acceptodds
Under review as a conference paper at ICLR 2027

NucleCast: Low-Cost Residual Flow-Matching Refinement for Radar Nowcasting

Abstract

Radar nowcasting requires predicting rapidly evolving precipitation fields while accounting for uncertainty in their future development. Deterministic forecasts often smooth over multiple plausible outcomes, whereas generative models represent alternative futures through ensembles. Repeated network evaluations make diffusion and flow-based ensemble sampling computationally and infrastructure-intensive. We introduce NucleCast, a two-stage probabilistic nowcasting framework that refines a deterministic forecast anchor rather than generating the entire future field from noise. An efficient Mamba-3-based forecaster first predicts the large-scale evolution. A pixel-space conditional flow matching model then directly predicts clean residual corrections using a mask-conditioned compound-Poisson source that proposes localized precipitation variations. Each ensemble member requires only four flow matching model evaluations, with the same deterministic anchor shared across members. Evaluated on forecast skill, spatial and spectral structure, NucleCast achieves performance competitive with state-of-the-art generative baselines while drastically reducing computational cost. Furthermore, we investigate newly emerging precipitation by explicitly distinguishing recovered events from false off-support predictions. Our sampling analysis reveals that finer numerical integration generates additional fine-scale structures and off-support precipitation.

Primary area:
generative models
Keywords:
Precipitation Nowcasting, Conditional Flow Matching, Residual Generative Modeling, Probabilistic Forecasting, Efficient Generative Models, Spatiotemporal Forecasting

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_bLi7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

NucleCast is a two-stage radar-nowcasting framework combining a frozen deterministic forecaster with a conditional flow-matching refiner. The refiner models pixel-space residuals, uses an anchor-guided compound-Poisson source, and combines ensemble members through a probability-matched mean for deterministic verification. Experiments on SEVIR and ARSO report improvements over the anchor and analyze intense precipitation, precipitation births, numerical integration, and longer-horizon forecasting.

Critical strengths

The complete system offers a practical inference configuration. Reported results show improvements over the anchor at low inference cost, with a smaller model than the compared FlowCast configuration; these deployment results have practical value.

Refinement benefits are supported on two radar datasets. Intense-precipitation forecasts improve, and the birth-recovery analysis measures newly introduced false components, making both benefits and costs visible.

Critical weaknesses

1.The structured source's independent benefit is not established. The source comparison is limited to one example, and its training setup is unclear, so it does not establish a general benefit. Please provide a fair aggregate source comparison or narrow the claim.

2.Probabilistic forecasting needs direct validation. The reported CRPS evaluates a Gaussian fitted to the members, while improvements after aggregation do not establish ensemble calibration. The claimed non-Gaussian or multimodal forecast quality therefore still lacks direct evidence. Please provide direct ensemble diagnostics using existing predictions.

Questions for authors

Q1 : What controlled evidence supports the structured source's independent benefit?

Q2 : What direct ensemble diagnostics support its probabilistic forecast quality?

Q3 : What cost evidence supports the training-cost advantage and residual refinement's efficiency contribution?

Q4 : Does feedback outperform output-only refinement under the same protocol?

1 comment

We thank the reviewer for the constructive questions. Each one led to a new experiment, and in two cases to a narrower claim: on what the structured source improves (Q1) and on the ensemble (Q2).

Q1. Controlled evidence for the structured source's independent benefit.

We trained an identical twin of the released refiner, also 50 epochs, in which only the source changes to Gaussian noise. We compare the two at matched epochs on SEVIR Cohort B, paired over storm timestamps (Section 4.3.1, Figure 4; exact values in Appendix Table 12).

  • At the final epoch (no checkpoint selection), the structured source does not change CSI-M. It raises the extreme-area budget (bias@219 +0.06, interval excluding zero), with CSI@219 (+0.0107) and POD@219 higher but not resolved.
  • At the deployed epoch 16 it leads on CSI-M (+0.0031 [+0.0011, +0.0048]) and CSI@219 (+0.0126 [+0.0061, +0.0186]), while the Gaussian twin over-forecasts extreme area (bias@219 +0.66). Single draws show a larger margin (+0.0088 CSI-M, +0.0163 CSI@219; Appendix Table 12). The over-forecasting twin has the lower fair CRPS at that epoch, but negative Brier skill over the anchor at VIL 160 and 219 (−0.08, −0.19, against +0.05, +0.01 for NucleCast; Appendix A.10). Epoch 16 was selected on validation for NucleCast, not for the twin, so we also report the final epoch as the selection-free comparison.

Following the reviewer's suggestion, we narrowed the claim: the source improves the prediction of extreme cores, not general skill. Figure 4 also shows where general skill comes from: the residual target (+0.0105 CSI-M) and the anchor clip (+0.0044). Appendix A.9 reports the proposal-rate variants.

Q2. Direct ensemble diagnostics.

Section 4.3.2 and Appendix A.10 now report fair ensemble CRPS, rank histograms, spread–skill and reliability/Brier, for NucleCast and for FlowCast's members.

  • Both ensembles are under-dispersed. Spread/RMSE (ensemble spread over the error of the ensemble mean; 1 if the spread matches the error) is 0.51 for NucleCast and 0.54 for FlowCast, and both rank histograms are symmetric U-shapes.
  • CRPS. NucleCast's fair CRPS is 27% below the anchor's MAE but 9% above FlowCast's.
  • Brier skill over the deterministic anchor is positive at VIL 133, 160 and 219, and below FlowCast's.

We therefore describe NucleCast's ensemble as a low-cost probabilistic product with dispersion comparable to FlowCast's, not a calibrated one, and we removed the "multimodal" claim. Table 2 now reports the fair ensemble CRPS instead of the Gaussian-fit estimator the reviewer pointed out.

Q3. Cost evidence: training cost and the efficiency contribution of residual refinement.

Appendix A.12 now lists the training cost of every model as run (Table 16) and end-to-end inference latency (Table 17).

  • Latency. For eight members, NucleCast takes 0.92 s and FlowCast 15.25 s, a factor of 17. Run at FlowCast's 384² output grid, NucleCast takes 5.87 s, a factor of 2.6. One DiffCast draw takes 15.03 s with the same harness, and eight sequential draws take 120 s.
  • Training. On the same 4×H100 node, measured step time × schedule gives 36.8 h for NucleCast's full anchor and refiner schedules against 118.3 h for FlowCast's latent model alone, 3.2× more, without counting the training of its VAE (Table 15). The appendix states the remaining asymmetries.
  • Residual refinement's contribution. At an identical inference budget, regenerating the full field instead of refining the anchor loses 0.0105 CSI-M (Figure 4).
  • Step count. Table 5 shows that PM8's categorical skill plateaus from four evaluations.

Q4. Does feedback outperform output-only refinement under the same protocol?

Yes. Appendix A.8 now reports both NucleCast modes, feedback and output-only refinement, under the same 12→36-lead protocol on the same 96 events, with the released checkpoint and paired tests.

  • Feedback beats output-only refinement by +0.021 CSI-M over three hours [+0.016, +0.027].
  • Output-only refinement improves on the anchor for two hours, then falls below it (−0.004, interval excluding zero).
  • Replication. A second (20-epoch) checkpoint gives the same picture (+0.026 for feedback); there, output-only refinement falls below the anchor from the second hour.

We are grateful for these questions; the paper's claims are now tied to direct evidence.

Reviewer_ch5ZOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents NucleCast, a two-stage probabilistic radar nowcasting framework that combines a frozen Mamba-3-based deterministic forecaster (STVMamba3) with a conditional flow-matching model for residual refinement in pixel space. The flow-matching model adopts a clean-residual prediction parameterization, where the predicted residual is analytically converted into a velocity field for sampling. It further employs a structured source distribution combining Gaussian noise with mask-guided compound-Poisson proposals to introduce localized precipitation variations. The framework uses four Euler steps per ensemble member and aggregates eight independently generated forecasts through probability-matched mean (PM-mean). Experiments on SEVIR and ARSO demonstrate a favorable efficiency–accuracy trade-off compared with existing generative baselines, alongside analyses of newly emerging precipitation events, off-support generation, and the effects of varying the number of sampling steps.

Critical strengths
  1. Computational Efficiency: The paper presents a lightweight probabilistic nowcasting framework that combines a deterministic Mamba-3 backbone with few-step residual flow matching, achieving substantially lower inference latency and parameter counts than existing generative baselines.

  2. Comprehensive Evaluation: The experiments cover two radar datasets (SEVIR and ARSO) and a broad range of metrics, including probabilistic forecasting quality, extreme precipitation detection, spatial structure, and computational cost.

  3. Insightful Additional Analysis: The studies on sampling steps (NFE), long-horizon feedback, and the trade-off between newly recovered precipitation events and false predictions provide useful insights into the behavior and limitations of the proposed framework.

Critical weaknesses
  1. Limited methodological novelty. The core idea of combining deterministic forecasting with generative residual refinement has already been explored in DiffCast, while clean-data prediction and pixel-space flow matching have also been investigated in prior work. NucleCast mainly combines these existing techniques with a lightweight Mamba-based architecture and a structured noise source. The paper does not sufficiently clarify its distinct methodological contributions beyond this integration.

  2. Insufficient motivation and justification for key design choices. The paper does not clearly explain why clean-residual prediction is preferred over direct velocity prediction in the flow-matching framework. Although the mathematical relationship between these parameterizations is discussed in the appendix, their practical differences and the specific benefits of clean-residual prediction for radar nowcasting remain unclear. Furthermore, the lack of direct ablation studies makes it difficult to assess whether this design meaningfully contributes to the reported performance.

  3. Poor presentation and figure quality. The presentation lacks clarity in communicating the core methodology and key innovations. In particular, the main framework figure (Figure 1) contains excessive textual information, inconsistent font styles and sizes, and densely packed annotations, resulting in a cluttered and visually disorganized layout. The lack of a clear visual hierarchy makes it difficult to distinguish the main components and follow the proposed pipeline. The figures should be substantially improved with more concise annotations, consistent typography, and clearer organization to effectively communicate the methodology.

Questions for authors
  1. Methodological novelty: Given that DiffCast has already explored generative residual refinement and PixelFlowCast has investigated pixel-space clean prediction with deterministic guidance, what are the key methodological innovations of NucleCast beyond combining these existing techniques? Could the authors clarify the distinct contributions and provide more direct comparisons with these closely related methods?

  2. Motivation for clean-residual prediction: Why is clean-residual prediction preferred over direct velocity prediction in this framework? Although their mathematical relationship is discussed in the appendix, the paper does not sufficiently explain the practical advantages of this parameterization. Could the authors provide a controlled ablation comparing clean-residual and direct velocity prediction under identical model architectures and training settings?

  3. Effectiveness of the structured source: How much does the proposed mask-conditioned compound-Poisson source contribute to the final performance compared with standard Gaussian noise? Could the authors provide quantitative ablations and clarify the sensitivity to key source parameters, such as proposal rate, amplitude, and spatial scale?

  4. Justification for few-step sampling: Why is NFE = 4 selected as the default sampling configuration? Could the authors provide comparisons with fewer sampling steps (e.g., NFE = 1 or 2) and further explain why increasing NFE leads to performance degradation? This would help clarify whether the reported efficiency results from the proposed design or the particular sampling configuration.

  5. Clarity and presentation: Could the authors improve the explanation and visualization of the proposed framework, particularly Figure 1? The figure contains excessive textual information, inconsistent font styles and sizes, and densely packed annotations, making it difficult to follow the overall pipeline and identify the key innovations. A more concise and consistently formatted illustration would significantly improve the clarity of the paper.

4 comments

We thank the reviewer for the insightful questions. We ran an experiment for each of Q1–Q4; for Q1 the comparison is an in-house ablation, because PixelFlowCast has no public code.

Q1. Methodological novelty beyond DiffCast and PixelFlowCast; direct comparisons.

The reviewer is right that DiffCast already generates residuals and that PixelFlowCast already uses pixel-space clean prediction around a deterministic forecast. Section 2 now says so explicitly and lists what NucleCast does differently:

  • an anchor-conditioned compound-Poisson source instead of Gaussian noise;
  • an anchor clipped at a core threshold, with its core mask as conditioning;
  • a 9.4 M-parameter Mamba-3 first stage;
  • four network evaluations per member, with probabilistically verified ensembles.

We found no public code or checkpoints for PixelFlowCast (its paper links none), so we isolated the differences in-house, under identical architecture, data and training. Figure 4 (exact values in Appendix Table 12) reports the gain of each choice, NucleCast − ablated variant, for PM8 on SEVIR Cohort B, with 95% intervals:

Choice (ablated variant)gain in CSI-Mother gains
Residual target (full-field target, anchor kept as conditioning)+0.0105 [+0.0086, +0.0117]CSI@219 +0.0125
Anchor clip (unclipped residual, as in PixelFlowCast)+0.0044 [+0.0031, +0.0055]POD@219 +0.016
Clean prediction (velocity head)+0.0003 (PM8 tie)single draws +0.0077; the velocity head adds sampler noise and 3.3× the extreme area through false alarms
Structured source (Gaussian), deployed epoch 16+0.0031 [+0.0011, +0.0048]CSI@219 +0.0126
Structured source (Gaussian), final epoch 49−0.0002 (tie)CSI@219 +0.0107 (not resolved)

The first stage is compared too: STVMamba3 beats SimVP-gSTA, from the SimVP family that PixelFlowCast uses as its first stage, by +0.0126 CSI-M with 5× fewer parameters (Table 2). Two of these differences (source, clip) are ablated inside the cascade; the first stage is compared as a standalone forecaster, and the four-step sampler is characterised by the NFE sweep (Q4) but not against MeanFlow. We compare with DiffCast directly in the main-text Tables 2–3 and in the long-horizon study.

Q2. Why clean-residual rather than velocity prediction? Controlled ablation.

We trained a velocity-head twin with identical architecture and training. Its PM8 CSI-M ties NucleCast's because averaging hides the noise, but each member shows the failure:

  • single-draw CSI-M −0.0077: clean prediction keeps single draws skilful;
  • high-frequency power 0.55 against 0.33: uncancelled sampler noise in the terminal state;
  • 3.3× the extreme area (bias@219 1.21 against 0.36) through false alarms (FAR@219 0.83 against 0.69), with no gain in CSI@219.

The mechanism (Section 3.3, Appendix A.1.1): with either head, the last Euler step at tN=0.999t_N = 0.999 moves 99.6% of the way to the implied clean estimate XN−1+(1−tN−1)vX_{N-1} + (1-t_{N-1})v. Clean prediction supplies the −XN−1-X_{N-1} term of v=(R^−Xt)/(1−t)v = (\hat R - X_t)/(1-t) analytically, so XN=0.004 XN−1+0.996 R^N−1X_N = 0.004\,X_{N-1} + 0.996\,\hat R_{N-1} cancels the source noise left in XN−1X_{N-1} by construction. A velocity head, XN=XN−1+Δt vθX_N = X_{N-1} + \Delta t\,v_\theta, must learn to cancel that noise through its output, and the twin shows what it leaves. Its wider ensemble (spread/RMSE 0.57 vs 0.47; spread/RMSE is the ensemble spread divided by the error of the ensemble mean, 1 for a well-dispersed ensemble and below 1 when it is over-confident) gives a slightly lower fair CRPS (4.19 vs 4.21 raw VIL), at the cost of single-draw skill.

Q3. Contribution of the structured source and sensitivity to its parameters.

We trained an identical 50-epoch twin with Gaussian noise as the only change and compared the two at matched epochs:

  • CSI@219 is higher with the structured source at both epochs: +0.0126 [+0.0061, +0.0186] at the deployed epoch 16 (11% relative) and +0.0107 at the final epoch 49 (interval including zero).
  • CSI-M improves by +0.0031 [+0.0011, +0.0048] at epoch 16 (+0.0088 for single draws) and ties at the final epoch. Epoch 16 was selected on validation for NucleCast, not for the twin; the final epoch involves no selection.
  • At epoch 16 the Gaussian twin over-forecasts extreme area (bias@219 +0.66), with more false alarms (FAR@219 +0.085); at the final epoch the structured source keeps the larger extreme-area budget (bias@219 +0.06, interval excluding zero).

The source therefore acts where it was designed to act, on extreme cores; general skill comes from the residual target and the anchor clip, and the paper now states the claim in exactly these terms. Appendix Table 13 adds the source parameters, each changed once in a 20-epoch retraining and compared with the 20-epoch NucleCast on Cohort B (variant − NucleCast):

  • Spatial scale ×2 (30 instead of 15 km): −0.0072 CSI@219 [−0.0127, −0.0012] and −0.0029 CSI-M, with less extreme area (POD@219 −0.016): wider proposals spread intensity instead of seeding cores.
  • Amplitude ×0.5 (×2 is impossible: the range already ends at the 255-VIL ceiling): CSI@219 unchanged (−0.0009 [−0.006, +0.006]); CSI-M +0.0019 [+0.0008, +0.0033], gained at VIL 133–181 with more false alarms.
  • Proposal rate: λ=64\lambda = 64 scores slightly above λ=192\lambda = 192 (development windows).

NucleCast is thus sensitive to blob scale and not to amplitude at the extreme threshold. One run per arm, so the intervals reflect test sampling only.

Q4. Why NFE = 4? NFE 1 or 2? Why does more NFE degrade?

Table 5 sweeps 1, 2, 3, 4, 6 and 8 evaluations per member with identical source draws:

  • One evaluation returns the network's estimate of the conditional mean. The members collapse (spread/RMSE 0.05), PM8 equals a single draw (CSI-M 0.3686 vs 0.3689), and extreme cores almost vanish (bias@219 0.14).
  • More steps monotonically raise dispersion, high-frequency power and extreme area.
  • PM8 CSI-M plateaus from 4 evaluations onward (0.3940; +0.0006 at 6, resolved but negligible; −0.0005 at 8), an empirical result of the sweep. Single-draw CSI-M peaks at 2–3 evaluations and then declines, because the added structure is often misplaced; the PM8 averaging absorbs it.
  • The ensemble's fair CRPS keeps improving slowly with more steps (4.21 at 4, 4.10 at 8, in raw VIL; 0.0165 and 0.0161 on Table 2's [0, 1] scale).

Four evaluations is thus the cheapest setting at which the ensemble reaches its categorical skill: 0.92 s for eight members, against 0.35 s at 1 evaluation and 1.68 s at 8. The efficiency comes from the residual design (Q1) rather than from the step count alone.

Q5 / W3. Presentation and Figure 1.

We redrew Figure 1 following the reviewer's points:

  • One pipeline in two rows: Stage 1 is the frozen deterministic anchor; Stage 2 is the residual flow refiner with its four-step loop.
  • Less text and consistent typography: one font in three sizes, one label and at most one sub-label per block, and no floating notes. The loss and colour bars were removed from the figure (the loss is Eq. 8), and the legend notes moved to the caption.
  • A clear hierarchy: one tinted band per stage. The network internals (former panel 3) are left to the appendix figure.
  • The key innovations marked: three badges (a–c) single out the clipped anchor with its core mask, the compound-Poisson source, and four-step clean-residual prediction, the three choices ablated in Figure 4.

We also corrected typos and made the method section state the actual time grid and the sampler identity.

We thank the reviewer again; the requested ablations substantially sharpened what the paper claims. We hope that the controlled evidence for each design choice, the redrawn Figure 1 and the corrected presentation address the soundness and presentation concerns.

Reviewer_ch5ZOpenReview ↗

Response after rebuttal

We thank the authors for their detailed response and additional experiments, which have addressed several of our concerns. However, our primary concern regarding methodological novelty remains unresolved. The proposed framework still appears to be largely an integration of existing techniques, particularly residual refinement from DiffCast and pixel-space clean prediction from PixelFlowCast. Although the authors provide component-wise ablations to highlight the differences from PixelFlowCast, these experiments do not fully establish the methodological advantages of NucleCast over this closely related approach. Therefore we will maintain our original score unchanged.

We thank the reviewer for the follow-up and are glad the new experiments addressed several concerns. On the remaining point, three facts may help the final assessment.

What is new relative to DiffCast and PixelFlowCast. We agree that residual generation (DiffCast) and pixel-space clean prediction (introduced by Li & He, 2026, on which PixelFlowCast also builds) are prior work, and Section 2 credits both. NucleCast adds what neither paper has: an anchor-conditioned compound-Poisson source instead of Gaussian noise, and a residual taken against an anchor clipped at a core threshold, with its core mask as conditioning. With a 9.4 M-parameter Mamba-3 first stage and four evaluations per member, an eight-member forecast takes 0.92 s, against 15.25 s for FlowCast. Each choice was ablated alone with everything else fixed (Figure 4, Tables 12–13): the residual target is worth +0.0105 CSI-M, the anchor clip +0.0044 CSI-M, and the structured source +0.0126 CSI@219 at the deployed checkpoint.

Why there is no head-to-head with PixelFlowCast. PixelFlowCast is an arXiv preprint (single version, 11 May 2026) with no released code or weights, so it cannot be run under our protocol. We therefore reproduced two of its design choices inside our pipeline, each changed alone: the Gaussian source and the unclipped residual both lose. Its SimVP-family first stage is compared as a standalone forecaster (STVMamba3 +0.0126 CSI-M, Table 2); its MeanFlow sampler and KAN-based conditioning network remain untested. We note that the ICLR 2027 reviewer guidelines do not require comparisons with papers solely on arXiv and state that "the lack of such comparisons cannot be a basis for rejection." We cite and discuss PixelFlowCast nonetheless, as the closest related work.

Reproducibility. Our code, trained weights and the configuration of every ablation, including the two source-parameter variants added during the rebuttal, are public in the anonymous repository, so any reader can rerun these comparisons, or add PixelFlowCast's remaining choices once its code is released.

We respect the reviewer's judgement on novelty and thank them again for the review, which led to the component ablations now in the paper.

Reviewer_aGtaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Many works consider a combination of deterministic and probabilistic forecasts. In these works, the deterministic models are giving a smooth prediction, with the crisp details filled in by generative models. This work proposes a refinement of such an approach, which is using a combination of Mamba-3-based forecaster for large-scale evolution with pixel-space conditional flow-matching model.

Critical strengths

Significance: The motivation of the proposed work is strong, leading to a good significance.

Quality: the experimental results generally (on the reservations, see the weaknesses) show that this work is competitive against the state-of-the-art. I have not spotted problems related to correctness.

Clarity: the paper is clearly written

Critical weaknesses

Originality/soundness concerns: the authors need to show why this approach with residual models is principally different from the state-of-the-art, and what the significant advantages are. I would invite the authors to say how they could address the motivation. For example,

Li et al (2026) Event-Guided Spatiotemporal Transformer with Conditional Diffusion Refinement for High-Intensity Precipitation Nowcasting, MDPI remote sensing present a similar motivation

Park et al (2026) Diffusion-Based Refinement for Kilometer-Scale Probabilistic Precipitation Nowcasting

I understand these works are recent and it would be right to argue these are competitive works, however, DiffCast also uses a similar strategy. Therefore, the clear message why the proposed approach is needed and how is it different from the previously proposed residual blending methods would be critical.

Quality: the argument that clearly says that the proposed method has an edge is not clearly visible. It does not mean that I would wish it to win over every benchmark, but stating the clear reason why the proposed method ensures, as stated in the claims, 'A compute-aware evaluation of stochastic refinement' and ' Efficient residual flow matching for probabilistic nowcasting.' Where the analysis claiming efficiency is shown, it seems incomplete: "FlowCast remains stronger on most general metrics, reaching 0.4263 CSI-M with PM8 compared to 0.3940 for NucleCast, but requires 13.05 s and 160.5 M parameters compared with 17.2 M for NucleCast. " But this does not say how much time is spent on the same task for NucleCast. Therefore, refining the message saying why the world needs this method/this analysis is crucial.

Questions for authors

Addressing the concerns above would exactly bridge the gap between the current state of the paper and its acceptance.

2 comments

We are grateful to the reviewer for acknowledging the clarity and motivation of our manuscript and for raising two constructive questions.

1. Why is residual refinement principally different from the state of the art (DiffCast; Li et al., 2026; Park et al., 2026)?

We agree that residual generation itself is not new, and Section 2 now says so explicitly. It positions NucleCast against DiffCast, PixelFlowCast and the two works the reviewer cites. Li et al. (EGN-Nowcast, Remote Sensing, Aug. 2026) and Park et al. (exPreCast-ENS, arXiv, 1 Sep. 2026) appeared within weeks of the submission deadline, so we discuss them as concurrent work.

The principal difference is how the residual is sampled. DiffCast, EGN-Nowcast and exPreCast-ENS refine with conditional diffusion samplers: DiffCast uses 250 DDIM steps (15.03 s per draw, timed with our harness on the same A100) and exPreCast-ENS 20 Heun steps (3.4 s per member at 1024×1024, as reported in their paper). NucleCast instead learns pixel-space flow matching on the residual of a core-clipped Mamba-3 anchor. The network predicts the clean residual R^\hat R, and the velocity follows analytically as vθ=(R^−Xt)/(1−t)v_\theta = (\hat R - X_t)/(1-t). At the fourth and last Euler step this gives XN=0.004 XN−1+0.996 R^N−1X_N = 0.004\,X_{N-1} + 0.996\,\hat R_{N-1}: the source noise left in XN−1X_{N-1} is cancelled analytically, whereas a velocity head must learn to cancel it. The NFE sweep shows that four evaluations reach the ensemble's categorical skill. One NucleCast draw takes 0.22 s including the anchor, and eight take 0.92 s.

To answer the question with evidence rather than argument, we added a controlled ablation (Figure 4). Under identical architecture, source, data and budget, we trained the refiner to generate the full future field while keeping the deterministic forecast and its core mask as conditioning, i.e. using the forecast as guidance instead of refining it. This loses 0.0105 PM8 CSI-M [0.0086, 0.0117] and 0.0125 CSI@219, and worsens the ensemble CRPS. Refining the anchor is therefore not equivalent to conditioning on it.

Two further choices are specific to NucleCast and are also ablated:

  • Clipping the anchor at a core threshold before residualization, which makes the refiner responsible for intense cores: removing it costs 0.0044 CSI-M [0.0031, 0.0055].
  • An anchor-conditioned compound-Poisson source: against an identically trained Gaussian-source twin it raises CSI@219 at both matched epochs (+0.0126 at the deployed checkpoint; +0.0107, not resolved, at the end of training); CSI-M improves by 0.0031 at the deployed checkpoint and ties at the end of training.

2. The efficiency analysis seemed incomplete.

The reviewer is right that this sentence named only FlowCast's time (NucleCast's 0.92 s was only in Table 2). We rewrote it, and in doing so re-timed FlowCast end to end with the same harness as NucleCast (Appendix A.12, Table 17). FlowCast's own timer excludes its VAE encode and decode; it gives 13.02 s on our re-run (13.05 s in the submission).

latency per 8-member forecast
NucleCast, native 128²0.92 s
FlowCast, end to end15.25 s
NucleCast, at FlowCast's 384² output grid5.87 s

So NucleCast is 17× faster end to end at native resolution, with 9× fewer parameters (17.2 M vs 160.5 M). Because FlowCast's output grid has nine times more pixels, we also ran NucleCast at that grid and measured the matched-resolution latency: 5.87 s, still 2.6× faster. Training cost is in Table 16. FlowCast's paper reports 4×H100 for 7 days for its latent model; NucleCast's refiner reached the released checkpoint after 14 B200-hours. Because device-hours across GPU generations are not comparable, we also measured training throughput on the same 4×H100 node (Table 15): NucleCast's full anchor and refiner schedules take 36.8 h, against 118.3 h for FlowCast's latent model alone, 3.2× more before its VAE is counted.

3. Why the world needs this method.

The use case is an operational centre that must issue probabilistic nowcasts every five minutes on modest hardware. The ARSO experiments (Slovenian Environment Agency) are that setting. An eight-member NucleCast forecast takes 0.92 s at SEVIR's 128×128 grid and about 5 s at ARSO's native grid (Table 17), far inside a five-minute cycle, and on ARSO it matches FlowCast's published CSI-M with the same ensemble mean (0.422 vs 0.420; 0.428 as PM8). The first contribution of the introduction now gives these compute numbers.

We thank the reviewer for the encouraging remark that these points bridge the gap to acceptance; we believe the revision now addresses both of them directly.

Reviewer_aGtaOpenReview ↗

The comments addressed partially

On 'why the world needs this method' and the efficiency analysis, I think it clarifies the claims, and I would like to thank the authors for that part.

However, on originality, I think my comments are also mirrored by the ones of Reviewer ch5Z. The rebuttal still does not convince that to achieve the goal of low-cost refinement of deterministic forecast anchor, the authors proposed something that goes beyond, essentially, a combination of DiffCast and PixelFlowCast. It is necessary but not sufficient to only give the numerical argument that the proposed method wins over the metrics (i.e., speed accompanied competitive performance), I would expect some insight into what constitutes a step change in understanding how to construct the low-cost refinement of deterministic anchor. The argument essentially goes that it works competitively, the individual architectural choices work, and it is much faster, without saying why this happens.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: new ablations, ensemble diagnostics and revised paper

We thank the three reviewers for careful and constructive reviews. They converge on one request: show which component of NucleCast earns its results. As requested by the reviewers, we conducted all additional experiments, except re-running PixelFlowCast (no public code). All new SEVIR ablations are paired on Cohort B with the same bitwise-identical anchor and 95% bootstrap intervals clustered by storm timestamp; the long-horizon study uses its own 96-event cohort, clustered by event.

The new evidence supports NucleCast's central claims and narrows two of them:

  • Refining beats regenerating. At the same architecture and inference budget, generating the full field instead of refining the anchor costs 0.0105 CSI-M.
  • NucleCast is efficient. An eight-member forecast is 17× faster than FlowCast's end to end (0.92 s against 15.25 s) and still 2.6× faster when NucleCast is run at FlowCast's 384² output grid. On the same 4×H100 node, its estimated training time is 3.2× shorter than that of FlowCast's latent model alone.
  • Every NucleCast-specific choice is ablated. The anchor clip, clean prediction and the structured source each contribute, and the paper now says exactly what each one buys. Where a claim went beyond this evidence (the word "multimodal", the calibration of the ensemble), we narrowed it.

New experiments

  1. Design-choice ablation (Section 4.3.1, Figure 4; exact values in Appendix Table 12). We remove one choice at a time and report the gain of each choice in PM8 CSI-M: NucleCast − ablated variant, with its 95% interval in brackets:
    • residual target, against a full-field target with the anchor kept as conditioning: +0.0105 [+0.0086, +0.0117];
    • anchor clip: +0.0044 [+0.0031, +0.0055];
    • clean prediction, against a velocity head: tie on PM8, but +0.0077 for single draws, which the velocity head leaves with high-frequency sampler noise and 3.3× the extreme area, through more false alarms;
    • compound-Poisson source, against an identically trained 50-epoch Gaussian twin: CSI@219 is higher at both matched epochs (+0.0126 at the deployed epoch 16, interval excluding zero; +0.0107 at the final epoch, not resolved); CSI-M gains +0.0031 at epoch 16 (selected for NucleCast) and ties at the final epoch.
    • source parameters (Table 13): doubling the blob scale costs 0.0072 CSI@219; halving the amplitude leaves it unchanged.
  2. Direct ensemble diagnostics (Section 4.3.2, Appendix A.10).
    • Fair ensemble CRPS, rank histograms, spread–skill and reliability/Brier, for NucleCast and FlowCast.
    • Table 2 now reports the fair ensemble CRPS. The submitted value was the CRPS of a Gaussian fitted to the ensemble's mean and spread (FlowCast's estimator), not of the ensemble itself.
  3. Sampler steps (Sections 3.3 and 4.3.4, Table 5). We sweep 1–8 network evaluations per member; the ensemble's categorical skill plateaus from four evaluations onward.
  4. Efficiency (Section 4.2, Appendix A.12).
    • End-to-end latency with one harness for all systems, at native and matched resolution.
    • A training-cost table, plus same-hardware training throughput on one 4×H100 node.
  5. Long-horizon feedback (Appendix A.8). In a three-hour rollout, refining each hour's output only versus feeding the refined forecast back as the next hour's input, under the same protocol, with paired tests.

Protocol harmonization and reproducibility updates

  • Training details. Section 4.1.3 now gives the full training record (4×B200, 50-epoch OneCycle, every epoch kept; the released epoch 16 was selected on validation by a rule fixed before test scoring) and reports the final epoch, so that refiner comparisons are made at matched epochs.
  • One timing harness for all systems. FlowCast's built-in timer covers only its latent sampling loops, and DiffCast's submitted 19.15 s used deterministic kernels (reproduced at 18.93 s). NucleCast, FlowCast and DiffCast are now timed end to end with one harness on one A100: eight members take 0.92 s for NucleCast and 15.25 s for FlowCast (13.02 s with its own timer); one DiffCast draw takes 15.03 s.
  • ARSO. The warm-started ARSO refiner is replaced by a from-scratch run with the SEVIR schedule (50-epoch OneCycle, read out at epoch 16) and a source rescaled to dBZ (white noise 25 dBZ, blobs 15 km, λ=18\lambda = 18). With the arithmetic ensemble mean FlowCast's paper reports, it matches FlowCast's published CSI-M (0.422 vs 0.420) with fewer false alarms at 39 dBZ (0.544 vs 0.589); as PM8 it reaches 0.428 CSI-M, 0.256 CSI@36 and 0.182 CSI@39 (Table 3 shows both means).
  • Long horizon. The rollout study now uses the released checkpoint of Tables 2–3; the earlier 20-epoch checkpoint is kept as a replication.
  • Presentation. Related work now positions NucleCast against DiffCast, PixelFlowCast and two concurrent works. Figure 1 is redrawn as one two-stage pipeline with consistent typography.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.