiments is controlled simultaneously across all spatial scales, rather than only at the input level.

Third, compared with our proposed method, the raw three-frame concatenation variant removes only the temporal difference operations and the subsequent GELU activations at all spatial scales, while keeping all other components unchanged. This design ensures that the variables are controlled as rigorously as possible for a fair comparison.

Finally, our characterization of the proposed method as a "new paradigm" is motivated by the following considerations. Most existing weather forecasting methods place little emphasis on explicitly exploiting temporal patterns in historical weather data. However, our results demonstrate that even simple temporal difference operations can effectively improve forecasting accuracy without introducing substantial computational overhead. Moreover, the proposed approach is lightweight and modular. Through this work, we hope to encourage future weather forecasting research to place greater emphasis on temporal feature modeling.

W2, Q2: Unfair Baseline Comparisons

Pangu-Weather provides different model variants with 6-hour and 24-hour forecasting intervals. According to Extended Data Fig. 7 of the original Nature paper [1], the 24-hour variant achieves higher forecasting accuracy than the 6-hour variant. We therefore selected the 24-hour variant to ensure the strongest possible baseline performance.

Due to space limitations, the initial version of our manuscript did not provide complete reproduction configurations for the three baselines. We sincerely apologize for this omission. We have now included the relevant details in the appendix of the revised manuscript and added corresponding explanations in the main text. These details demonstrate that all reproduced baselines were adequately trained and had essentially converged.

W2, Q3: Band Power Ratio and Spectral Correlation

We agree that evaluating high-frequency RMSE alone is insufficient to demonstrate the effective preservation of fine-grained spatial details. Following your suggestion, we additionally computed the band power ratio and spectral correlation for Z500, T850, and T2M, comparing TIDE with its variant without the adaptive gating mechanism. The results are presented in the following table.

VariableMethodL4 Band Power RatioL4 Spectral CorrelationL5 Band Power RatioL5 Spectral Correlation
Z500TIDE1.190.112.680.14
Z500Without adaptive gate1.730.084.500.04
T850TIDE0.540.280.820.31
T850Without adaptive gate0.380.250.490.27
T2MTIDE0.700.740.720.71
T2MWithout adaptive gate0.560.680.560.65

The results demonstrate that our proposed method preserves fine-grained spatial details in weather fields more effectively than the variant without adaptive gating, as evidenced by band power ratios closer to unity and consistently higher spectral correlations.

Regarding typhoon forecasting, both TIDE and all baseline models are deterministic forecasting models. Therefore, they can only produce deterministic typhoon trajectories rather than probabilistic forecasts. In future work, we plan to investigate probabilistic forecasting approaches to further enhance the practical applicability of our method.

Reference

[1] Bi, K., Xie, L., Zhang, H. et al. Accurate medium-range global weather forecasting with 3D neural networks. Nature 619, 533–538 (2023). https://doi.org/10.1038/s41586-023-06185-3

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

This paper proposes TIDE for global weather forecasting. It combines first- and second-order temporal differences through State-Tendency-Acceleration fusion with multiscale detail memories and gated reconstruction. Experiments on ERA5 report improvements over several weather forecasting models, supported by frequency-band analysis, typhoon case studies, and ablations.

Critical strengths

The paper addresses temporal evolution and spatial detail loss in weather forecasting. The motivation is clear.

The experiments cover multiple weather variables, forecast horizons, frequency bands, and ablations, providing useful evidence for the proposed design.

Critical weaknesses

First- and second-order temporal differences are standard operations, while detail memories and gated reconstruction are closely related to existing multiscale architectures.

The STA module essentially reparameterizes three historical frames without introducing additional information. The comparison with raw concatenation is insufficient to establish its advantage over more advanced temporal modeling methods.

Pangu-Weather uses a different prediction interval, and all baselines are retrained at reduced resolution. The effects of these differences are not sufficiently discussed.

The experimental protocol needs clarification. The main test set covers 2021, but the paper also evaluates a typhoon from 2026 without clearly explaining the corresponding data source and evaluation setup.

The evaluation of extreme weather remains limited to two typhoon cases.

Questions for authors

Clarify the novelty of STA fusion and detail-preserving reconstruction.

Explain the fairness of baseline configurations and prediction intervals.

Clarify the data source and evaluation protocol for the 2026 typhoon.

Provide evidence for the stability of the reported improvements.

2 comments

W1, Q1: Limited Novelty of the STA and Detail-Preserving Modules

Unlike pure time-series forecasting, weather forecasting involves high-resolution spatial information at each time step, resulting in substantial computational costs. Consequently, existing weather forecasting methods have not fully exploited temporal dependencies in weather data. To the best of our knowledge, TIDE is the first weather forecasting method to attempt explicit temporal dependency modeling.

Regarding the detail-preserving module, existing weather forecasting methods typically compress high-resolution weather fields into a low-resolution bottleneck through repeated spatial downsampling, leading to oversmoothing and the loss of fine-grained spatial details. In contrast, our proposed method is the first weather forecasting approach to explicitly model and preserve spatial details.

W2: Insufficient Evidence for the Effectiveness of the STA Module

As discussed above, the computational demands of high-resolution weather data prevent us from employing excessively complex temporal modeling techniques. The proposed STA module, based on multi-order temporal differences, introduces only modest additional computational overhead while effectively improving forecasting accuracy compared with directly feeding three consecutive frames into the model without additional processing.

W3, Q2: Unfair Baseline Comparisons

Pangu-Weather provides different model variants with 6-hour and 24-hour forecasting intervals. According to Extended Data Fig. 7 of the original Nature paper [1], the 24-hour variant achieves higher forecasting accuracy than the 6-hour variant. We therefore selected the 24-hour variant to ensure the strongest possible baseline performance. The other baseline methods provide only models with 6-hour intervals, so this issue does not apply to them.

Due to space limitations, the initial version of our manuscript did not provide complete reproduction configurations for the three baselines. We sincerely apologize for this omission. We have now included the relevant details in the appendix of the revised manuscript and added corresponding explanations in the main text. These details demonstrate that all reproduced baselines were adequately trained and had essentially converged.

W4, Q3: Source of the 2026 Data

The statement that the test set uses WeatherBench2 data from 2021 refers specifically to the line plots presenting statistical results and the tables reporting ablation experiments, all of which are based on 2021 data. Since Typhoon In-Fa occurred in 2021, its case study also uses this dataset.

Typhoon Bavi occurred in 2026, which falls outside the temporal coverage of WeatherBench2. Therefore, the data for this case study were obtained directly from cds.climate.copernicus.eu. Although the two datasets were downloaded from different sources, they are both derived from the same underlying ERA5 dataset.

W5, Q4: Insufficient Case Studies of Extreme Weather Events

Typhoons are representative weather systems characterized by rapid changes in atmospheric conditions within localized regions. Their low-pressure centers, strong wind circulations, and precipitation distributions exhibit pronounced spatial gradients and localized structures. Therefore, typhoons are particularly suitable for evaluating the forecasting capabilities of models under extreme weather conditions. In fact, both Pangu-Weather and GraphCast selected typhoon forecasting for their case studies.

Regarding the selection of specific cases, we chose one representative typhoon event from the 2021 test set and another more recent event from 2026. Both cases are highly representative and provide meaningful evaluations of the model's forecasting capabilities.

Reference

[1] Bi, K., Xie, L., Zhang, H. et al. Accurate medium-range global weather forecasting with 3D neural networks. Nature 619, 533–538 (2023). https://doi.org/10.1038/s41586-023-06185-3

Thanks for clarifying the source of the typhoon data and the choice of the 24-hour Pangu model. That answers my questions about the setup. The comparison with raw frame concatenation still leaves me unsure how much STA adds over another lightweight way of combining the same three frames. I’m keeping my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.