acceptodds
Under review as a conference paper at ICLR 2027

From Discrete Full-Field Labels to Any-Time Weather Forecasting

Abstract

Full-field labels for data-driven weather forecasting are typically available only at discrete lead times, which ties the learned prediction mappings to predefined forecast intervals, while many weather processes and downstream applications require weather states at arbitrary times. To enable any-time prediction under this discrete supervision, we propose Physics-Supervised Any-Time Forecasting (PSAF). PSAF recomposes internal representation changes formed during fixed-endpoint forecasting into latent states for requested times, and maps them to weather fields through local decoding aligned with the forecast semantics of neighboring supervised lead times. To supervise intermediate times without labels, we use diagnostic-space physical surrogates as training targets. On WeatherBench 2, PSAF reduces the mean RMSE across 1–6 h lead times by 39.7% relative to endpoint field interpolation and by 10.1% relative to the strongest learned endpoint-reconstruction baseline. At intermediate lead times, PSAF outperforms the compared continuous-time methods on independent RTMA-RU analyses and downstream wind-power forecasting. Physics Surrogate Supervision further reduces the geostrophic and kinetic-energy target errors by 25.6% and 7.2%, while also improving physical diagnostics not included in the training objective.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
weather forecasting, any-time forecasting, arbitrary lead time, physics-informed machine learning, spatiotemporal modeling

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_JxrAOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Physics-Supervised Any-Time Forecasting (PSAF) for predicting weather fields at fractional lead times using full-field labels available only at integer hours. The method recomposes the latent changes produced by a frozen 6-hour forecasting model and uses a local mixture of neighboring integer-time decoders to generate the requested field. Fractional queries are trained using interpolated geostrophic and kinetic-energy targets together with continuity regularization. The paper evaluates integer-time forecasting, fractional-time accuracy against RTMA-RU, physical diagnostics, spatial structure, and downstream wind-power forecasting.

Critical strengths
  1. The paper addresses a useful problem, since many downstream weather applications require forecasts at time intervals that do not match the model’s native forecast step.
  2. The method is clearly described and supported by a broad set of internal analyses, including component ablations, physical diagnostics, boundary continuity, spatial statistics, and a downstream wind-power task.
Critical weaknesses
  1. The fractional-time comparison is not well controlled because ClimODE and CEF use WeatherBench v1 inputs at 5.625-degree resolution, while PSAF is trained on WeatherBench 2 at 1.5-degree resolution and is only downsampled during evaluation.
  2. The evidence for arbitrary-time forecasting is limited because the real fractional-time evaluation only tests hourly midpoints, while the stronger and more comparable endpoint-reconstruction baselines are evaluated only at integer lead times.
Questions for authors
  1. Can ModAFNO, AMT, UPR-Net, and field interpolation be evaluated on the same fractional RTMA-RU queries?

  2. Since RTMA-RU is available every 15 minutes, can the authors report results at 0.25, 0.5, and 0.75 within each hourly interval?

  3. Can ClimODE, CEF, and PSAF be compared using matched training data and input resolution?

  4. How much of the Local Decoder MoE improvement remains when compared with a parameter-matched time-conditioned shared decoder?

  5. How does the method perform during rapidly evolving weather events, where linear interpolation of the physical diagnostics may be inaccurate?

[Additional Comments] The recursive composition result also shows a noticeable gap between a direct 1-hour query and two successive 0.5-hour queries, suggesting that the learned forecasts are not yet temporally composition-consistent.

7 comments
AuthorsOpenReview ↗

Response to Reviewer JxrA

Thank you for the constructive comments. We provide additional experiments and analyses addressing fractional-time accuracy, baseline comparability, decoder design, and rapid changes.

1. Learned baselines and dense queries (W2/Q1/Q2). We retain the paper's RTMA-RU setup and extend queries to 0.25, 0.50, and 0.75 h within each hour. Metrics and units follow the paper.

MethodT2mU10V10Wind vector
PSAF1.30921.11041.11691.7496
6h field interpolation1.41971.12351.14511.7821
ModAFNO1.37401.12141.12871.7673
AMT1.35291.12381.13261.7722
UPR-Net1.36541.13271.14161.7864

Grouped by query position within each hour, T2m / wind vector MAEs are:

Method0.25h0.50h0.75h
PSAF1.3135 / 1.76511.3023 / 1.73581.3118 / 1.7477
6h field interpolation1.4335 / 1.80061.4083 / 1.76531.4175 / 1.7805
ModAFNO1.3833 / 1.79021.3561 / 1.75171.3825 / 1.7600
AMT1.3640 / 1.78941.3436 / 1.75821.3512 / 1.7689
UPR-Net1.3783 / 1.80221.3573 / 1.77241.3606 / 1.7845

PSAF has lower errors on all four overall metrics and both reported metrics at each query position, extending the evidence beyond hourly midpoints.

2. Training data and input resolution (W1/Q3). WeatherBench v1 and WeatherBench 2 both derive from ERA5. Direct any-time forecasting of global multivariate weather fields remains relatively underexplored at finer spatial resolutions, and PSAF demonstrates this capability at 1.5-degree resolution, complementing existing research on continuous-time weather forecasting.

Adapting ClimODE and CEF to this resolution and retraining them under matched conditions requires additional time and computational resources. We apologize that, given the time and resources available, we have not yet completed this comparison.

The added ModAFNO, AMT and UPR-Net comparisons use the same WB2 data source, 1.5-degree grid and 2006–2015 integer training labels as PSAF Stage 2, and share the same 6h forecast endpoint. These comparisons further evaluate PSAF's predictive performance under the same data and resolution settings.

3. Local Decoder MoE versus a shared decoder (Q4). The paper's shared decoder receives time conditioning through the queried latent state. Removing Local Decoder MoE increases 1h RMSE by 10.8% and yields higher errors at 1–5h. The additional fractional RTMA-RU results are:

MethodT2mU10V10Wind vector
PSAF1.30921.11041.11691.7496
w/o Local Decoder MoE1.32471.11281.11951.7534

These results support Local Decoder MoE at both integer and fractional leads. The existing shared decoder does not match the experts' total parameter count. We will prioritize a parameter-matched, time-conditioned shared decoder comparison as compute becomes available, to separate the contributions of local experts and model capacity.

4. Rapid changes and Physics Surrogate Supervision (Q5). Linear interpolation of physical diagnostic targets provides approximate supervision. Full fields are learned jointly through integer full-field supervision, Physics Surrogate Supervision and Continuity Regularization.

We select 15-minute temperature or wind vector changes at or above each grid point's P95 threshold in RTMA truth. PSAF's change-MAE reductions against the four baselines are:

BaselineT2mWind vector
6h field interpolation7.434%1.219%
ModAFNO0.594%0.726%
AMT2.441%0.914%
UPR-Net2.694%0.732%

PSAF has the lowest change MAE for both temperature and wind vector on these rapid-change subsets.

5. Recursive composition consistency (additional comment). PSAF constructs each forecast from the supplied history and query time. A direct 1h query uses the original history, whereas two successive 0.5h steps use generated states to form the second query's history. This changes the conditioning context and carries forward prediction errors, so the recursive procedure does not inherit the composition property of an exact ODE flow. The appendix reports normalized forecast MAEs of 0.01578 and 0.02043, respectively.

We will incorporate these additional results and the corresponding experimental settings and clarifications into the revised manuscript shortly.

AuthorsOpenReview ↗

Manuscript revision update

Thank you again for your constructive feedback. We have updated the manuscript with additional results and clarifications from our response. The main text now includes denser RTMA-RU evaluation, interpolation comparisons on RTMA-RU and wind power, and the controlled 2h-to-1h label-holdout experiment. The appendix includes the RTMA-RU component ablations, AMT with historical input, cross-window consistency, and rapid-change evaluations. We have also clarified the method description and evaluation protocols.

Regarding Q4, the parameter-matched, time-conditioned shared decoder comparison is now underway. We will report the results once training and evaluation are complete.

AuthorsOpenReview ↗

Follow-up on our response

Thank you again for your time and consideration. We would appreciate your feedback on our response and briefly summarize the additional results below.

To address fractional-time evaluation and baseline comparability, we evaluated ModAFNO, AMT, UPR-Net, and field interpolation on the same RTMA-RU queries at 0.25, 0.50, and 0.75 h within each hour. The added learned baselines use the same data source, resolution, and integer training labels as PSAF. PSAF has lower errors than these baselines on all four aggregate metrics (Table 3 of the revised manuscript).

On rapid-change subsets selected from RTMA ground truth, PSAF yields lower temperature-change and wind-vector-change errors than 6h field interpolation and the learned baselines above (Table 17).

Regarding Q4, training of the parameter-matched, time-conditioned shared decoder is ongoing. To match the experts' total parameter count, this control uses a larger shared decoder without MoE routing. Each query runs through the full decoder, increasing computational cost and training time. We hope to complete training before the discussion period ends and will share the results as soon as they are available.

We would welcome your feedback and any remaining questions. If these additions address your concerns, we would be grateful if you could reconsider your assessment.

Reviewer_Fn4cOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PSAF for forecasting at arbitrary times within a 0–6-hour window using Latent Recomposition and a Local Decoder MoE. The overall approach is clear and practically relevant. However, the experimental evidence does not yet fully support the central claim of accurate forecasting at unsupervised fractional lead times.

Critical strengths
  1. The overall framework is clearly presented, combining latent recomposition and local decoding for continuous-time forecasting.
  2. Forecasting at arbitrary intermediate times is a practically relevant problem, and the proposed approach offers a potentially useful solution.
Critical weaknesses
  1. The 39.7% and 10.1% improvements in the abstract are based on average RMSE at integer lead times from 1 to 6 hours, where full ERA5 supervision is available. These results mainly demonstrate performance at supervised integer times rather than at unlabeled fractional times. The abstract should explicitly state "at supervised integer lead times" to avoid overstating the results.

  2. The most important missing comparison is between directly querying PSAF at 2.5 hours and interpolating PSAF's own predictions at 2 and 3 hours. The current 39.7% improvement is measured against interpolation between the 0-hour analysis and 6-hour forecast. This does not isolate the benefit of Latent Recomposition and Local Decoder MoE over interpolation between adjacent prediction anchors.

  3. The method supports continuous queries within a single 0–6-hour window, which should be stated more clearly in the introduction. Although rapidly evolving weather is a central motivation, the experiments do not specifically evaluate precipitation, convection, or extreme storms. The evidence mainly concerns large-scale atmospheric fields and near-surface temperature and wind. Furthermore, the RTMA-RU validation covers only near-surface variables over the United States. Direct fractional-time verification is limited to a few half-hour query points, leaving global, multivariable, and more general continuous-time performance insufficiently established.

  4. In Eq. (8), α_i is not theoretically guaranteed to lie within [0, 1] or vary monotonically with time. The actual range of α_i and the effects of monotonicity constraints, range constraints, or a shared linear schedule should be examined. Also, the latent changes C_i represent differences between network layers rather than physical temporal increments. Although the paper acknowledges that network depth is not equivalent to time, it remains unclear why reweighting these differences should recover realistic intermediate atmospheric states.

  5. Table 6 evaluates Latent Recomposition and Local Decoder MoE mainly using integer-time RMSE, although both modules are motivated by fractional-time prediction without full-field labels. Ablations at fractional lead times, using RTMA-RU or held-out labels, would be more informative. The continuity regularization and physical proxy supervision should also be ablated separately.

  6. Section 3.5 provides only the overall physical proxy and continuity losses, while key definitions appear in the appendix. Equation (17) in the main text and Eq. (28) in the appendix describe the same total continuity loss. The main text should explicitly refer to Appendix A.4 and use consistent notation.

Questions for authors
  1. How does direct PSAF prediction at 2.5 hours compare with linear interpolation between PSAF's own 2-hour and 3-hour predictions? This is the most important baseline for establishing the benefit of the proposed continuous-time modules.

  2. Can the authors provide ablations of Latent Recomposition, Local Decoder MoE, continuity regularization, and physical proxy supervision using RTMA-RU or held-out labels?

  3. What is the observed range of α_i across query times? How does the performance change with monotonicity constraints, range constraints, or a shared linear schedule? What justifies interpreting reweighted inter-layer latent differences as intermediate physical states?

  4. Can the authors evaluate denser unseen query times, consistency across 6-hour window boundaries, and agreement between direct queries and recursive forecast composition?

  5. Can the authors provide direct evaluations of rapidly evolving weather processes, such as precipitation, convection, or extreme storms, to better support the stated motivation?

5 comments
AuthorsOpenReview ↗

Response to Reviewer Fn4c

Thank you for the constructive comments. We provide additional experiments and analyses addressing prediction at unlabeled times, component roles and design choices, temporal consistency and rapid changes.

1. Interpolation and dense queries (W2/Q1/Q4). Using the paper's RTMA-RU protocol, we test 18 leads (0.25/0.50/0.75h per hour, including 2.5h). AMT is Table 2's strongest learned endpoint-reconstruction baseline.

MethodT2mU10V10Wind vector
PSAF1.30921.11041.11691.7496
PSAF hourly interpolation1.30381.10931.11631.7482
6h field interpolation1.41971.12351.14511.7821
AMT1.35291.12381.13261.7722

PSAF has lower MAEs than 6h field interpolation and AMT; its MAEs are 0.05%–0.41% higher than its own hourly interpolation.

Under the paper's wind-power setup, direct queries reduce mean NMAE/NRMSE by 0.26%/0.45%:

Weather inputNMAENRMSE
PSAF direct queries0.093053 ± 0.0037240.138948 ± 0.005960
PSAF hourly interpolation0.093300 ± 0.0038960.139573 ± 0.006874

In both tasks, PSAF hourly interpolation tests the added benefit of direct queries; the other RTMA baselines assess the overall approach.

2. Module roles and fractional ablations (W3/W5/Q2). The modules serve both supervised anchors and unlabeled queries. Table 6 shows gains from Latent Recomposition and Local Decoder MoE at supervised anchors. Continuity analyses assess Continuity Regularization; physical diagnostics and the 2h-to-1h test assess Physics Surrogate Supervision. For Q2, ablations on the same RTMA-RU queries give:

MethodT2mU10V10Wind vector
PSAF1.30921.11041.11691.7496
w/o Latent Recomposition1.29251.11161.11751.7508
w/o Local Decoder MoE1.32471.11281.11951.7534
w/o Continuity Regularization1.30641.10891.11771.7489
w/o Physics Surrogate1.31421.11201.11831.7518

The 2h-to-1h test uses even-hour supervision and evaluates odd-hour fields against ground truth. Global 69-channel RMSEs under the Table 2 test protocol are:

LeadPSAFw/o Physics SurrogatePSAF 2h-anchor interpolation
1h0.0271970.0363680.028967
2h0.0376250.0372750.037625
3h0.0458590.0511450.046817
4h0.0492400.0490960.049240
5h0.0519300.0540160.053229
6h0.0551680.0551680.055168

At 1/3/5h, PSAF reduces mean RMSE by 11.69% versus w/o Physics Surrogate and 3.12% versus its own interpolation. This extends verification beyond RTMA to global full fields at leads withheld from full-field supervision, showing the role of Physics Surrogate Supervision.

3. α and latent interpretation (W4/Q3). With s=τ/6s=\tau/6, αi=s+2s(1−s)tanh⁡di\alpha_i=s+2s(1-s)\tanh d_i fixes both endpoints. Its theoretical range is [-0.125,1.125]; the observed range is [-0.125,1.1249]. Monotonicity and [0,1] bounds are not prerequisites for learned recomposition. A shared linear schedule yields a linear latent path; ours adapts to time, sample, layer and channel.

We use CiC_i as features learned by the frozen 6h forecaster, without assigning physical time to network depth. The controller and Local Decoder MoE learn to map these features and query time to atmospheric fields. Integer full-field labels anchor this mapping to real atmospheric states, while Physics Surrogate Supervision and Continuity Regularization constrain predictions at unlabeled times. We also observe shallow-block contributions strengthening earlier and deep-block contributions later, consistent with our intuition of progressively incorporating learned forecast changes.

4. Cross-window and recursive consistency (Q4). Dense queries are covered above. For the context switch [-6h,0h] → [0h,predicted 6h], cross-window normalized MAEs are:

ComparisonPSAFw/o Continuity Regularization
Same 6h time: old endpoint vs new start reconstruction0.0043140.004565
Boundary sides: 5.99h vs 6.01h0.0043170.007238

Regularization reduces both differences. The appendix reports normalized forecast MAEs of 0.01578 for direct 1h queries and 0.02043 for two 0.5h steps. PSAF does not enforce an exact ODE flow's composition property; recursion rebuilds history from generated states and performs worse here.

5. Rapid changes (W3/Q5). Although PSAF hourly interpolation has slightly lower overall field MAEs, direct queries better capture rapid 15-minute changes. We select changes at or above each grid point's P95 threshold in RTMA truth. PSAF's change-MAE reductions are:

BaselineT2mWind vector
PSAF hourly interpolation0.843%0.164%
6h field interpolation7.434%1.219%
ModAFNO0.594%0.726%
AMT2.441%0.914%
UPR-Net2.694%0.732%

Dedicated precipitation, convection and extreme-storm evaluations remain future work.

6. Revisions (W1/W3/W6). We will shortly add these results to the manuscript and clarify the supervised 1–6h mean RMSE gains, 0–6h window, and Section 3.5/Appendix A.4 references and notation in Eqs. (17)/(28).

AuthorsOpenReview ↗

Manuscript revision update

Thank you again for your constructive feedback. We have updated the manuscript with additional results and clarifications from our response. The main text now includes denser RTMA-RU evaluation, interpolation comparisons on RTMA-RU and wind power, and the controlled 2h-to-1h label-holdout experiment. The appendix includes the RTMA-RU component ablations, AMT with historical input, cross-window consistency, and rapid-change evaluations. We have also clarified the method description and evaluation protocols.

Regarding Q3, we are also training a control that replaces the learned recomposition weights with a shared linear schedule while retaining the other components. We will report the results once training and evaluation are complete.

AuthorsOpenReview ↗

Follow-up on our response

Thank you again for your time. We would like to share the additional results below.

Following up on Q3, we evaluated a control with recomposition weights fixed to αi(τ)=τ/6\alpha_i(\tau)=\tau/6, retaining Local Decoder MoE and the remaining training objectives. We also include w/o Latent Recomposition, which uses the endpoint latent representation zLz_L for all queries, for comparison.

The normalized RMSEs under the paper's test protocol are:

Method1h2h3h4h5h6h
w/o Latent Recomposition0.0271600.0396320.0464040.0496770.0514780.055168
Shared linear schedule0.0263560.0394510.0469790.0505260.0523130.055168
PSAF0.0240570.0363670.0439860.0486040.0513560.055168

Compared with using the endpoint latent representation for all queries, the shared linear schedule yields lower errors at 1–2h but slightly higher errors at 3–5h, while PSAF outperforms both fixed schemes at 1–5h.

We hope these additions further address your concerns. We would be happy to clarify any remaining questions and would welcome your feedback.

Reviewer_oD89OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper freezes the FengWu-style 6h forecaster and adds an any-time head to it. Then, a query latent is computed by reweighting the per-block increments and it is then decoded by hourly decoders D0,…,D6D_0,\dots,D_6. When at fractional time τ=k+δ\tau=k+\delta, only the two neighbor decoders will be mixed. These time are then trained with linearly interpolated geostrophic residual and kinetic energy targets with continuity regularizations, because there are no field labels for them.

Critical strengths
  1. The additional head added is light-weight and does not touch the 6h forecast.
  2. Beyond ERA5, this work also evaluates 0.5h leads against indepentent RTMA-RU analyses and a downstream wind-power task. This looks good as a practical check.
  3. The results reported are transparent. The physics ablation is split into fractional and integer queries. The fixed-1h is also included though it beats the proposed method at 1-3h.
Critical weaknesses

My main concern is that I wonder if the PSAF actually learn meaningful continuous-time dynamics, or just a smooth and complex interpolation between supervised anchors? A high-quality interpolation itself can be a valuable contribution, but I think the paper should show it has real and robust advantage over simple interpolations.

  1. Missing baselines for the fractional-time results. The Tab.11 shows the midpoint fields (by construction in the paper, it is a convex mix of DkD_k and Dk+1D_{k+1}) match linear interp of PSAF's own hourly anchor, and it is close on all four structure stats. So, a natural baseline to me here at fractional times is this interpolation, but I didn't see it in Table 3 and 4. Besides, the advantage over CEF at matched 5.625∘5.625^\circ resolution seems to depend on the downsampling choice, i.e., the PSAF is worse than CEF on all four RTMA-RU metrics with conservative regridding. Adding the anchor-interp baseline to Tab. 3 and 4 could clarify what the fractional machinery adds.
  2. Physical-consistency claim is not well supported by the current results. In Table 5, to my understanding, the geostrophic and KE target errors at frational times was reported, but they are against the linearly interpolated diagnostics used in the training objective Lphysics\mathcal{L}_{\text{physics}}. So, these two numbers show that the surrogate was fit, not that the outputs are more "physical". It is also a bit unclear to me whether the held-out target MAE are computed against the interpolated targets. The Table 9 reports results on integer anchors, and it shows much smaller held-out gains compared to Table 5. These numbers seem to come from a single random seed as no std are reported (also from line 656, only a single random seed was reported). The effect on RMSE is also not reported, which I believe should be also very important to support the claim. I would suggest a lable hold-out ablation, e.g., train with labels only at even hours, and treat odd hours as fractional.
  3. The integer-time comparison is not well matched. According to A.5, PSAF uses the encoder’s latents and skip features computed from (X_{t-6}, X_t), while AMT, UPR-Net, and ModAFNO receive only (X_t, \hat X_{t+6}). So, the reported 10.1% gain over AMT is mixing architechtural differences, input information, and pretrained representations. PSAF (fixed-1h) is better than PSAF at 1-3h, but it was trained on 2006-2015 for 50 epoches, compared with 1979-2015 for 70 epochs for the backbone. This makes it difficult to say the performance gain is because of the forecasting approach. Besides, in table 1, it evaluates the stage-1 backbone, which is frozen in stage-2, rather than the proposed any-time mechanism. A lead-time conditioned forecaster trained with the same hourly labels, or giving endpoint baselines the same context, would make the comparison more informative.
Questions for authors
  1. Fractional-time accuracy. How does the PSAF compare with linear interpolation between its own hourly predictions on RTMA-RU and wind power? How sensitive are the reported gains over CEF to the re-gridding method?
  2. Physics surrogate supervision. Can you report MAE/RMSE with and without the physics loss? Are the held-out diagnostic targets in Table 5 computed from interpolated diagnostics or GT? Could you conduct a controlled label hold-out experiment to clarify whether the surrogate improves prediction at unlabeled times?
  3. Baseline fairness. How would endpoint baselines perform if given the same context as PSAF?
6 comments
AuthorsOpenReview ↗

Response to Reviewer oD89

Thank you for the constructive comments. We have added comparisons addressing interpolation, Physics Surrogate Supervision, and input context.

1. Fractional-time accuracy and regridding (W1/Q1). We retain the paper's RTMA-RU setup and extend queries to 0.25, 0.50, and 0.75 h within each hour. We include AMT, the strongest learned endpoint-reconstruction baseline in Table 2. Metrics and units follow the paper.

MethodT2mU10V10Wind vector
PSAF1.30921.11041.11691.7496
PSAF hourly interpolation1.30381.10931.11631.7482
6h field interpolation1.41971.12351.14511.7821
AMT1.35291.12381.13261.7722

PSAF outperforms 6h field interpolation and AMT on all four MAEs. Its own hourly interpolation is slightly better, with relative differences of 0.05%–0.41%.

Under the paper's wind-power setup, direct queries improve mean NMAE/NRMSE by 0.26%/0.45% over PSAF hourly interpolation:

Weather inputNMAENRMSE
PSAF direct queries0.093053 ± 0.0037240.138948 ± 0.005960
PSAF hourly interpolation0.093300 ± 0.0038960.139573 ± 0.006874

For both RTMA and wind power, hourly interpolation uses PSAF's learned hourly predictions, so the comparison measures the additional benefit of direct queries. The RTMA comparisons with 6h field interpolation and AMT assess the overall approach.

Regridding and spatial evaluation do affect the ranking. In the paper's pointwise evaluation, PSAF (conservative) has higher errors than CEF, whereas PSAF (bilinear) has lower errors. In the appendix's grid-cell-average evaluation, PSAF (conservative) has the lowest error on all four metrics. Bilinear regridding is more closely aligned with our main-text evaluation at forecast-grid centers, whereas conservative regridding better matches the grid-cell-average evaluation in the appendix.

2. Physics targets and the 2h-to-1h comparison (W2/Q2). Fractional diagnostic targets, including held-out ones, are linearly interpolated from diagnostics of adjacent ground-truth fields. “Held-out” denotes diagnostics excluded from the training losses. At integer times, where no physical loss is applied, we evaluate diagnostics against targets computed directly from ground-truth fields. The appendix reports error reductions of 22.2%/3.6% for geostrophic/KE diagnostics and 1.9%–2.5% for the four held-out diagnostics.

We also compare field RMSE with and without Physics Surrogate Supervision using the paper's test set and initialization times. The integer RMSEs are close, consistent with both models receiving the same integer full-field supervision:

LeadPSAFw/o Physics Surrogate
1h0.0240570.024081
2h0.0363670.036346
3h0.0439860.044040
4h0.0486040.048546
5h0.0513560.051021
6h0.0551680.055168

Following your suggestion, the 2h-to-1h experiment uses even-hour supervision anchors and evaluates intermediate odd-hour fields against ground truth. Under the same test protocol, RMSEs are:

LeadPSAFw/o Physics SurrogatePSAF 2h-anchor interpolation
1h0.0271970.0363680.028967
2h0.0376250.0372750.037625
3h0.0458590.0511450.046817
4h0.0492400.0490960.049240
5h0.0519300.0540160.053229
6h0.0551680.0551680.055168

At 1/3/5h, PSAF reduces mean RMSE by 11.69% versus w/o Physics Surrogate and 3.12% versus its own interpolation, providing direct field-accuracy evidence.

3. Input context and the fixed-1h reference (W3/Q3). We add Xt−6X_{t-6} to AMT while retaining XtX_t and X^t+6\widehat X_{t+6}, and continue training. Test RMSEs under the paper's protocol are:

LeadAMTAMT+historyPSAF
1h0.0292210.0286980.024057
2h0.0441870.0436590.036367
3h0.0519590.0514840.043986
4h0.0542150.0538430.048604
5h0.0538330.0536040.051356
6h0.0551680.0551680.055168

The reduction decreases from 1.79% at 1h to 0.43% at 5h, consistent with historical context supplying more useful information at shorter leads. The gains are modest; PSAF retains lower errors at 1–5h.

Stage 2 and PSAF (fixed-1h) both use 2006–2015. Hourly sampling yields six times as many samples per year: about 87,600 per epoch for this 10-year set, versus 54,100 for the 37-year 6h set. PSAF (fixed-1h) serves as a specialized reference for what the same architecture can achieve on the fixed-1h task; longer leads require recursion.

We will incorporate these additional results and the corresponding clarifications into the revised manuscript shortly.

AuthorsOpenReview ↗

Manuscript revision update

Thank you again for your constructive feedback. We have updated the manuscript with additional results and clarifications from our response. The main text now includes denser RTMA-RU evaluation, interpolation comparisons on RTMA-RU and wind power, and the controlled 2h-to-1h label-holdout experiment. The appendix includes the RTMA-RU component ablations, AMT with historical input, cross-window consistency, and rapid-change evaluations. We have also clarified the method description and evaluation protocols.

AuthorsOpenReview ↗

Follow-up on our response

Thank you again for your time and consideration. We would appreciate your feedback on our response and briefly summarize the additional evidence below.

Regarding the benefit of direct queries over interpolation, Tables 3 and 4 in the revised manuscript include comparisons with interpolation between PSAF's own hourly predictions. The overall errors on RTMA-RU are comparable, while direct queries perform slightly better on wind power. The controlled 2h-to-1h label-holdout experiment further withholds full-field labels at 1, 3, and 5h during training and evaluates predictions against ground truth at these leads. PSAF outperforms linear interpolation between its own predictions at all three leads, reducing mean RMSE by 3.12% (Table 6). This provides more direct field-accuracy evidence of the benefit of direct queries.

Regarding Physics Surrogate Supervision, the integer-field RMSEs reported in our earlier response are close, consistent with both models receiving the same integer full-field supervision. At the held-out leads in the 2h-to-1h experiment, however, PSAF reduces mean RMSE by 11.69% relative to w/o Physics Surrogate, showing its benefit at leads without full-field supervision. The RTMA-RU component ablations also provide verification against independent external analyses (Table 16).

Regarding physical consistency, integer leads receive full-field supervision but no physical loss. Their diagnostic targets are computed directly from ground-truth fields at the corresponding times, while fractional diagnostic targets, including those for held-out diagnostics, are obtained by linearly interpolating diagnostics of adjacent ground-truth fields. The improvements in held-out diagnostics and at integer leads provide additional evidence that the benefits of Physics Surrogate Supervision extend beyond fitting its direct training targets.

Table 11 also reports AMT with historical input, addressing the effect of input context.

We would welcome your feedback and any further questions. If these additions address your concerns, we would be grateful if you could reconsider your assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.