acceptodds
Under review as a conference paper at ICLR 2027

CloudFlow: Physical Evolution-Guided Flow Matching for Multispectral Geostationary Meteorological Satellite Data Nowcasting

Abstract

The increasing frequency of extreme weather events necessitates precise real-time nowcasting for effective disaster prevention and mitigation, a task that relies heavily on high spatiotemporal resolution observations from geostationary meteorological satellites (GMS). However, existing deep learning models face a persistent dilemma where many deterministic methods suffer from severe blurring, while many generative approaches are plagued by severe structural inconsistency and radiance intensity deviations. To overcome this dilemma, we propose CloudFlow, a two-stage framework that constructs a physics-inspired evolution prior and then performs physics-anchored conditional-flow refinement. In the first stage, the proposed Evolution Network is motivated by the continuity equation and implemented with a differentiable semi-Lagrangian transport operator, together with motion and source-sink parameterization. In the second stage, the proposed Physics-Anchored Flow Matching utilizes the coarse-grained prediction of the Evolution Network as a structural anchor to guide the reconstruction of fine-grained details. Extensive experiments on the Himawari-8/9 and GOES-16 datasets show that CloudFlow unifies predictive accuracy, visual fidelity, inference efficiency, robustness and evolution interpretability.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
geostationary satellite nowcasting, multispectral forecasting, spatiotemporal prediction, conditional flow matching, semi-Lagrangian transport, cloud evolution

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

The paper forecasts 8-band geostationary satellite imagery (Himawari-8/9 and GOES-16) three hours ahead at 30-min steps. CloudFlow has two stages. An Evolution Network first warps the last frame with a predicted velocity field and adds a learned source term, in a semi-Lagrangian style. A flow-matching model then starts from this coarse forecast and refines it with a few deterministic Euler steps. The paper reports the best MSE and FVD on both datasets, as well as results on a cross-region test set and several robustness tests.

Critical strengths
  1. Forecasting multispectral satellite imagery is useful where radar coverage is poor, and the method is tested on two satellites and an additional region.
  2. Splitting the evolution into motion and source terms makes the model easy to inspect, and the learned speeds look plausible (Fig. 12).
  3. The refined forecasts are much sharper than those of the baselines (lowest FVD), at far lower cost than MCVD and PreDiff.
  4. The appendix includes per-band results and tests on seasons, input noise and missing frames.
Critical weaknesses
  1. The task is satellite image forecasting, not precipitation nowcasting, yet the paper is motivated by extreme weather and rapid convection. At 2–4 km and 30-min steps, with daytime data only, it is unclear how much convective-scale information is left, and summer is the worst season (Fig. 22). The evaluation relies on image metrics only. MSE/MAE are averaged over bands with different units, there are no event-based scores, and no variance over runs is reported.

  2. Baseline training is not described at all (adaptation to 8 bands, epochs, tuning), and some numbers suggest under-trained baselines: MCVD reaches MSE 496.6 and 1274, and PreDiff 227.3 on GOES-16. Even the simplest stage-1 variant (SSD only in Table 2, MSE 53.73, SSIM 0.682) beats every baseline in Table 1. CloudFlow also has 193.7M parameters, while most baselines have under 35M.

  3. Both stages largely follow existing work. An evolution network that predicts a motion field and an intensity residual with semi-Lagrangian advection, followed by a generative refiner conditioned on its output, is very close to NowcastNet, which is not cited. Refining a deterministic nowcast with a flow has also been explored before (e.g., PercpCast), and image restoration methods such as InDI and I2SB also start the flow from a degraded estimate. PA-FM itself is built from standard parts.

  4. The physics part is weakly supported. Nothing is actually constrained: S is free and the velocity field is not regularized. The physically motivated DFI ordering gives higher MSE than the reversed ordering (44.55 vs. 43.76, Table 12), and α barely changes the results (Table 13), since S can absorb it. No ablation replaces the Evolution Network with a generic deterministic prior, so it is unclear whether the physics matters for the final model.

  5. The refinement stage trades accuracy for sharpness. In Table 11, MSE rises from 45.89 at N=1 to 53.87 at the default N=10 and keeps increasing with N. Table 2 also suggests that the Evolution Network alone reaches 44.55. This contradicts the claim that CloudFlow breaks the perception–accuracy trade-off. In addition, the output is deterministic and gives no uncertainty estimate, and the model fails when the latest frame is missing .

  6. The efficiency analysis only covers parameters and GFLOPs (Table 17). CloudFlow needs 3–7.5× the GFLOPs of SimVP, SimVPv2 and Earthformer for small gains (e.g., 0.7% lower MSE than SimVP). The 812× saving over PreDiff assumes 1000 sampling steps for PreDiff. Latency, memory and training cost are not reported. Table 10 lists CFG, which the text never mentions and which would double the refinement cost if used at inference.

Questions for authors
  1. How were the baselines adapted to and trained on the 8-band data?
  2. Can you report the stage-1 prediction in Table 1 under the same protocol, and state the evaluation setting of each ablation?
  3. What happens if PA-FM starts from a generic deterministic prior (e.g., SimVP) instead of the Evolution Network?
  4. How does CloudFlow compare with NowcastNet and with deterministic-plus-generative refiners such as DiffCast, CasCast or PercpCast?
  5. Is CFG used at inference? Please also report latency, memory and training time, and explain how N was chosen for each dataset.
  6. Can you add event-based scores (e.g., CSI on cold cloud-top thresholds) and results at night? If these points are addressed, I am willing to raise my score.
6 comments
AuthorsOpenReview ↗

Rebuttal Response 1/6

Dear Reviewer KZFz,

We sincerely thank you for your careful review and for recognizing the forecasting task, inspectable evolution components, and sharper predictions. To further address your concerns, we provide the following clarifications and additional experimental evidence. All changes described below have been incorporated into the revised manuscript. All section, table, and figure references below refer to this revised version.

1. Baseline adaptation, training, and convergence (W2, Q1)

Thank you for suggesting that we provide implementation details. Appendix D.3 now documents baseline adaptation and training. All baselines share evaluation data and forecast targets: six eight-band frames from six historical frames at 256 × 256. Adaptations differ as follows.

  • ConvLSTM/PredRNN. Recurrence uses a 32 × 32 grid with bilinear down/up-sampling to avoid full-resolution memory failure, limiting fine-scale information.
  • Earthformer/SimVP. Earthformer uses [T, H, W, C]; SimVP uses spatial/temporal depths 4/8. Two-dimensional backbones merge time and bands where needed.
  • DGMR. A shared single-band network processes the last four frames with bands folded into the batch, excluding joint eight-band modeling and full six-frame context.
  • Refinement baselines. DiffCast/CasCast freeze the listed SimVP; PercpCast uses official Phy-DNet, distinct from the standalone PhyDNet row.

Internal resolution, context, capacity, and budgets differ. The ten original deterministic baselines use 100 epochs, AdamW, effective batch size 64, and warm-up/cosine decay; Appendix D.3 supplies model-specific settings.

To address your concern of insufficient training of PreDiff and MCVD, we added loss curves (Figures 9–10), which flatten substantially on both datasets without establishing optimal tuning. On GOES-16, MCVD is trained for 1,000,000 steps; PreDiff’s VAE/diffusion trained 250/480 epochs. These figures show that the losses have converged.

The source-sink-only comparison alone is insufficient to assess baseline convergence. NowcastNet$^{\mathrm{S1}}$ has lower MSE than this variant, while added competitive refiners extend the comparison beyond MCVD and PreDiff.

2. First-stage results and ablation settings (Q2)

Thank you for suggesting that we report first-stage results. Table 1 now includes our CloudFlow$^{\mathrm{S1}}$ under the full-model protocol. Section 3 gives Himawari results; ablations follow.

  • Evolution Network (Table 3). Added motion transport, gradient inputs, and DFI progressively improve pixel and structural scores, although DFI increases FVD. All variants run without PA-FM on Himawari-8/9. SSD, MD, GI, and DFI denote the source-sink decoder, motion decoder, gradient input, and Dynamics Feature Integration.
ConfigurationChangeMSE ↓\downarrowFVD ↓\downarrow
SSDSource-sink-only branch53.73315.1
SSD + MDAdd motion transport47.41140.3
SSD + MD + GIAdd spatial gradients45.20117.7
SSD + MD + GI + DFICouple motion descriptors to intensity correction44.55130.4

Our added common-width test favors our Evolution Network over NowcastNet$^{\mathrm{S1}}$ in MSE, MAE, PSNR, FVD, and CSI (Appendix E.8, Table 19). Both use 32 base channels and 100 epochs.

  • PA-FM (Table 4). Forecast initialization improves all five metrics over Gaussian noise under fixed historical conditioning (I/V). All variants use Himawari-8/9, 1,000,000 training steps, and N=10N=10, varying the starting state and condition.
VariantStarting stateConditionMSE ↓\downarrowFVD ↓\downarrow
IGaussian noiseFpast\mathcal F_{past}79.5078.49
IIGaussian noiseF^phy\hat{\mathcal F}_{phy}65.8965.41
IIIGaussian noiseFpast\mathcal F_{past} and F^phy\hat{\mathcal F}_{phy}71.6384.31
IVFpast\mathcal F_{past}F^phy\hat{\mathcal F}_{phy}78.20110.3
VF^phy\hat{\mathcal F}_{phy}Fpast\mathcal F_{past}53.8767.56

Variant V leads pixel/structural metrics; II has slightly lower FVD. II/V and IV/V change both inputs, so neither isolates initialization nor establishes a conditioning-weight trend.

AuthorsOpenReview ↗

Rebuttal Response 2/6

3. Prior work, modern comparators, and model capacity (W2, W3, Q4)

Following your suggestions, we added NowcastNet, DiffCast, CasCast, and PercpCast comparisons; our CloudFlow has lower pixel errors than these complete refiners (Table 1). The revision credits their designs and clarifies the differences.

Differences from NowcastNet's evolution network (Section 3.1 and Appendices B and D.3):

  • Motion–intensity coupling. NowcastNet’s Evolution Network predicts displacement and intensity residuals through parallel decoders with shared encoder features. Our Evolution Network also uses DFI to supply the intensity decoder with speed, deformation, vorticity, and divergence features derived from the predicted displacement field.
  • Displacement field decomposition. NowcastNet’s motion decoder directly predicts a dense displacement field. Ours combines global translation and rotation from a rigid-motion head with a dense residual from a local decoder.
  • Gradient input for cloud-boundary. NowcastNet’s encoder takes historical frames as input. Ours also takes their spatial gradients, giving both decoders explicit cloud-boundary cues.
  • Attention gate skip connections. NowcastNet’s decoders concatenate encoder features directly with upsampled decoder features. Our local displacement decoder uses attention gates to select encoder features before fusion.
  • Evolution update. NowcastNet’s Evolution Network uses nearest-neighbor warping during recursion and bilinear warping for supervision. Ours uses bilinear warping during recursion and refines each updated frame with a 5×5 depthwise convolution followed by a 1×1 convolution.

At the same base width of 32 base channels and training budget of 100 epochs, our Evolution Network outperforms NowcastNet$^{\mathrm{S1}}$ on MSE, MAE, PSNR, FVD, and cold-cloud CSI (Table 19). These gains persist after matching width, although our model uses more parameters and GFLOPs and has slightly lower SSIM.

ModelMSE ↓\downarrowMAE ↓\downarrowPSNR ↑\uparrowSSIM ↑\uparrowFVD ↓\downarrowCSI235 K10.4 μm ⁣↑\mathrm{CSI}_{235\,\mathrm{K}}^{10.4\,\mu\mathrm{m}}\!\uparrowParams (M) ↓\downarrowGFLOPs ↓\downarrow
Evolution Network47.183.37324.140.683111.40.49011.9760.74
NowcastNet$^{\mathrm{S1}}$51.983.53023.790.690124.70.4588.8951.62

Relative to existing refinement methods (Introduction and Section 3.2 / 5.1):

  • DiffCast/CasCast. These cascades refine deterministic forecasts with diffusion; our PA-FM integrates conditional flow from the evolution forecast itself.
  • PercpCast. Both start from deterministic forecasts. Our PA-FM jointly processes leads with clean-endpoint parameterization; PercpCast refines leads separately with LPIPS supervision. Figure 3 compares their step-dependent behavior.
  • InDI/I2SB. Following your literature suggestions, the Section 3.2 Physics now cites InDI [1] and I2SB [2] as informative-input image-restoration precedents.

Selected Himawari-8/9 results from Table 1 follow.

ModelMSE ↓\downarrowMAE ↓\downarrowPSNR ↑\uparrowSSIM ↑\uparrowFVD ↓\downarrowCSI235 K10.4 μm ⁣↑\mathrm{CSI}_{235\,\mathrm{K}}^{10.4\,\mu\mathrm{m}}\!\uparrowRank ↓\downarrow
SimVP54.254.17123.710.637264.40.3887
NowcastNet$^{\mathrm{S1}}$51.983.53023.790.690124.70.4583
PercpCast$_{N=2}^{\ddagger}$74.124.85522.910.63552.150.4816
DiffCast$^{\dagger}$70.054.30922.510.60829.110.4248
CasCast$^{\dagger}$70.864.68022.660.61590.920.38913
NowcastNet67.104.82122.830.626692.30.43616
CorrDiff64.744.44322.890.62124.900.4794
CloudFlow$^{\mathrm{S1}}$44.553.29624.420.701130.40.5041
CloudFlow$_{N=10}$53.874.12923.790.67267.560.4912

S1^{\mathrm{S1}}: first stage; †^{\dagger}: listed SimVP backbone; ‡^{\ddagger}: official Phy-DNet. Ranks follow full Table 1. Newly added stochastic models use the mean of separately evaluated scores over three sampling seeds, not ensemble-mean predictions.

The added shared-backbone comparison shows our PA-FM best preserving MSE/PSNR/SSIM/CSI; DiffCast leads MAE/FVD (Table 2). Refiners use the same Evolution Network and 200 epochs, with architectures and compute differing. DiffCast and PercpCast attain lower FVD.

AuthorsOpenReview ↗

Rebuttal Response 3/6

4. Physics support, DFI, α, and a generic prior (W4, Q3)

To clarify, our Evolution Network is inspired by physical advection, without claiming physical consistency (Section 3.1; Appendix H). Its displacement field and source term SS are learned operator intermediates optimized for prediction.

  • DFI ordering. The proposed order improves MAE/SSIM; reversing it lowers MSE (Table 14). This supports an empirical compromise, not a unique physical ordering.
  • Source scale. Sensitivity to α initialization is modest (Table 15). A small initial value favors early transport, but the unconstrained source branch can absorb scale changes, precluding guaranteed sparsity or physical bounds.
  • Motion diagnostics. To address your physics concern, our added comparison with operational Himawari-9 atmospheric motion vectors (AMVs) supports the plausibility of the learned displacement fields. Displacement-derived velocities show broad directional agreement and lower RMSVD than zero motion in all four selected 10.4 μm cases (Appendix E.6; Table 17; Figures 19–23).
DateRMSVD (m/s) ↓\downarrowZero-motion RMSVD (m/s) ↓\downarrowMedian angular error (°) ↓\downarrowSpeed bias (m/s)
2023-01-1510.85823.4448.290-1.776
2023-06-1513.60829.2375.063-4.172
2023-07-1413.28117.75411.405-6.798
2023-10-1510.06512.54816.846-7.089

Speed magnitudes remain underestimated.

  • Source interventions. Added interventions show lower 10.4 μm MAE when retaining the source than removing/reversing it in each selected case, supporting post-transport correction (Table 18; Appendix E.7). Weights and motion are fixed.

Thank you for suggesting SimVP as a prior; in the added test, our Evolution Network + PA-FM performs better on all six metrics under the same refiner epoch count and inference steps. Our PA-FM lowers SimVP's FVD, with little CSI change and poorer pixel and structural scores. This supports the chosen prior's value in the complete system, while first-stage architecture, capacity, and forecast quality remain part of the comparison.

Table 20 (Appendix E.8) uses 200 refiner epochs and N=10N=10 for both priors; our Evolution Network retains its default base width 64.

ModelMSE ↓\downarrowMAE ↓\downarrowPSNR ↑\uparrowSSIM ↑\uparrowFVD ↓\downarrowCSI235 K10.4 μm ⁣↑\mathrm{CSI}_{235\,\mathrm{K}}^{10.4\,\mu\mathrm{m}}\!\uparrow
SimVP54.254.17123.710.637264.40.388
SimVP + PA-FM73.964.88922.730.618112.20.387
Evolution Network44.553.29624.420.701130.40.504
Evolution Network + PA-FM56.414.43723.600.66483.490.490

5. Refinement trade-off and deterministic output (W5)

Our CloudFlow model offers a favorable empirical perception–accuracy operating point among the evaluated models (Figure 1). Against SimVP, it maintains similar MSE with substantially lower FVD; against CorrDiff, it has lower MSE but higher FVD.

Refinement sacrifices pixel accuracy for lower FVD, as the first-stage rows in Table 1 show. Figure 1's connecting lines are comparisons, not a proven bound; refinement does not eliminate the accuracy–perception trade-off.

Our CloudFlow model targets deterministic prediction, with fixed initialization and integration yielding reproducible outputs. Flow matching describes the training objective, not a claim of probabilistic forecasting. The lack of uncertainty estimates limits risk-sensitive applications but is distinct from performance on the stated deterministic task. Point-mass CRPS equals MAE and does not measure uncertainty calibration (Appendix C).

Reviewer_p9qVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CloudFlow, a two-stage framework for multispectral satellite nowcasting. It combines physics-inspired cloud evolution modeling with Physics-Anchored Flow Matching (PA-FM) for prediction refinement. Experiments on Himawari-8/9 and GOES-16 demonstrate promising improvements in accuracy, perceptual quality, and computational efficiency. However, concerns remain regarding methodological novelty, physical validity, and experimental rigor.

Critical strengths

The paper addresses an important satellite nowcasting problem by balancing prediction accuracy, visual quality, and computational efficiency.

The two-stage design is intuitive, and the experiments cover multiple datasets, cross-region generalization, and robustness scenarios.

Critical weaknesses
  1. The framework combines existing motion modeling and flow-matching techniques, while its methodological advances over related approaches remain unclear.

  2. The learned velocity field is not quantitatively compared with independent atmospheric motion estimates, and the physical interpretation of the source-sink term lacks independent validation.

  3. Although experiments cover two satellite datasets, cross-region evaluation is conducted within the Himawari-8/9 domain. The transferability of a model trained on one satellite platform to another without retraining remains unclear. 

4.The evaluation does not include recent satellite nowcasting approaches, such as diffusion-based models published in 2026.

Questions for authors

How does PA-FM differ from existing conditional flow-matching methods, such as FlowCast (ICLR 2026)? Can the authors isolate the contribution of physics-inspired initialization through controlled comparisons?

Can the authors quantitatively validate the learned velocity field against independent atmospheric motion estimates and provide further evidence supporting the physical interpretation of the source-sink term?

4 comments
AuthorsOpenReview ↗

Rebuttal Response 1/4

Dear Reviewer p9qV,

We sincerely thank you for your careful review and for recognizing the nowcasting problem, two-stage design, and breadth of the evaluation. To further address your concerns, we provide the following clarifications and additional experimental evidence. All changes described below have been incorporated into the revised manuscript. All section, table, and figure references below refer to this revised version.

1. Methodological contribution and controlled comparisons (W1, Q1)

Our CloudFlow model couples the two stages through the forecast itself. Our Evolution Network uses its estimated motion to update cloud intensity after transport. Our PA-FM takes this forecast sequence as the starting state of its flow and refines all six lead times jointly.

First stage: differences from NowcastNet's evolution network

Both models use motion and intensity branches with semi-Lagrangian transport. Our Evolution Network changes how motion is estimated and how it informs the intensity update (Section 3.1 and Appendices B.1–B.4).

  • Motion–intensity coupling. NowcastNet’s Evolution Network predicts displacement and intensity residuals through parallel decoders with shared encoder features. Our Evolution Network also uses DFI to supply the intensity decoder with speed, deformation, vorticity, and divergence features derived from the predicted displacement field.
  • Displacement field decomposition. NowcastNet’s motion decoder directly predicts a dense displacement field. Ours combines global translation and rotation from a rigid-motion head with a dense residual from a local decoder.
  • Gradient input for cloud-boundary. NowcastNet’s encoder takes historical frames as input. Ours also takes their spatial gradients, giving both decoders explicit cloud-boundary cues.
  • Attention gate skip connections. NowcastNet’s decoders concatenate encoder features directly with upsampled decoder features. Our local displacement decoder uses attention gates to select encoder features before fusion.
  • Evolution update. NowcastNet’s Evolution Network uses nearest-neighbor warping during recursion and bilinear warping for supervision. Ours uses bilinear warping during recursion and refines each updated frame with a 5×5 depthwise convolution followed by a 1×1 convolution.

Component ablations. Adding motion transport, gradient inputs, and DFI successively improves pixel and structural scores (Table 3). DFI further reduces pixel error, although it increases FVD.

Comparison at matched width. To address your question about the first-stage contribution, we added a comparison using 32 base channels and 100 training epochs for both models (Appendix E.8, Table 19). Our Evolution Network improves MSE, MAE, PSNR, FVD, and cold-cloud CSI over NowcastNet$^{\mathrm{S1}}$, with slightly lower SSIM and higher parameter and GFLOP counts.

ModelMSE ↓\downarrowMAE ↓\downarrowPSNR ↑\uparrowSSIM ↑\uparrowFVD ↓\downarrowCSI235 K10.4 μm ⁣↑\mathrm{CSI}_{235\,\mathrm{K}}^{10.4\,\mu\mathrm{m}}\!\uparrowParams (M) ↓\downarrowGFLOPs ↓\downarrow
Evolution Network47.183.37324.140.683111.40.49011.9760.74
NowcastNet$^{\mathrm{S1}}$51.983.53023.790.690124.70.4588.8951.62

Second stage: differences from existing flow-matching methods like FlowCast and PercpCast

Our PA-FM starts from our Evolution Network’s deterministic forecast and jointly refines all six lead times in observation space (Section 3.2 and Appendix B.5).

  • Compared with FlowCast [1]. FlowCast starts from Gaussian noise in VAE latent space. Its cuboid-attention U-Net predicts the vector field directly and minimizes squared vector-field error. Our PA-FM starts from forecast cloud locations and intensities. Fixed initialization and integration yield one reproducible forecast, while FlowCast can sample multiple forecasts.
  • Compared with PercpCast [6]. Both methods start from deterministic forecasts. PercpCast refines each lead time separately using a vector-field loss with additional LPIPS supervision. Our PA-FM’s RefineFormer embeds patches with time folded into channels, allowing its DiT-style blocks to use information across all six lead times. It predicts the clean sequence, derives the vector field from that endpoint, and applies a robust loss to the resulting field.

With the same Evolution Network prior, our PA-FM achieves lower MSE and higher PSNR, SSIM, and cold-cloud CSI than PercpCast across the tested step counts. PercpCast achieves lower MAE and FVD (Section 5.1, Figure 3).

AuthorsOpenReview ↗

Rebuttal Response 2/4

Controlled refinement comparisons

  1. Initialization. Forecast initialization improves all five metrics over Gaussian noise under fixed historical conditioning (Table 4, I/V). Table 4 fixes training at 1,000,000 steps and inference at N=10N=10.
  2. Prior choice. We added an alternative-prior comparison to assess initialization; our Evolution Network prior outperforms SimVP on all six metrics with PA-FM (Table 20; Appendix E.8). Both refiners use 200 epochs and N=10N=10, supporting the complete evolution prior, including its structure and forecast quality.

Himawari-8/9 results (Table 20; Appendix E.8); our Evolution Network uses its default base width 64.

ModelMSE ↓\downarrowMAE ↓\downarrowPSNR ↑\uparrowSSIM ↑\uparrowFVD ↓\downarrowCSI235 K10.4 μm ⁣↑\mathrm{CSI}_{235\,\mathrm{K}}^{10.4\,\mu\mathrm{m}}\!\uparrow
SimVP54.254.17123.710.637264.40.388
SimVP + PA-FM73.964.88922.730.618112.20.387
Evolution Network44.553.29624.420.701130.40.504
Evolution Network + PA-FM56.414.43723.600.66483.490.490
  1. Refiner choice. Thank you for suggesting controlled comparisons. The added refiner test shows our PA-FM best preserving MSE, PSNR, SSIM, and cold-cloud CSI (Table 2). All use the same frozen Evolution Network and 200 Himawari-8/9 epochs; FlowCast’s autoencoder and flow stage each receive 200 epochs. Architectures and compute differ.
ModelMSE ↓\downarrowMAE ↓\downarrowPSNR ↑\uparrowSSIM ↑\uparrowFVD ↓\downarrowCSI235 K10.4 μm ⁣↑\mathrm{CSI}_{235\,\mathrm{K}}^{10.4\,\mu\mathrm{m}}\!\uparrowRank ↓\downarrow
+DiffCast58.983.87323.160.62524.780.4662
+FlowCast70.674.59322.800.63686.630.4105
+CasCast59.384.21823.420.64066.020.4693
+PercpCast$_{N=10}$72.764.42621.950.53654.830.4524
+PA-FM$_{N=10}^{200 \mathrm{E}}$56.414.43723.600.66483.490.4901

DiffCast achieves lower MAE and FVD. All refiners improve FVD while degrading the other metrics relative to our unrefined Evolution Network. The spectral analysis shows that our PA-FM reduces the first-stage power deficit at intermediate wavenumbers (8≤k≤648\le k\le64, Appendix F.5).

2. Independent motion comparison and source-term interpretation (W2, Q2)

To clarify, our Evolution Network is inspired by physical advection, without claiming physical consistency or recovery of atmospheric processes. “Evolution interpretability” concerns cloud-motion directions in intermediate displacement fields, supported by visualizations and a preliminary check of plausible motion speeds (Appendix E.5).

Comparison with atmospheric motion vectors

Thank you for recommending independent motion validation. Following your suggestion, we compare the predicted 10.4 μm\mu m band's displacement fields with Himawari-9 AMV products. As shown in Table 17 and Figures 19–23 of Appendix E.6, our new AMV comparison shows broad directional agreement and lower RMSVD than zero motion in all four selected cases, although it underestimates speed magnitudes.

DateRMSVD (m/s) ↓\downarrowZero-motion RMSVD (m/s) ↓\downarrowMedian angular error (°) ↓\downarrowSpeed bias (m/s)
2023-01-1510.85823.4448.290-1.776
2023-06-1513.60829.2375.063-4.172
2023-07-1413.28117.75411.405-6.798
2023-10-1510.06512.54816.846-7.089
AuthorsOpenReview ↗

Rebuttal Response 3/4

Functional evidence for the source-sink correction

Thank you for suggesting further tests of the source term. Two factors limit direct physical validation.

  1. No matching satellite product. Operational Himawari products do not directly observe this correction, which combines non-advective radiometric changes and transport/interpolation errors. It cannot be uniquely identified with condensation or evaporation rates.
  2. Existing evidence is indirect. L-CNN [3] uses optical-flow-aligned radar differences; LUPIN [4] jointly learns motion and advection-free changes. Both evaluate forecast skill. A NowcastNet follow-up [5] proposes optical-flow residuals as growth/decay targets. These motion-dependent residuals do not independently validate microphysical rates.

Our added interventions show that retaining the source gives the lowest MAE in every selected case, supporting an effective post-transport correction (Table 18; Appendix E.7).

Appendix E.7 tests the correction by retaining, removing, or reversing it with fixed weights, displacement/source sequences, and initial frame, without retraining. Interventions affect all eight bands; Table 18 reports 10.4 μm MAE (K) averaged over six leads.

DateOriginal sourceZero sourceReversed source
2023-01-152.9895.92111.039
2023-06-157.4029.37714.010
2023-07-147.5808.45111.345
2023-10-151.8883.1174.681
Equal-case mean4.9656.71610.269

3. Cross-satellite transfer without retraining (W3)

Following your suggestion, we added cross-satellite transfer tests; our CloudFlow$^{\mathrm{S1}}$ and the full model rank second and third, with the first stage leading SSIM and cold-cloud CSI.

Table 26 (Appendix G.2) reports direct transfer from Himawari-8/9 to GOES-16 without retraining or fine-tuning. Channel-wise normalization uses GOES-16 training-set minima and maxima.

ModelMSE ↓\downarrowMAE ↓\downarrowPSNR ↑\uparrowSSIM ↑\uparrowFVD ↓\downarrowCSI235 K10.35 μm ⁣↑\mathrm{CSI}_{235\,\mathrm{K}}^{10.35\,\mu\mathrm{m}}\!\uparrowRank ↓\downarrow
ConvLSTM245.5210.36120.560.698308.50.21810
PredRNN266.9010.90719.980.701316.80.24014
PhyDNet302.3511.39419.470.680294.90.21717
Rainformer248.0910.51220.630.662236.00.2368
Earthformer269.3611.22818.920.715314.10.3049
SimVP230.848.92520.860.679283.30.2885
Earthfarseer358.8511.78419.260.661272.20.21619
EPM221.279.99120.490.699294.80.2687
SimVPv2350.1413.30018.220.702295.40.32312
AlphaPre252.8810.31320.360.686238.20.2706
NowcastNet$^{\mathrm{S1}}$92.835.42423.870.728107.60.3741
PercpCast$_{N=2}^{\ddagger}$327.4311.54819.360.578144.80.22415
MCVD1266.6524.27112.110.458534.50.08821
DGMR100.635.72823.660.704313.30.3704
PreDiff315.4511.52319.560.585242.10.21816
DiffCast$^{\dagger}$441.7610.65818.080.606218.90.27211
CasCast$^{\dagger}$491.7413.70418.390.458337.40.18220
NowcastNet256.7910.58120.310.533713.20.23318
CorrDiff351.8411.29818.950.569147.50.24113
CloudFlow$^{\mathrm{S1}}$113.266.04423.190.762115.40.3962
CloudFlow$_{N=4}$195.478.96520.560.674132.80.3023

S1^{\mathrm{S1}} denotes the first stage; †^{\dagger} uses SimVP; ‡^{\ddagger} uses the official Phy-DNet configuration.

Refinement with our PA-FM is less effective under this cross-satellite shift.

AuthorsOpenReview ↗

Rebuttal Response 4/4

4. A recent diffusion-based satellite nowcasting baseline (W4)

Thank you for recommending recent satellite baselines; our added CorrDiff (2026) [2] comparison favors both stages of our CloudFlow model in MSE, MAE, PSNR, SSIM, and cold-cloud CSI. Our PA-FM improves first-stage perceptual quality while preserving these advantages.

Appendix D.3 documents the eight-band, six-step CorrDiff adaptation. The following Himawari-8/9 results are reproduced from Table 1.

ModelMSE ↓\downarrowMAE ↓\downarrowPSNR ↑\uparrowSSIM ↑\uparrowFVD ↓\downarrowCSI235 K10.4 μm ⁣↑\mathrm{CSI}_{235\,\mathrm{K}}^{10.4\,\mu\mathrm{m}}\!\uparrowRank ↓\downarrow
CorrDiff64.744.44322.890.62124.900.4794
CloudFlow$^{\mathrm{S1}}$44.553.29624.420.701130.40.5041
CloudFlow$_{N=10}$53.874.12923.790.67267.560.4912

S1^{\mathrm{S1}} denotes the first stage; ranks follow the complete Table 1 comparison.

CorrDiff has lower FVD and closer mean PSD agreement at 8≤k≤648\le k\le64 (Appendix F.5).

The expanded Table 1 comparison reported here covers Himawari-8/9. Despite our best efforts during the rebuttal period, the corresponding GOES-16 runs remain in progress because several diffusion- and flow-matching baselines require lengthy training. We appreciate your understanding.

Thank you again for your careful assessment of our work.

References

[1] Ribeiro, B. P., and Pucer, J. F. (2026). FlowCast: Advancing Precipitation Nowcasting with Conditional Flow Matching. International Conference on Learning Representations. arXiv: 2511.09731.

[2] Chase, R. J., Haynes, K., Ver Hoef, L., and Ebert-Uphoff, I. (2026). How to Use Score-Based Diffusion in Earth System Science: A Satellite Nowcasting Example. Artificial Intelligence for the Earth Systems, 5(3). doi: 10.1175/AIES-D-25-0046.1.

[3] Ritvanen, J., Harnist, B., Aldana, M., Mäkinen, T., and Pulkkinen, S. (2023). Advection-Free Convolutional Neural Network for Convective Rainfall Nowcasting. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 16, 1654–1667. doi: 10.1109/JSTARS.2023.3238016.

[4] Pavlík, P., Výboh, M., Ezzeddine, A. B., and Rozinajová, V. (2025). Fully differentiable Lagrangian convolutional neural network for physics-informed precipitation nowcasting. Applied Computing and Geosciences, 28, 100296. doi: 10.1016/j.acags.2025.100296.

[5] Brunet, D., Civitarese, D. S., and Davignon, D. (2025). An Improved Implementation of NowcastNet with Fine-Tuning for Canada. Extended Abstracts, 41st International Conference on Radar Meteorology, Toronto, Ontario, Canada, American Meteorological Society, Poster 7.

[6] Feng, W., Li, X., Wu, Z., Lin, K., Yu, D., Ye, Y., and Wang, Y. (2025). Perceptually Constrained Precipitation Nowcasting Model. Proceedings of the 42nd International Conference on Machine Learning, Proceedings of Machine Learning Research, 267, 16777–16795.

Reviewer_nqW7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CloudFlow, a two-stage framework for multispectral geostationary meteorological satellite (GMS) data nowcasting that predicts six future frames of 8-band Himawari-8/9 and GOES-16 observations at 30-minute intervals from six past frames. In the first stage, an Evolution Network motivated by the advective form of the continuity equation decomposes cloud evolution into a velocity field vv (a global rotation/translation component plus a local residual) and a source-sink term SS, and produces a coarse prediction through a differentiable semi-Lagrangian update X^t+1(p)=W(Xt,v)+αS(p)\hat{X}_{t+1}(p)=\mathcal{W}(X_t,v)+\alpha S(p), with gradient inputs and kinematic descriptors such as vorticity and divergence injected at scale-specific decoder stages. In the second stage, Physics-Anchored Flow Matching (PA-FM) starts from this coarse prediction F^phy\hat{\mathcal{F}}_{\mathrm{phy}} instead of Gaussian noise and learns a prior-to-data trajectory with an x-prediction parameterization, so that inference with N=10N=10 Euler steps is fully deterministic. The authors report the lowest MSE and FVD among ten deterministic and three generative baselines on both datasets, an approximately 812×812\times reduction in GFLOPs relative to PreDiff, and additional analyses on the interpretability of the learned motion fields, cross-region transfer, extreme weather cases, and robustness to noise and missing input frames.

Critical strengths
  1. Design that embeds physical intuition into the architecture: Motivated by the advective form of the continuity equation, the paper decomposes cloud evolution into motion (velocity field vv) and intensity change (source-sink term SS). This yields an interpretable intermediate representation rather than a black-box prediction.

  2. Flow matching initialized from an informative prior: Instead of starting from Gaussian noise, the refinement starts from the first-stage coarse prediction, so the model can focus on residual correction. The ablation in Table 3 systematically shows that using the prior as the starting distribution is more effective than using it only as a condition.

  3. Stability from deterministic outputs: Since inference involves no noise sampling, the model always produces the same output for the same input. This mitigates the instability of diffusion-based models, such as the mean shifts observed in MCVD and the structural fragmentation observed in PreDiff.

Critical weaknesses
  1. Missing key prior work and overstated novelty:   * Evolution Network: NowcastNet already proposed a neural implementation of advection based on the 2D continuity equation, in which the field is advected by motion fields and intensity residuals are then added. The first stage of CloudFlow follows nearly the same structure, and its unique components are limited to the gradient input, the global motion head, and DFI.   * Two-stage structure: DiffCast adds a diffusion-based residual to the coarse prediction of a deterministic backbone, and CasCast also combines deterministic and probabilistic components in a cascaded manner. The only substantive difference of CloudFlow is deterministic refinement starting from a prior instead of noise.

  2. Inconsistency between the claim of "breaking the perception-accuracy trade-off" and the paper's own results: In Table 2, the Evolution Network alone outperforms the final CloudFlow on all pixel-level metrics. This means that PA-FM sacrifices approximately 21% in accuracy to gain FVD, and Table 11 shows that NN directly controls this trade-off. Since the Evolution Network-only result is missing from the main table, this point is not visible.

  3. "Flow matching" framing and lack of uncertainty estimation: Since inference is fully deterministic, the model is essentially a learned iterative refiner. It cannot provide probabilistic forecasts, which is a key advantage of generative models, and while CRPS is reported for the generative baselines, CloudFlow is excluded from the probabilistic evaluation.

  4. Evaluation metrics not suited to meteorological applications: The only perceptual metric is FVD, which is based on I3D trained on natural videos, and no justification is given for applying it to 8-band satellite radiance fields. Without domain-specific metrics such as brightness-temperature-threshold-based CSI/POD/FAR or power spectra, it is impossible to tell whether the generated details are real structures or merely plausible textures. The claim of capturing extreme weather relies only on qualitative figures.

Questions for authors
  1. Differences from prior work: What are the structural differences between the Evolution Network of CloudFlow and the evolution network of NowcastNet? In addition, could you provide quantitative comparisons with NowcastNet, DiffCast, and CasCast on the same datasets?

  2. Isolating the contribution of PA-FM: What advantages does PA-FM provide compared with using the Evolution Network as the deterministic backbone of DiffCast or CasCast?

  3. Trade-off claim: Since the Evolution Network alone outperforms the final CloudFlow on all pixel-level metrics, how do you justify the claim of "breaking the perception-accuracy trade-off"? Could you include the Evolution Network-only results in the main table?

  4. Accuracy of the generated details: How can one verify that the high-frequency details produced in the refinement stage are consistent with the actual observed structures? Could you provide results based on brightness-temperature-threshold-based CSI/POD/FAR or power spectrum comparisons?

3 comments
AuthorsOpenReview ↗

Rebuttal Response 1/3

Dear Reviewer nqW7,

We sincerely thank you for your careful review and for recognizing the physical intuition, informative initialization, and reproducibility of deterministic inference. To further address your concerns, we provide the following clarifications and additional experimental evidence. All changes described below have been incorporated into the revised manuscript. All section, table, and figure references below refer to this revised version.

1. Differences from prior work and new baseline comparisons (W1, Q1)

Our CloudFlow model couples a structured evolution prior with deterministic flow refinement. We acknowledge NowcastNet's transport–intensity formulation and DiffCast/CasCast's refinement.

Differences between NowcastNet's Evolution Network and Ours are:

  • Motion–intensity coupling. NowcastNet’s Evolution Network predicts displacement and intensity residuals through parallel decoders with shared encoder features. Our Evolution Network also uses DFI to supply the intensity decoder with speed, deformation, vorticity, and divergence features derived from the predicted displacement field.
  • Displacement field decomposition. NowcastNet’s motion decoder directly predicts a dense displacement field. Ours combines global translation and rotation from a rigid-motion head with a dense residual from a local decoder.
  • Gradient input for cloud-boundary. NowcastNet’s encoder takes historical frames as input. Ours also takes their spatial gradients, giving both decoders explicit cloud-boundary cues.
  • Attention gate skip connections. NowcastNet’s decoders concatenate encoder features directly with upsampled decoder features. Our local displacement decoder uses attention gates to select encoder features before fusion.
  • Evolution update. NowcastNet’s Evolution Network uses nearest-neighbor warping during recursion and bilinear warping for supervision. Ours uses bilinear warping during recursion and refines each updated frame with a 5×5 depthwise convolution followed by a 1×1 convolution.

At the same base width of 32 base channels and training budget of 100 epochs, our Evolution Network outperforms NowcastNet$^{\mathrm{S1}}$ on MSE, MAE, PSNR, FVD, and cold-cloud CSI (Table 19). These gains persist after matching width, although our model uses more parameters and GFLOPs and has slightly lower SSIM.

ModelMSE ↓\downarrowMAE ↓\downarrowPSNR ↑\uparrowSSIM ↑\uparrowFVD ↓\downarrowCSI235 K10.4 μm ⁣↑\mathrm{CSI}_{235\,\mathrm{K}}^{10.4\,\mu\mathrm{m}}\!\uparrowParams (M) ↓\downarrowGFLOPs ↓\downarrow
Evolution Network47.183.37324.140.683111.40.49011.9760.74
NowcastNet$^{\mathrm{S1}}$51.983.53023.790.690124.70.4588.8951.62

Comparisons with new baselines

Thank you for suggesting NowcastNet, DiffCast, and CasCast. Our Evolution Network leads on MSE, MAE, PSNR, SSIM, and cold-cloud CSI (Table 1). Our PA-FM lowers FVD relative to our first stage, while our full CloudFlow model retains lower pixel errors than all three baselines. DiffCast achieves the lowest FVD among these three. Selected Himawari-8/9 results from Table 1:

ModelMSE ↓\downarrowMAE ↓\downarrowPSNR ↑\uparrowSSIM ↑\uparrowFVD ↓\downarrowCSI235 K10.4 μm ⁣↑\mathrm{CSI}_{235\,\mathrm{K}}^{10.4\,\mu\mathrm{m}}\!\uparrowRank ↓\downarrow
SimVP54.254.17123.710.637264.40.3887
NowcastNet$^{\mathrm{S1}}$51.983.53023.790.690124.70.4583
PercpCast$_{N=2}^{\ddagger}$74.124.85522.910.63552.150.4816
DiffCast$^{\dagger}$70.054.30922.510.60829.110.4248
CasCast$^{\dagger}$70.864.68022.660.61590.920.38913
NowcastNet67.104.82122.830.626692.30.43616
CorrDiff64.744.44322.890.62124.900.4794
CloudFlow$^{\mathrm{S1}}$44.553.29624.420.701130.40.5041
CloudFlow$_{N=10}$53.874.12923.790.67267.560.4912

S1^{\mathrm{S1}}: first stage; †^{\dagger}: listed SimVP backbone; ‡^{\ddagger}: official Phy-DNet. Ranks follow full Table 1. Newly added stochastic models over three sampling seeds.

AuthorsOpenReview ↗

Rebuttal Response 2/3

2. Isolating the contribution of PA-FM (Q2)

Thank you for suggesting controlled comparisons. As shown in the table below, our PA-FM best preserves MSE, PSNR, SSIM, and CSI with the lowest per-step GFLOPs in our added test.

To compare refiners on the same initial forecasts, we use our frozen Evolution Network (Table 2). Each refiner trains for 200 epochs on Himawari-8/9, with 200 epochs for each of FlowCast’s two stages.

ModelMSE ↓\downarrowMAE ↓\downarrowPSNR ↑\uparrowSSIM ↑\uparrowFVD ↓\downarrowCSI235 K10.4 μm ⁣↑\mathrm{CSI}_{235\,\mathrm{K}}^{10.4\,\mu\mathrm{m}}\!\uparrowParams (M) ↓\downarrowGFLOPs / step ↓\downarrowRank ↓\downarrow
+DiffCast58.983.87323.160.62524.780.46646.56276.602
+FlowCast70.674.59322.800.63686.630.410160.431,349.245
+CasCast59.384.21823.420.64066.020.469391.10256.293
+PercpCast$_{N=10}$72.764.42621.950.53654.830.45238.071,383.324
+PA-FM$_{N=10}^{200 \mathrm{E}}$56.414.43723.600.66483.490.490146.2451.751

DiffCast leads MAE/FVD. All refiners improve FVD but worsen other metrics versus Stage1. Our PA-FM uses 200 epochs here versus 1,000,000 updates in Table 1.

Our PA-FM and PercpCast refine deterministic forecasts with two main differences.

  • Sequence processing. Our PA-FM jointly processes leads for sequence interactions [B, T$\times$C, H, W]; PercpCast refines each separately [B$\times$T, C, H, W].
  • Learning target. Our PA-FM uses xx-prediction to parameterize its vector field; PercpCast uses the original vector field loss and additionally adds LPIPS loss for perceptual supervision.

3. The perception–accuracy claim and first-stage results (W2, Q3)

With first-stage results added as requested (Table 1), our CloudFlow model breaks existing methods’ empirical perception–accuracy boundary (Figure 1).

  • Existing methods. Our CloudFlow model’s FVD is below the NowcastNet$^{\mathrm{S1}}$–CorrDiff line segment at its MSE, breaking this empirical boundary.
  • Including our first stage. Our CloudFlow model also lies below the new Evolution Network–CorrDiff line-segment boundary.

Stage 1 achieves the lowest MSE of all evaluated methods. Its MSE lies outside the range spanned by the empirical boundary in Figure 1, so the boundary comparison applies to our full CloudFlow model.

Table 13 and Figure 11 further show how inference steps control the accuracy, perception, and computation trade-off.

4. Flow-matching terminology, deterministic inference, and CRPS (W3)

We recognize ensemble forecasting and uncertainty quantification as important advantages of probabilistic models. Our CloudFlow model targets deterministic forecasting, using flow matching to refine a fixed evolution forecast into a single reproducible prediction (Section 3.2). Ensemble generation is outside this prediction objective.

For a deterministic point forecast, CRPS reduces to absolute error, whose mean is MAE (Appendix C). Reporting CRPS would therefore duplicate MAE.

AuthorsOpenReview ↗

Rebuttal Response 3/3

5. Meteorological evaluation and generated details (W4, Q4)

  • FVD. We use FVD to compare the spatiotemporal structure of predicted and observed sequences alongside pixel errors and cold-cloud CSI. PreDiff [1] and FACL [2] also report FVD for precipitation nowcasting. Following PreDiff, we replicate each single-band frame across three channels for the pretrained RGB I3D network, use it without fine-tuning, and average FVD across satellite bands (Appendix C).
  • Cold-cloud CSI. Thank you for suggesting an event-based metric. We added cold-cloud CSI to measure the overlap between predicted and observed cold cloud tops, identified by 10.4 μm brightness temperatures of T_{b} ≤ 235 K [3]. Such cold cloud tops may indicate deep convection (Section 4.2 and Appendix C). On Himawari-8/9, our Evolution Network ranks first and our full CloudFlow model ranks second on this metric (Table 1).
  • Extreme weather. To address your concern, we expanded the Typhoon Mawar analysis in Figure 5 with 10.4 μm forecasts, cold-cloud masks, and curves for MSE, MAE, PSNR, SSIM, and CSI across forecast lead times. Our Evolution Network achieves the lowest mean MSE and the highest cold-cloud CSI through 90 min. Our full CloudFlow model achieves the second-lowest mean MSE, and its 10.4 μm forecasts show finer cloud structure after refinement. We omit MCVD from the curves because its much larger errors obscure differences among the other models.
  • Spatial spectra analysis. Following your suggestion, we added a power spectral density (PSD) analysis (Appendix F.5). Our PA-FM brings the first-stage spectrum closer to observations at intermediate wavenumbers (8≤ k≤64), supporting our claim that it refines cloud structure at these scales. At 8≤ k≤32, the spectrum of our CloudFlow model is also closer to observations than those of ConvLSTM, SimVP, PreDiff, and CasCast. Both stages of CloudFlow underestimate power at k≥64. The larger deficit after refinement at the highest wavenumbers may partly reflect grid effects from RefineFormer’s patch embedding and reconstruction.

The expanded Table 1 comparison reported here covers Himawari-8/9. Despite our best efforts during the rebuttal period, the corresponding GOES-16 runs remain in progress because several diffusion- and flow-matching baselines require lengthy training. We appreciate your understanding.

Thank you again for your careful assessment of our work.

References

[1] Gao, Z., Shi, X., Han, B., Wang, H., Jin, X., Maddix, D., Zhu, Y., Li, M., and Wang, Y. (2023). PreDiff: Precipitation Nowcasting with Latent Diffusion Models. Advances in Neural Information Processing Systems, 36, 78621–78656. doi: 10.52202/075280-3439.

[2] Yan, C.-W., Foo, S. Q., Trinh, V. H., Yeung, D.-Y., Wong, K.-H., and Wong, W.-K. (2024). Fourier Amplitude and Correlation Loss: Beyond Using L2 Loss for Skillful Precipitation Nowcasting. Advances in Neural Information Processing Systems, 37, 100007–100041. doi: 10.52202/079017-3173.

[3] Pan, L. L., Atlas, E. L., Honomichl, S. B., Smith, W. P., Kinnison, D. E., Solomon, S., Santee, M. L., Saiz-Lopez, A., Laube, J. C., Wang, B., Ueyama, R., Bresch, J. F., Hornbrook, R. S., Apel, E. C., Hills, A. J., Treadaway, V., Smith, K., Schauffler, S., Donnelly, S., Hendershot, R., Lueb, R., Campos, T., Viciani, S., D'Amato, F., Bianchini, G., Barucci, M., Podolske, J. R., Iraci, L. T., Gurganus, C., Bui, P., Dean-Day, J. M., Millán, L., Ryoo, J.-M., Barletta, B., Koo, J.-H., Kim, J., Liang, Q., Randel, W. J., Thornberry, T., and Newman, P. A. (2024). East Asian summer monsoon delivers large abundances of very short-lived organic chlorine substances to the lower stratosphere. Proceedings of the National Academy of Sciences, 121(12), e2318716121. doi: 10.1073/pnas.2318716121.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.