acceptodds
Under review as a conference paper at ICLR 2027

Regional Trajectory Conditioning for Hourly Kilometer-Scale Forecasts from Global Forecast

Abstract

Global weather forecasts predict large-scale evolution over days, but many applications need hourly, kilometer-scale probabilistic forecasts for a specific region. Regional history describes local conditions at initialization, while an issued global forecast provides information about weather entering the region later. That forecast can differ from the regional future even after spatial alignment, making it guidance rather than an exact regional target. When fine regional fields are generated in shorter windows, each ensemble member still forms an hourly sequence, but the global forecast and regional history do not specify which uncertain regional evolution its windows should share. We introduce Regional Trajectory Conditioning (RTC), a hierarchical conditional diffusion model that samples each member's regional trajectory to coordinate shorter fine windows. Conditioned on the global forecast and regional history, its coarse generator jointly samples that trajectory as hourly averages over blocks of neighboring regional-grid cells. The detail generator uses corresponding segments to produce 3 km fields that preserve those block averages exactly. Overlapping denoising contexts extend this coordinated generation from 24-hour training sequences to 10-day forecasts. Across four U.S. regions, RTC achieves lower state CRPS and trajectory Energy Score than every baseline at both 24-hour and 10-day horizons, and its 10-day scores at hourly lead times are 6.2% and 8.8% lower, respectively, than the lowest baseline score for each metric.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Weather Forecasting, Regional Weather Forecasting, Generative Modeling

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_Y8R9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Regional Trajectory Conditioning (RTC) uses global forecasts and regional history to first generate an hourly coarse trajectory for each ensemble member. It then generates fine-scale fields that preserve the trajectory’s block averages and uses overlapping denoising windows to extend the forecast horizon. Across four U.S. regions, RTC improves state CRPS, regional-mean trajectory Energy Score, and ensemble-mean errors for hourly 3 km forecasts.

Critical strengths

The method directly addresses how to coordinate predictions across temporal windows. The shared regional trajectory gives fine-scale generation a common condition, while the block-average constraint makes the relationship between the two scales explicit.

Critical weaknesses
  1. The controls in Table 3 show benefits in temporal-change metrics, but they do not report state CRPS or regional-mean trajectory Energy Score on the main 2025 test set. This makes it hard to connect these improvements to the paper’s main performance gains, especially since the 2024 validation results show that local coarse conditioning can achieve lower state CRPS. All Table 3 variants also preserve block averages, so this comparison does not establish the constraint’s contribution. Reporting the main scores for these variants and a matched constraint ablation would help clarify what each design choice adds.

  2. Fine-scale changes at window boundaries still need a closer look. RTC performs well on the broader detail-trajectory scores, but its boundary-change score is worse than several baselines, and adding overlap helps over 24 hours but not over 10 days. As Appendix A.1 explains, the selected 10-day boundary pairs also coincide with changes in the contributing outer contexts, so the current scores cannot separate these two effects. Comparing transitions where only one window schedule changes, alongside matched interior time pairs, would help identify where the remaining errors come from.

Questions for authors

Can the authors separately evaluate fine-window boundaries, outer-context boundaries, and matched interior times?

Field sequences or time series from individual members would help show where the problem occurs and which adjustments actually improve it. Overall, this paper requires further improvement regarding its presentation and experimental content.

5 comments
AuthorsOpenReview ↗

Response to Reviewer Y8R9 (P1)

We thank the reviewer for the careful review and for recognizing that the shared regional trajectory gives fine-scale generation a common condition with an explicit link between the two scales (S1). We first restate what RTC does and contributes, and then answer the two weaknesses. Each design choice has a distinct role, including the constraint. Reweighting the overlapping contexts, without retraining, brings the full-field boundary scores to the level of the best baselines, and shifting one window schedule shows that the remaining detail elevation follows the fine-window boundary.

Setup. The new results use the paper's 2025 test cases and scores, with 16 members. Lower is better, percentages are relative differences on the same cases, and trajectory ES denotes the regional-mean trajectory Energy Score. Detail is a field minus its 8×8 block average. B12 denotes the pairs of consecutive hours (12+24k, 13+24k), B00 the pairs (24k, 24k+1), and interior all other pairs. The cross-fade lets each of the overlapping 24-hour contexts of a 10-day forecast fade in and out with the weight w(τ) = sin²(π(τ+0.5)/24) instead of counting equally, with the trained models unchanged.


P1. Presentation

We restate here, as directly as we can, what RTC does and what the paper contributes. We will revise the paper so that these points are as easy to follow there.

  • The problem. The task is an hourly 3 km ensemble forecast for one region over several days. Four points explain why it needs the approach we take.
    • A regional forecast needs a global forecast. Weather systems enter the region through its boundaries during the forecast, so the history of the region alone does not determine its future. The issued global forecast carries this information.
    • The global trajectory is not the regional trajectory. The global forecast is given every 6 hours on a grid of tens of kilometers. It comes from a different modeling system and it has forecast errors. Even after it is aligned to the region and averaged over the same blocks, it differs from the regional evolution that verifies. On the regional grid, the aligned GraphCast forecast has a state CRPS of 0.287 at 24 hours and 0.445 over 10 days, against 0.148 and 0.331 for RTC. The regional evolution is therefore uncertain given the global forecast, and each ensemble member needs an hourly regional evolution of its own.
    • Fine fields have to be generated in short windows. Hourly 3 km fields over many hours are too large to generate jointly, so a generator produces them in windows of a few hours.
    • What this leaves open. Existing approaches represent the regional evolution inside each generation window, as in generative downscaling, or through step-by-step updates of the state, as in limited-area and autoregressive models. The global forecast and the regional history guide every window, but they do not say which of the possible regional evolutions the windows of one member should share.
  • The idea. The evolution at the regional scale has 64 times fewer values than the fine fields, so it can be sampled jointly over a whole day. RTC samples it first, once for each member, and gives it to all fine windows of that member as their common condition.
  • A member and its regional trajectory. An ensemble member is one sampled future of the region. Its regional trajectory U is the hourly sequence of that member's block averages, where each block averages 8×8 neighboring 3 km cells.
  • What RTC does. RTC first samples U jointly over 24 hours. It then generates the 3 km detail in two 12-hour windows. Each window follows its segment of the same U, so both windows describe the same regional evolution, and the generated fields reproduce U exactly in their block averages. For forecasts longer than 24 hours, 24-hour contexts that start every 12 hours are combined.
  • What the paper contributes. First, RTC uses the issued global forecast and the regional history to sample an hourly regional trajectory before it generates the kilometer-scale detail in multi-hour windows. Second, this trajectory is a shared probabilistic condition for the fine windows of each member. The generated fields preserve its block averages exactly, and composing denoising contexts extends the forecast beyond the 24-hour training sequences to 10 days. Third, RTC has lower state CRPS and regional-mean trajectory Energy Score than the baselines at both forecast horizons, and with fine-window overlap fixed, the shared coarse context lowers one-hour and 3-hour change CRPS.
  • How the experiments are organized. They ask three questions. Is the complete system more accurate than the alternatives? What changes when one design choice is replaced? At which lead times do the remaining errors occur? The answers to W1 and W2 below address the second and third questions.
AuthorsOpenReview ↗

Response to Reviewer Y8R9 (W1, Q2)

W1. Main scores of the design variants, and the constraint

We scored the Table 3 variants on the main scores of the 2025 test set and added a variant without the constraint. Each design choice shows a distinct role.

All rows share RTC's mean model. The two coarse controls give every member the same coarse trajectory. Learned deterministic coarse is the control of the paper's Table 15, a coarse generator trained to output a single trajectory. Ensemble-mean coarse uses RTC's own trained models unchanged and gives every member the mean of the 16 coarse trajectories that RTC samples.

Variantstate CRPS, 24 h3-h change CRPS, 24 htrajectory ES, 24 hstate CRPS, 10 d3-h change CRPS, 10 dtrajectory ES, 10 d
Local coarse + fine overlap0.145470.163370.050620.325430.217080.31485
Shared coarse + fine overlap0.145360.162960.049760.330560.211220.31289
RTC0.145370.163090.049760.330290.210980.31289
RTC without projection0.145550.163130.049560.331860.211500.31394
Learned deterministic coarse0.165950.183120.062350.368550.238850.33043
Ensemble-mean coarse0.169280.187470.063360.371550.244900.33240
Full-path residual diffusion0.147530.163350.056190.348160.225250.33715
Autoregressive 12-hour windows0.147870.163600.057540.329570.214190.32385

1. Shared against local coarse context (fine overlap fixed). The local variant has 1.6% lower 10-day state CRPS, in line with the 2024 result the reviewer cites, while the shared context is lower on the other five scores, with 2.7% lower 10-day change CRPS and 0.6% lower trajectory ES. Under the cross-fade the pattern holds and the gap at the boundary pairs widens. The shared context has 3.9% lower change CRPS, 1.0% lower trajectory ES and a full-field score at (12+24k, 13+24k) of 0.120 against 0.169, and the local one has 1.5% lower state CRPS.

2. The constraint. We added the matched ablation the reviewer asked for, RTC without the projection. Without the projection, block averages deviate from the sampled trajectory by up to 0.20 at 24 hours and 0.24 over 10 days, against numerical precision for RTC. The 24-hour scores differ by less than 0.5%, while the 10-day scores are 0.25–0.48% higher. The constraint therefore mainly makes U each member's actual regional trajectory, with a small 10-day score gain.

3. The stochastic coarse trajectory. RTC has lower scores than the learned deterministic coarse control on all three metrics at both horizons. The ensemble-mean control removes only the coarse spread and keeps RTC's trained models. The loss is as large (state CRPS +16.5% and +12.5%, trajectory ES +27.3% and +6.2%), so it comes from the missing coarse spread and not from the changed training objective.

4. No separate coarse-trajectory sampler (last two rows). Against these two variants RTC has 11.4% and 13.5% lower trajectory ES at 24 hours and 7.2% and 3.4% lower over 10 days. The autoregressive variant has 0.22% lower 10-day state CRPS, and RTC is lower on the other five scores.

Each choice therefore has a distinct role. The stochastic coarse trajectory lowers all three scores. Its separate generator lowers the trajectory ES against same-mean models. Sharing it over 24 hours lowers change and trajectory scores at a 1.5–1.6% cost in 10-day state CRPS. The constraint makes U the trajectory each member follows. The ensemble-mean control changes the coarse spread only, with the trained model unchanged.


Q2. Member-level figures

We added the two figures the reviewer asked for to the revised PDF, as Figures R1 and R2 in Appendix A.1 (marked in red). Figures cannot be shown in this comment, so we describe them and give their numbers here. The first shows all 16 members and the HRRR analysis at one grid point for hours 1–120. The second shows, for each pair of consecutive hours, the error of the members' one-hour block-average changes.

In the example case (2025-01-20, region 0), the mean absolute error of the one-hour T2m change over 16 sites and 16 members falls from 1.52 to 0.57 K at B12 and from 1.86 to 0.91 K at B00 with the cross-fade, and from 0.74 to 0.71 K at interior pairs. These are mean-member absolute errors of one case, not CRPS. They illustrate the table in our answer to W2 and Q1, which gives the scores over all test cases.

Over all test cases the same quantity falls from 1.82 to 0.68 K at B12 (lower in every case), from 1.47 to 0.78 K at B00 (lower in all cases but one), and from 0.66 to 0.63 K at interior pairs.

AuthorsOpenReview ↗

Response to Reviewer Y8R9 (W2, Q1)

W2 and Q1. Fine-window boundaries, outer-context changes and interior pairs

Over 10 days RTC has the lowest state CRPS and trajectory ES of all methods, and it has the lowest one-hour change scores at the 220 interior pairs. Changing the weights of overlapping contexts, without retraining, brings the full-field scores at the 19 boundary pairs to the level of the best baselines and makes RTC's 3-hour change CRPS the lowest as well. Shifting one window schedule separates the two effects, as the reviewer asked. The remaining detail elevation follows the fine-window boundary.

Main scores over 10 days:

state CRPS3-h change CRPStrajectory ES
Vanilla diffusion0.36140.23210.3537
CorrDiff0.35200.21240.3559
STVD0.38490.21310.3935
EnScale0.36360.21070.3431
StormCast0.51310.22330.4809
RTC0.33030.21100.3129
RTC + cross-fade0.32630.20000.3097

One-hour change CRPS over 10 days by pair class, for the full field and for the within-block detail. Both B12 and B00 contain a fine-window boundary and a change of outer contexts.

Full-field interiorFull-field B12Full-field B00Detail interiorDetail B12Detail B00
Vanilla diffusion0.13740.21130.24710.06700.07150.0822
CorrDiff0.14140.14140.16000.07470.07440.0829
STVD0.12630.13050.15330.06900.06910.0803
EnScale0.12680.12200.14650.06990.06870.0798
StormCast0.12820.12240.14900.06950.06790.0799
RTC0.12250.19480.20420.06640.07890.0861
RTC + cross-fade0.12060.12340.14420.06620.07480.0829

1. Fine-window boundary alone, at 24 hours. The pair (12, 13) crosses a fine-window boundary with no change of outer contexts. Its detail score is 0.0734, and 0.0684 with fine overlap.

2. Context weights. Over 10 days RTC averages overlapping 24-hour contexts with uniform weights. Changing only these weights, with the same checkpoints, lowers the block-average change CRPS at B12 from 0.177 to 0.095.

3. Cross-fade. With the fixed weight w(τ) = sin²(π(τ+0.5)/24), applied without any tuning, the full-field B12 and B00 fall by 37% and 29%, and the 10-day state CRPS, 3-hour change CRPS and trajectory ES by 1.2%, 5.2% and 1.0%. State CRPS at the 6-hour lead times rises by only 0.15% (0.14% in 2024). On 2024 validation data, the three scores fall by 1.6%, 4.6% and 1.4%, and the full-field one-hour change CRPS at (12+24k, 13+24k) falls by 34% (0.186 to 0.123).

4. Fine overlap and detail weighting over 10 days. Fine overlap alone does not lower the detail B12 over 10 days (0.0789 to 0.0810 with uniform weights, 0.0748 to 0.0760 with the cross-fade on the mean path and the coarse generator), but with the cross-fade also on the detail branch it falls to 0.0706, inside the baselines' range (0.068–0.074). This costs 0.09% in state CRPS and 2.3% in detail-state CRPS.

5. Transitions where only one window schedule changes (Q1). We moved the fine windows by 6 hours. Their boundaries then fall at lead times 6+12k, and the changes of outer contexts stay at 12k, so each kind of pair contains one transition only. Every pair is compared with the same pair, at the same lead time, in the default schedule. RTC + cross-fade, one-hour change CRPS, lead times 25–216:

Pairs at 12kPairs at 6+12kOther pairs
Detail, default schedule0.078400.062230.06668
Detail, fine windows shifted by 6 hours0.070440.070770.06673
Full field, default schedule0.132940.113770.12076
Full field, fine windows shifted by 6 hours0.128650.118660.12082

In the default schedule both transitions fall on the 12k pairs. After the shift the 12k pairs keep only the change of outer contexts, and the 6+12k pairs hold the fine-window boundary. The detail score falls by 10.2% where the fine-window boundary was and rises by 13.7% where it now is, so the remaining detail elevation follows the fine-window boundary.

The full-field score changes by −3.2% and +4.3%. With uniform weights the full-field score at the 12k pairs does not change (+0.05%), so the full-field elevation of the paper's setting belongs to the change of outer contexts.

The shift leaves the main scores unchanged (state CRPS −0.01%, 3-hour change CRPS +0.01%), so the default alignment of the two schedules costs no accuracy.

The full-field boundary scores thus reach the range of the best baselines, and away from the boundary pairs RTC has the lowest full-field and detail scores in the table.

Reviewer_p74POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies hourly, 3 km probabilistic regional forecasting conditioned on an issued global weather forecast and regional analysis history. The proposed Regional Trajectory Conditioning (RTC) first jointly samples a member-specific trajectory of regional block averages using a coarse diffusion generator. A detail diffusion generator then produces high-resolution structure in shorter temporal windows, while a zero-block-average projection ensures exact agreement with the sampled coarse trajectory. Overlapping denoising contexts extend generation from 24-hour training sequences to 10-day forecasts. On the 2025 test set covering four U.S. regions and four surface variables, RTC outperforms the reported baselines in state CRPS and regional-mean trajectory Energy Score at both horizons, with approximately 6.2% and 8.8% improvements, respectively, over 10 days at hourly lead times. Validation also includes temporal-context ablations, variable- and region-wise results, and a nine-variable extension.

Critical strengths

1. A well-defined shared-trajectory mechanism

RTC explicitly separates regional coarse-scale evolution from local detail generation, allowing short detail windows to follow a common member-specific stochastic trajectory. The conditional-center and projection construction ensures that the final fine fields exactly match that trajectory in their block averages, and this property is preserved by Heun sampling and temporal overlap averaging. This is a consequential technical strength because it provides a clear and verifiable mechanism for coordinating regional states across windows.

2. Forecasting gains across metrics, variables, and regions

RTC achieves the lowest state CRPS and regional-mean trajectory Energy Score among the reported baselines at both 24-hour and 10-day horizons, together with the lowest ensemble-mean RMSE for all four variables. Variable-wise, region-wise, and nine-variable extension results provide further support. This is the main empirical basis for my weak-accept recommendation, because the improvements extend beyond a single aggregate metric or region.

Critical weaknesses

1. weaker performance at window boundaries

RTC’s detail B12 is 0.07341 over 24 hours, worse than four of the five baselines, and 0.07893 over 10 days, worse than all five. The selected 10-day pairs involve both detail-window and outer-context transitions, so the current analysis does not identify the source of this weakness. This directly limits the evidence for temporal coordination across windows. I recommend separately scoring detail-window boundaries, outer-context transitions, and matched interior pairs, while independently varying the offsets and overlaps of the two window schedules.

2. confounded component ablations

The 24-hour and 96-hour training protocols differ in their initialization cohorts, update counts, and outer sampling contexts, so Table 5 does not isolate the effect of training-sequence length. The deterministic-coarse and no-projection controls are also reported only on 2024 validation data, and the authors acknowledge that these interventions change multiple aspects of the system. I recommend comparing training lengths under a common training cohort and sampling protocol, and adding stochasticity and projection controls on the 2025 test set while keeping other conditions as similar as possible.

3. baseline adaptation and budget matching require further validation

Equal update counts per training stage do not match total compute or target-frame exposure. In addition, EnScale uses a replacement backbone, and StormCast is adapted from its 99-variable source task to a four-variable task. These differences make the current comparison stronger evidence for complete-system gains than for architectural superiority under matched compute. Regarding whether the adapted implementations adequately retain the capabilities of the source methods. I recommend adding compute-matched results, baseline convergence curves, and targeted comparisons of the major adaptations.

4. conditional mismatch in detail generation

The detail generator is trained with reference block averages, U=AYU=AY, but is conditioned on generated UU during inference. Hard projection guarantees aggregate agreement, but does not ensure that detail quality remains reliable under erroneous coarse conditions. Regarding the empirical impact of this mismatch. I recommend comparing detail and boundary scores under reference, generated, and controllably perturbed coarse trajectories, and testing whether generated or perturbed conditioning during training improves robustness.

5. the necessity of a shared stochastic trajectory needs more direct support

The paper clearly explains why a global forecast does not exactly specify the future regional state. However, this primarily motivates regional correction; it does not by itself establish the need to jointly sample a shared stochastic coarse trajectory before generating detail. Table 3 supports benefits of shared coarse context for several temporal scores, but local coarse contexts achieve lower 10-day state CRPS in the 2024 results in Table 15, indicating dependence on the metric and evaluation period. I recommend distinguishing the motivation for regional correction from that for stochastic trajectory coordination, and using common-mean, budget-matched controls to identify which temporal-dependence gains arise from the latter.

Questions for authors

See Weakness

6 comments
AuthorsOpenReview ↗

Response to Reviewer p74P (W1)

We thank the reviewer for the precise review and for recognizing the well-defined shared-trajectory mechanism (S1) and the gains across metrics, variables and regions (S2). The boundary weakness comes from one inference setting, the uniform averaging of overlapping 24-hour contexts, and not from trajectory conditioning. With the trained model unchanged, a standard cross-fade brings RTC's full-field boundary scores to the level of the best baselines. Removing only the coarse spread at inference raises the three main scores by 6–27%, and STVD trained for longer than RTC stays 6–11% above it in state CRPS.

Setup. The new results use the paper's 2025 test cases and scores, with 16 members. Lower is better, percentages are relative differences on the same cases, and trajectory ES denotes the regional-mean trajectory Energy Score. Detail is a field minus its 8×8 block average. B12 denotes the pairs of consecutive hours (12+24k, 13+24k), B00 the pairs (24k, 24k+1), and interior all other pairs. The cross-fade lets each of the overlapping 24-hour contexts of a 10-day forecast fade in and out with the weight w(τ) = sin²(π(τ+0.5)/24) instead of counting equally, with the trained models unchanged.


W1. Boundary scores

Over 10 days RTC has the lowest state CRPS of all methods and the lowest one-hour change CRPS at the 220 interior pairs, and with the cross-fade also the lowest 3-hour change CRPS. We scored the three kinds of pairs separately and varied the two schedules, as the reviewer recommended.

1. All methods on the same pairs. In the default schedule B12 and B00 both contain a fine-window boundary and a change of outer contexts. The first two columns are 10-day scores, and the others are one-hour change CRPS by pair class.

Methodstate CRPS3-h change CRPSFull-field interiorFull-field B12Full-field B00Detail B12
Vanilla diffusion0.36140.23210.13740.21130.24710.0715
CorrDiff0.35200.21240.14140.14140.16000.0744
STVD0.38490.21310.12630.13050.15330.0691
EnScale0.36360.21070.12680.12200.14650.0687
StormCast0.51310.22330.12820.12240.14900.0679
RTC0.33030.21100.12250.19480.20420.0789
RTC + cross-fade0.32630.20000.12060.12340.14420.0748

2. Source of the full-field elevation. The paper averages the two contexts that cover each hour with equal weights, so every 12 hours the average switches abruptly from one pair of contexts to the next. Changing only these weights, with the same checkpoints, lowers the block-average change CRPS at B12 from 0.177 to 0.095. Vanilla diffusion has no coarse trajectory and shows the same elevation (0.2113 at B12), so trajectory conditioning is not what produces it.

3. A cross-fade is the direct remedy. Each context fades in and out instead of switching, and nothing has to be tuned. The full-field B12 falls from 0.1948 to 0.1234, 1.2% above the lowest baseline. The three main 10-day scores fall as well, by 1.0–5.2% (1.4–4.6% on 2024 validation data), with only a 0.15% increase in state CRPS at the 6-hour lead times.

The same cross-fade does not make other methods match RTC. Applied to vanilla diffusion and STVD, which combine contexts in the same way, it leaves their state CRPS 9.5% and 17.9% above RTC's and their full-field change CRPS above RTC's at all three kinds of pairs. RTC's temporal consistency therefore comes from the model itself.

4. Fine overlap. With overlapping fine windows and the detail branch cross-faded too, the detail B12 falls from 0.0748 to 0.0706, within the baseline range (0.068–0.074). This costs 0.09% in state CRPS and 2.3% in detail-state CRPS.

5. Offset of the two schedules. We moved the fine windows by 6 hours, so that their boundaries (6+12k) no longer coincide with the changes of outer contexts (12k), and compared each pair at the same lead time. RTC + cross-fade, one-hour change CRPS:

Pairs at 12kPairs at 6+12kOther pairs
Detail, default schedule0.078400.062230.06668
Detail, fine windows shifted by 6 hours0.070440.070770.06673
Full field, default schedule0.132940.113770.12076
Full field, fine windows shifted by 6 hours0.128650.118660.12082

The detail score falls by 10.2% at the 12k pairs and rises by 13.7% at the 6+12k pairs, so the detail elevation follows the fine-window boundary. The shift leaves the main scores unchanged (state CRPS −0.01%), so the default alignment costs no accuracy.

6. Whole trajectories. The detail Energy Score of the paper's RTC is lower than that of all five baselines at both horizons (0.1038 against 0.1047 for the best baseline at 24 hours, 0.1227 against 0.1269 over 10 days).

AuthorsOpenReview ↗

Response to Reviewer p74P (W2, W5)

W2a. Training-sequence length

Table 5 compares RTC trained on 24-hour sequences with the same model trained on 96-hour sequences. Following the reviewer's recommendation, we sampled both in the same way, with the main setting's 24-hour contexts. The difference stays as in Table 5, so it does not come from the sampling context.

The 96-hour protocol has 2.7% lower 10-day state CRPS and 2.4% lower trajectory ES than the main RTC, its change CRPS differs by +0.2%, and its first-day state CRPS is 2.1% higher (0.1598 against 0.1566). The longer protocol thus trades first-day for long-horizon accuracy, and the main RTC reaches 10 days with 24-hour training sequences only.

We will add this result to the revised paper and describe Table 5 as a comparison of the two training protocols, whose training cohorts and update counts differ.


W2b. Stochasticity and projection controls on the 2025 test set

On the 2025 test set both controls show the same pattern as on the 2024 validation data in Table 15. The table also lists the design variants discussed in W5, and all rows share RTC's mean model.

Variantstate CRPS, 24 h3-h change CRPS, 24 htrajectory ES, 24 hstate CRPS, 10 d3-h change CRPS, 10 dtrajectory ES, 10 d
Local coarse + fine overlap0.145470.163370.050620.325430.217080.31485
Shared coarse + fine overlap0.145360.162960.049760.330560.211220.31289
RTC0.145370.163090.049760.330290.210980.31289
RTC without projection0.145550.163130.049560.331860.211500.31394
Learned deterministic coarse0.165950.183120.062350.368550.238850.33043
Ensemble-mean coarse0.169280.187470.063360.371550.244900.33240
Full-path residual diffusion0.147530.163350.056190.348160.225250.33715
Autoregressive 12-hour windows0.147870.163600.057540.329570.214190.32385

1. Stochasticity, with other conditions kept the same. Both stochasticity controls give every member the same coarse trajectory. Learned deterministic coarse is the control of the paper's Table 15, a coarse generator trained to output a single trajectory, so it also changes the coarse training objective. Ensemble-mean coarse keeps RTC's trained models, sampler and detail noise and gives every member the mean of the 16 coarse trajectories that RTC samples.

With it the three scores rise by 16.5%, 15.0% and 27.3% at 24 hours and by 12.5%, 16.1% and 6.2% over 10 days, as much as with the learned deterministic control. The loss therefore comes from removing the coarse spread.

2. Projection. With the projection, block averages equal the sampled trajectory to numerical precision. Without it they deviate by up to 0.20 and 0.24, and the 24-hour scores differ by less than 0.5%, while the 10-day scores are 0.25–0.48% higher.


W5. Necessity of a shared stochastic trajectory

We separate three gains, each with a control that shares RTC's mean model (table in W2b).

1. Regional correction. The aligned GraphCast input scores 0.287 at 24 hours and 0.445 over 10 days at its 6-hour lead times, against 0.148 and 0.331 for RTC.

2. A stochastic coarse trajectory, sampled first. The full-path residual diffusion and the autoregressive-window variant (last two rows) use RTC's mean model and generate the full field in one stage. Against them RTC has 11.4% and 13.5% lower trajectory ES at 24 hours, 7.2% and 3.4% lower over 10 days, and 6.3% and 1.5% lower 10-day change CRPS.

The autoregressive variant has 0.22% lower 10-day state CRPS, and RTC is lower on the other five scores of the table. The ensemble-mean control of W2b shows that the spread of the coarse trajectory is needed as well.

3. Sharing it across the fine windows. The local variant has 1.6% lower 10-day state CRPS, as in Table 15, while the shared context is lower on the other five scores, with 2.7% lower 10-day change CRPS and 0.6% lower trajectory ES. With the cross-fade the gap at the boundary pairs widens, to a full-field B12 of 0.120 against 0.169.

Regional correction accounts for the gain over the global input, and the stochastic coarse trajectory for the further gain over models with the same mean. Against the local variant, sharing it trades 1.5–1.6% of 10-day state CRPS for lower change and trajectory scores.

AuthorsOpenReview ↗

Response to Reviewer p74P (W3)

W3. Baseline budget, convergence and adaptations

RTC keeps lower state CRPS and trajectory ES than all five baselines. Three already train longer than RTC, and the other two stay behind after extended training, including STVD trained for longer than RTC.

CorrDiff, EnScale and StormCast train for 8.4, 10.0 and 12.1 hours per seed against 6.9 for RTC. Vanilla diffusion and STVD train for 2.6 and 4.2 hours, so we trained them further, until each had trained for as long as RTC at its own training speed. This gives 105k updates for vanilla diffusion (6.88/2.61 × 40k) and 65k for STVD (6.88/4.21 × 40k).

1. STVD at 65k updates (RTC's training time). It is resumed from the original training state. Its state CRPS is 3.0% lower than the paper's STVD at 24 hours and 1.6% lower over 10 days. Against RTC it is 8.3% and 14.7% higher, and its trajectory ES 36% and 22% higher. Its 10-day 3-hour change CRPS is level with RTC's (−0.05%).

2. STVD at 90k updates (longer than RTC). We also continued STVD to 90k updates, about 9.5 hours at its original training speed against 6.9 for RTC.

STVD against RTCstate CRPS, 24 h / 10 d3-h change CRPS, 24 h / 10 dtrajectory ES, 24 h / 10 d
65k updates+8.3% / +14.7%+2.4% / −0.05%+36% / +22%
90k updates+6.1% / +10.8%+1.4% / −0.7%+29% / +14%

For each of the three training seeds, the 10-day gap at 90k is clear (state CRPS +7.9% to +12.3%, trajectory ES +10.2% to +15.7%).

3. With the cross-fade on both methods. STVD composes its 10-day forecasts from overlapping contexts in the same way as RTC, so we applied the cross-fade of W1 to it as well:

10 days, cross-fade on bothstate CRPS3-h change CRPStrajectory ES19 boundary pairs, STVD / RTC
STVD 65k against RTC+16.0%+5.4%+23.5%0.138 / 0.133
STVD 90k against RTC+12.1%+4.8%+14.8%0.135 / 0.133

RTC then also has the lower score at the boundary pairs, where the paper's RTC was behind, and only the detail score at B12 remains lower for STVD (0.0677 and 0.0664 against 0.0748).

4. Vanilla diffusion at 105k updates (RTC's training time). It improves state CRPS by 6.1%/2.4% at 24 hours/10 days. It remains 20.1%/6.7% above RTC, with trajectory ES 86%/11.7% higher and 10-day change CRPS 6.4% higher. With the cross-fade, its full-field B12 is level with RTC's.

5. Convergence curves. On 2024 validation data, both baselines gain little from the later updates and stay well above RTC (0.1459 on the same cases). STVD's state CRPS is 0.1601 at 40k, 0.1562 at 65k and 0.1552 at 90k. The last 25k updates lower it by 0.7%, and it remains 6.3% above RTC.

Vanilla diffusion's is 0.1828 at 40k, 0.1752 at 70k and 0.1730 at 105k. The last 35k updates lower it by 1.2%, and it remains 18.6% above RTC. Training STVD for longer than RTC therefore does not close the gap.

A staged variant of STVD, which scores better than the joint one, also stays above RTC (10-day state CRPS 0.372 against 0.330).

6. Adaptations. We kept each baseline as close to its source method as the task allows. StormCast uses the official architecture, condition builder, loss and sampler, and only its target is the four variables of this task. EnScale keeps its Energy-Score objective, its two stages and its autoregressive conditioning, with a shared convolutional backbone. All five baselines use the same inputs, training data and evaluation as RTC, so Table 1 compares complete systems under this common protocol.

Reviewer_d8AbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies an important problem of global weather prediction, focusing on hourly kilometer-scale forecasting. The authors propose regional trajectory conditioning, a hierarchical conditional diffusion model based on regional trajectory sampling. Generally, this paper is well-structured. In addition, experiments show the effectiveness of the proposals.

Critical strengths
  1. This paper is well-written and easy to follow.
  2. A hierarchical conditional diffusion model, RTC, is proposed for hourly kilometer-scale global weather forecasting. Extensive experiments show the effectiveness of the proposal.
Critical weaknesses
  1. After carefully reviewing this paper, I have noticed that some concepts are unclear. It is suggested to include more clarification regarding these unclear concepts. For example, it is hard to understand the meaning of "each member's regional trajectory" in the abstract. It would be better to give more explanation. In addition, it would be better to proofread the manuscript and correct similar problems.
  2. Recently, large language model and foudation model based methods are widely used for global weather prediction. It would be better to include more corresponding baselines as follows.

[1]. Accurate medium-range global weather forecasting with 3D neural networks, Nature 2023.

In addition, it is suggested to include the following method as a baseline.

[2]. Learning skillful medium-range global weather forecasting, Science 2023.

Further, according to Appendix D.4, the training and evaluation cost of the proposed RTC is high. It would be better to provide theoretical computational analysis, i.e., space and time complexities. Further, it is suggested to give more explanations for why we need this method, given its poor efficiency.

Questions for authors

Please see the weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer d8Ab (W1, W2)

We thank the reviewer for finding the paper well written and the experiments extensive (S1, S2). Following the reviewer's suggestions, we scored five global forecasts on the paper's test cases, ran RTC with a second global model as its input, and compared cost with accuracy for every method. On the regional grid RTC has 47–52% lower state CRPS than the five global forecasts at 24 hours and 23–29% lower over 10 days, and no baseline that is cheaper to train or to run reaches its scores.

Setup. The new results use the paper's 2025 test cases and scores, with 16 members. Lower is better, percentages are relative differences on the same cases, and trajectory ES denotes the regional-mean trajectory Energy Score.


W1. "Each member's regional trajectory"

It is the hourly evolution of the region, at the scale of about 24 km, that one ensemble member follows. We explain the three words in turn.

  • Member. An ensemble forecast consists of several forecasts, and each of them is a member. RTC produces 16 members, and each is one possible future of the region.
  • Regional. The values are not on the 3 km grid itself. Each value is the average over a block of 8×8 neighboring 3 km cells, which we call a block average.
  • Trajectory. It is the sequence of these block averages at every hour of the forecast, so it describes how the region evolves over time and not its state at one time.

Why each member has its own. A global forecast does not determine the future of the region exactly, so several regional evolutions remain possible. RTC therefore samples a separate trajectory for every member. It is not the global forecast interpolated to the region. RTC samples it anew, with the global forecast as a condition.

What RTC does with it. RTC samples the trajectory of a member jointly over 24 hours. It then generates the 3 km detail in two 12-hour windows. Both windows receive the same trajectory, so they describe the same regional evolution although they are generated separately, and the block averages of the generated fields equal the sampled trajectory.

We will state this definition in the abstract, use the terms consistently and proofread the manuscript, so that the concepts are easier to follow.


W2. Global foundation-model baselines

On the regional 3 km grid, RTC has 47–52% lower state CRPS than five global forecasts at 24 hours and 23–29% lower over 10 days. Driven by Pangu-Weather without retraining, it is 45% below Pangu-Weather itself at 24 hours and 24% below over 10 days.

The task is regional forecasting from an issued global forecast. RTC and every baseline take GFS-initialized GraphCast Operational forecasts (0.25°, 6-hourly) as input and produce hourly 3 km ensembles for a region. Global models are therefore the input of this task, and their accuracy on the regional grid is the level that RTC has to improve on.

We scored Pangu-Weather [1] and GraphCast [2] directly, and added FourCastNet v2, the foundation model Aurora and the physics-based GFS forecast. All are GFS-initialized 0.25° forecasts from public archives. They are interpolated to the regional grid in the same way as RTC's input and compared with the HRRR analysis on the paper's 2025 test cases, at the 6-hour lead times that global models provide.

state CRPS at 6-hour lead times24 hours10 days
GFS (physics-based)0.3000.465
GraphCast (RTC's input)0.2870.445
Pangu-Weather0.2980.437
FourCastNet v2-small0.3060.451
Aurora0.2810.431
RTC0.1480.331
RTC with Pangu-Weather as input0.1630.332

The five global forecasts differ from one another by about 0.03, while RTC is 0.13 below the best of them at 24 hours and 0.10 below over 10 days. The gain of the regional model is three to five times the spread among the global models. RTC also has lower ensemble-mean RMSE than its GraphCast input for each of the four variables.

RTC with another global model as its input. We ran the RTC checkpoints, trained only with GraphCast input, with Pangu-Weather as input and no retraining. At 24 hours the score in the table's metric is 0.163, 45% below Pangu-Weather itself. With the changed input the hourly state CRPS is 8% higher (0.157) and the trajectory ES 31% higher (0.065), yet both remain below those of four of the five baselines of Table 1, which use the GraphCast input they were trained on (0.152–0.205 and 0.063–0.127).

Over 10 days it scores 0.332, 24% below Pangu-Weather itself (0.437). Its hourly state CRPS and trajectory ES do not differ from those of GraphCast-driven RTC, with a 3-hour change CRPS that is only 2.3% higher.

For a deterministic forecast, CRPS equals the mean absolute error. We will add these reference rows and the Pangu-Weather-driven result to the revised paper.

AuthorsOpenReview ↗

Response to Reviewer d8Ab (W3)

W3. Cost, complexity, and why RTC is needed

RTC trains faster than three of the five baselines and runs 8–13% longer than the two closest diffusion baselines. No baseline that is cheaper to train or to run reaches its state CRPS or trajectory ES, and its evaluation cost can be cut by a third with lower 10-day scores than in the paper's setting.

Costs are measured on H200 GPUs (Table 7), with training summed over all stages per seed and evaluation timed for one 16-member 10-day forecast. Scores are from Table 1.

MethodTrain (h/seed)Peak train memory (GiB)Eval 10 d (s)state CRPS, 24 h / 10 dtrajectory ES, 24 h / 10 d
Vanilla diffusion2.6112.37560.186 / 0.3610.100 / 0.354
CorrDiff8.364.33740.152 / 0.3520.063 / 0.356
STVD4.2144.77230.162 / 0.3850.072 / 0.393
EnScale10.036.8970.176 / 0.3640.080 / 0.343
StormCast12.074.911650.205 / 0.5130.127 / 0.481
RTC6.8812.48200.145 / 0.3300.050 / 0.313

1. Training. In the paper every method trains each of its stages for 40k updates. With this, CorrDiff, EnScale and StormCast train 1.2–1.8 times longer than RTC and still have 4–41% higher state CRPS at 24 hours and 7–55% higher over 10 days.

Vanilla diffusion and STVD train faster than RTC, so for this response we trained them further in two steps. First, we gave each of them RTC's training time of 6.9 hours, and RTC still has lower state CRPS and trajectory ES. Second, we trained STVD for longer than RTC, about 9.5 hours, and RTC is still lower.

Baseline trained furtherstate CRPS above RTC, 24 h / 10 dtrajectory ES above RTC, 24 h / 10 d
Vanilla diffusion, as long as RTC (105k updates)+20.1% / +6.7%+86% / +11.7%
STVD, as long as RTC (65k updates)+8.3% / +14.7%+36% / +22%
STVD, longer than RTC (90k updates)+6.1% / +10.8%+29% / +14%

2. Evaluation. A 16-member, hourly, 3 km, 10-day forecast takes RTC about 14 minutes on one GPU, 8–13% more than vanilla diffusion and STVD and 30% less than StormCast. The two faster baselines, CorrDiff and EnScale, are less accurate, with 4–21% higher state CRPS and 10–61% higher trajectory ES than RTC.

3. Fewer sampler steps, without retraining. With 12 or 9 sampler steps instead of 18, the denoiser calls and the measured sampling time fall by 34% and 51%. The three 24-hour scores change by less than 1% with 12 steps and by at most 2.4% with 9, and with 9 steps they remain below those of every baseline in Table 1. Over 10 days, 12 and 9 steps give 0.5% and 1.2% lower state CRPS than 18 steps, and 3-hour change CRPS changes by at most 0.5%.

4. Lower cost and lower scores together. A cross-fade lets the overlapping 24-hour contexts of a 10-day forecast fade in and out with the weight w(τ) = sin²(π(τ+0.5)/24), instead of counting equally, and uses the same trained models. With the cross-fade and 12 steps, RTC needs 34% fewer denoiser calls than in the paper and has lower scores on all three 10-day measures.

RTC over 10 daysDenoiser calls per windowstate CRPS3-h change CRPStrajectory ES
Paper setting (equal weights, 18 steps)350.33030.21100.3129
Cross-fade, 18 steps350.32630.20000.3097
Cross-fade, 12 steps230.32480.20030.3092

5. Why RTC is needed. RTC is the most accurate of the six methods at both horizons, with a training time in the middle of the baselines (6.9 hours against 2.6 to 12.1) and an evaluation cost that falls by a third without retraining. The two baselines that train faster do not reach its state CRPS or trajectory ES when they train for as long as RTC, and STVD does not when it trains for longer. The accuracy of RTC therefore comes from the framework itself, sampling each member's regional trajectory before the detail, and not from a larger budget.

6. Complexity. Sampling time grows linearly with the forecast length, the number of members and the number of sampler steps. A forecast uses KK contexts, and KK grows in proportion to its length. With MM members and NN denoiser calls per window, the work is approximately

KFμ+MNK (Fc+2Fd),K F_\mu + M N K \,(F_c + 2 F_d),

where FμF_\mu, FcF_c and FdF_d are the costs of one call of the mean, coarse and detail networks, and each context has two fine windows. The coarse network works on fields with 64 times fewer grid points than the detail network. The memory of the networks depends only on the window sizes, 24 hours for the coarse and 12 hours for the detail network, so it does not grow with the forecast length. Only the stored forecast grows, in proportion to members × variables × hours × grid points.

AuthorsOpenReview ↗

Discussion Follow-up

Dear Reviewer d8Ab,

As the discussion period is coming to an end, we would appreciate it if you could take a look at our additional experiments and analyses. Please let us know if you have any further questions or concerns. We would be happy to address them.

Thank you for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful reading and constructive feedback. We are encouraged that they found the shared-trajectory mechanism explicit and verifiable (p74P, Y8R9), the gains consistent across metrics, variables and regions (p74P), and the experiments extensive (d8Ab).

During the discussion period, we ran the analyses the reviewers recommended to address the main concerns:

  • Boundary pairs scored separately for all methods, with the weights and offsets of the two window schedules varied (p74P, Y8R9)
  • Main scores of the design variants, with stochasticity and projection controls (Y8R9, p74P)
  • Baselines trained as long as RTC or longer, and detail training on generated or perturbed trajectories (p74P)
  • Five global forecasts on the regional grid, a second global model as input, and cost against accuracy (d8Ab)

Summary of the Results

In short, RTC's gain comes from sampling each member's coarse trajectory before the detail, and a small auxiliary adjustment at inference, with the trained model unchanged, resolves its weaker scores at window boundaries.

1. RTC's advantage comes from the model, and a small adjustment at inference resolves the boundary scores (p74P, Y8R9).

  • In the paper's setting, RTC has the lowest state CRPS and trajectory ES of all methods at both horizons, and the lowest change scores away from the context boundaries.
  • As an auxiliary adjustment, a cross-fade of overlapping contexts lowers its full-field boundary scores by 29–37%, to the level of the best baselines (from 0.195 to 0.123 and from 0.204 to 0.144, against 0.122 and 0.147). It needs no retraining and nothing to tune, and the 10-day state CRPS, 3-hour change CRPS and trajectory ES also fall, by 1.2%, 5.2% and 1.0% (1.6%, 4.6% and 1.4% on 2024 validation data).
  • The same adjustment does not bring other methods to RTC's scores. Applied to vanilla diffusion and STVD, which combine contexts in the same way, it leaves their state CRPS 9.5% and 17.9% above RTC's. RTC's temporal consistency comes from the model itself.

2. Each design choice has a measured role (Y8R9, p74P).

  • Against the learned deterministic coarse generator of the paper's Table 15, RTC has 5–20% lower state CRPS, change CRPS and trajectory ES. It is also 6–21% lower than a control that removes only the spread of its sampled trajectories, with the trained models unchanged, so the gain comes from the coarse spread.
  • Against two variants that use RTC's mean model but generate the full field in one stage, RTC has 11.4% and 13.5% lower trajectory ES at 24 hours, 7.2% and 3.4% lower over 10 days, and 6.3% and 1.5% lower 10-day change CRPS. Both of its components are more accurate at 24 hours, with 1.7% and 2.1% lower state CRPS of the block averages and 1.0% and 0.7% lower state CRPS of the detail.
  • Separate 12-hour coarse segments have 1.6% lower 10-day state CRPS, while sharing the trajectory over 24 hours is lower on the other five main scores, with 2.7% lower 10-day change CRPS and 0.6% lower trajectory ES.
  • The projection makes every member match its sampled trajectory to numerical precision, and without it the main scores change by less than 0.5%.

3. No cheaper baseline and no global model reaches RTC's scores (p74P, d8Ab).

  • CorrDiff, EnScale and StormCast already train longer than RTC (8.4, 10.0 and 12.1 hours against 6.9) and have 4–41% higher state CRPS at 24 hours and 7–55% higher over 10 days. Trained for as long as RTC, vanilla diffusion keeps 20.1% and 6.7% higher state CRPS and STVD 8.3% and 14.7%. Trained for longer than RTC, about 9.5 hours, STVD keeps 6.1% and 10.8%.
  • With the cross-fade and 12 instead of 18 sampler steps, RTC needs 34% fewer denoiser calls, and all three 10-day scores are lower than in the paper's setting, by 1.7%, 5.1% and 1.2%.
  • On the regional grid, RTC has 47–52% lower state CRPS than five global forecasts at 24 hours and 23–29% lower over 10 days. GFS, GraphCast (RTC's input), Pangu-Weather, FourCastNet v2 and Aurora score 0.281–0.306 and 0.431–0.465 at their 6-hour lead times, against 0.148 and 0.331 for RTC. With Pangu-Weather as its input and no retraining, RTC is 45% and 24% below Pangu-Weather itself.

Setup. The new results use the paper's 2025 test cases with 16 members, 288 cases at 24 hours and 96 over 10 days, and the paper's Table 1 is unchanged. Lower is better, percentages are relative differences on the same cases, and trajectory ES denotes the regional-mean trajectory Energy Score. The cross-fade lets each of the overlapping 24-hour contexts of a 10-day forecast fade in and out with the weight w(τ) = sin²(π(τ+0.5)/24) instead of counting equally.

Per-experiment setups, full result tables and specific replies to each review are in the individual comments below. We will add these results to the revised paper and make the definitions of a member, its regional trajectory and the two window schedules easier to follow.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.