acceptodds
Under review as a conference paper at ICLR 2027

Diffusion Precipitation Generation with Medium-range Weather Forecasting

Abstract

Accurate daily precipitation forecasts are important for weather-sensitive decisions. Although recent Weather Foundation Models (WFMs) have achieved significant progress, they are trained on reanalysis data, which exhibit regional and intensity-dependent biases. Since retraining WFMs on observation-based targets is costly, we study Dynamics–Precipitation Decoupling, a plug-and-play framework that uses a separately trained generator to map non-precipitation forecasts from a frozen weather model to observation-aligned daily precipitation. Its key challenges are to capture heterogeneous rainfall distributions and align daily accumulations with atmospheric signals that evolve across space and time. To address them, we train a diffusion generator conditional on atmospheric states and satellite-gauge estimates of precipitation. A multi-scale recurrent encoder summarizes atmospheric evolution, while cross-attention aligns atmospheric and precipitation features during denoising. Extensive evaluation of various backbones, including Pangu-Weather, GraphCast, and HRES, on global and regional forecasting shows improved daily precipitation accuracy and event detection, with better representation of localized accumulation patterns. Global ensemble precipitation forecasting results demonstrate that our decoupling models outperform ENS and training-from-scratch GenCast. Remarkably, a heavy-rainfall case study using 2,288 stations in China further supports improved daily accumulation patterns and event detection against ground observations. Code is available at https://anonymous.4open.science/r/ICLR2027-code-9B4C.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
precipitation generation, conditional diffusion, ensemble forecast

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_fF5rOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper trains a conditional diffusion generator on paired reanalysis atmospheric states and satellite precipitation estimates, then conditions it on forecasts from frozen weather models to produce daily precipitation. A multiscale recurrent encoder represents atmospheric evolution, and cross-attention guides denoising. The evaluation covers global and regional forecasts, architecture comparisons, and a station-based rainfall case study.

Critical strengths

Separating atmospheric forecasting from precipitation supervision has clear practical value. In “Cross-Backbone Reuse and Fine-Tuning Diagnostic,” the paper states that different forecast models at the same resolution use the same generator checkpoint and variable interface; the main evaluation also includes both data-driven and numerical forecasts. This supports the claimed reusability more directly than an improvement shown with a particular forecasting backbone.

Critical weaknesses

The method motivates cross-attention through a “nonlocal spatiotemporal alignment challenge” and argues that precipitation features can selectively attend to atmospheric conditions. In the configuration comparison, attention fusion outperforms gated fusion, and the complete model outperforms variants with modules removed. These results establish that the design is useful, but changing the fusion operator also changes how features interact and how flexibly the model can represent their relation. The comparison does not establish whether the alternative has comparable cross-location interaction ability, and it does not examine whether attention actually uses upstream or surrounding atmospheric locations. The performance difference therefore cannot distinguish the proposed spatial alignment mechanism from the benefit of a more flexible fusion operator, limiting attribution of a central contribution.

The station case study uses a heavy-rainfall event to support improved local accumulations and event detection, and its figure displays local maxima. Yet the reported critical success indices cover only relatively low rainfall thresholds. The appendix adds station maps at other times within the same event but does not report intensity errors in the high-rainfall range or missed and false heavy-rainfall events. Better detection at low thresholds could result from more accurate rain-area coverage even if the magnitudes of local peaks remain biased. Thus, the station evidence does not directly establish how much the method improves heavy rainfall accumulation, and visually sharper peaks should not be equated with more accurate heavy-rainfall forecasts.

Questions for authors

Could the authors provide a fusion comparison with comparable cross-location interaction range and model capacity, or analyze which atmospheric locations the attention actually uses, to clarify how much of the gain comes from the proposed spatial alignment mechanism?

Could the authors use the existing station observations to report intensity errors in the high-rainfall range and missed and false heavy-rainfall events, so that the apparent improvement in visualized peaks can be checked against actual heavy-rainfall forecast skill?

7 comments
AuthorsOpenReview ↗

Response for Reviewer fF5r (1/3)

We thank the reviewer for the constructive feedback and will incorporate the additional fusion ablations, attention visualizations, and gridded and station-based heavy-rainfall evaluations presented below into the revised manuscript.

W1 & Q1. Disentangling Spatial Alignment from Fusion Flexibility.

We added ablations to assess whether cross-attention provides greater gains than alternative fusion operators.

  • ConvGRU + attention (ours) uses ConvGRU to encode the atmospheric sequence and cross-attention to aggregate atmospheric features with input-dependent spatial weights.
  • ConvGRU + global-mixing CNN retains ConvGRU and replaces cross-attention with learned, input-independent global token mixing followed by CNN fusion.
  • Channel-stacked CNN + attention replaces ConvGRU with a CNN that jointly processes the two atmospheric states stacked along the channel dimension, while retaining cross-attention.

All three variants access the same module parameter size (12M) and the same global grid of 15 × 30 atmospheric tokens, providing comparable cross-location interaction coverage.

Across Days 1–10, ConvGRU + attention achieves higher CSI25 and FSS25 than both alternatives at every lead time under both backbones.

Table R1a. CSI at 25 mm/24 h, Days 1–10. Higher is better.

BackboneVariantDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCastConvGRU + attention0.32230.26460.24160.22480.20700.18920.17390.15880.14520.1336
GraphCastConvGRU + global-mixing CNN0.28800.23810.21800.20630.19240.17670.16310.15330.14110.1312
GraphCastchannel-stacked CNN + attention0.28270.23370.21130.20090.18680.17240.15730.14730.13750.1275
HRESConvGRU + attention0.34210.32000.30750.29280.27440.25330.23460.21670.20290.1881
HRESConvGRU + global-mixing CNN0.30610.28610.27480.26390.25030.23050.21330.19860.18560.1722
HRESchannel-stacked CNN + attention0.29930.27950.26870.25950.24840.22770.21090.19730.18770.1742

Table R1b. FSS at 25 mm/24 h, Days 1–10. Higher is better.

BackboneVariantDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCastConvGRU + attention0.57440.47170.42390.38270.34260.30230.26250.23470.19980.1763
GraphCastConvGRU + global-mixing CNN0.50170.40960.36400.33420.30090.26490.23040.21060.18070.1617
GraphCastchannel-stacked CNN + attention0.48340.39750.35050.32680.29540.26120.22510.20800.18200.1625
HRESConvGRU + attention0.59770.55400.52670.49320.45980.41450.37330.32900.29350.2598
HRESConvGRU + global-mixing CNN0.51470.47560.44820.42230.39590.35540.31830.28340.25190.2228
HRESchannel-stacked CNN + attention0.49750.45790.43130.40830.38690.34750.30940.27860.25170.2239

The comparison with global-mixing CNN supports input-dependent nonlocal fusion under comparable spatial coverage and parameter budgets. The comparison with channel-stacked CNN supports the effectiveness of ConvGRU-based temporal encoding.

W1 & Q1. Which atmospheric locations the attention actually uses?

To examine the spatial behavior of cross-attention, we visualize the atmospheric key locations attended to by rain-core queries in two precipitation events over South China and Pakistan in Figure 15 and 16 in revised manuscript. We will upload the revised revised manuscript as soon as possible.

W2 & Q2. Heavy-Rainfall Evaluation Using Gridded and Station Observations.

Thanks for the suggestion. We extend heavy-rainfall verification to 25 and 50 mm/day over Days 1–10, using gridded IMERG and the 2,288-station network.

In the gridded evaluation, our framework improves both CSI and FSS at every lead time under GraphCast, HRES, and Pangu conditioning (Tables R2a–R2d).

Table R2a. Gridded CSI at 25 mm. Higher is better.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.43580.38020.31450.28260.25450.23110.21030.18900.17190.1576
HRES0.40870.35020.33530.32990.30550.28120.25390.23250.21290.1980
GenCast0.40120.32520.32410.29120.24410.23620.22510.19070.18740.1744
ENS Mean0.39450.30150.25150.20600.16200.12470.09210.06900.05110.0388
GraphCast+Ours0.50000.45570.41250.36830.32750.30410.27510.25050.22850.2072
HRES+Ours0.48290.44870.42660.39870.37180.34300.31510.29020.26900.2478
Pangu+Ours0.47300.41170.38480.35860.32740.28950.25670.22650.20420.1837
AuthorsOpenReview ↗

Response for Reviewer fF5r (2/3)

Table R2b. Gridded CSI at 50 mm.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.24080.17440.14330.12100.10000.08770.07550.06600.05780.0519
HRES0.22130.17620.15510.13890.12600.11280.10160.08680.07740.0692
GenCast0.23510.16420.15890.14640.13850.12740.11430.09320.08690.0542
ENS Mean0.17480.10210.06950.04520.02510.01210.00590.00290.00170.0010
GraphCast+Ours0.34320.29430.24800.20600.17560.15520.13320.11730.10580.0941
HRES+Ours0.31780.27830.25630.23640.21520.19480.17360.15550.14100.1291
Pangu+Ours0.30730.24420.21920.19680.17500.14570.12050.10060.08540.0746

Table R2c. Gridded FSS at 25 mm. Higher is better.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.73030.65860.60660.55180.49580.44150.38850.34350.30420.2695
HRES0.73870.67150.62790.58060.51740.44370.37230.30680.26350.2239
ENS Mean0.73820.64360.57640.50020.41910.33100.24350.17340.12260.0820
GraphCast+Ours0.79750.73800.66640.61110.56490.51080.45710.40430.36230.3270
HRES+Ours0.76560.71550.67560.63320.58770.53550.47940.42500.37770.3397
Pangu+Ours0.75680.70070.65620.59010.52600.47550.41400.36480.31720.2776

Table R2d. Gridded FSS at 50 mm.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.53890.42840.36160.30850.25080.21560.17300.15180.12750.1065
HRES0.53420.45610.40420.36230.32570.28270.23720.20070.16770.1418
ENS Mean0.46640.33640.24860.16980.09460.04170.01490.00530.00270.0017
GraphCast+Ours0.59750.49690.44070.39700.35020.29660.24700.20790.17800.1492
HRES+Ours0.60500.54300.49940.45820.41450.36780.31590.27140.23280.2058
Pangu+Ours0.64830.57350.48560.40360.32970.26330.21420.16410.13240.1059

Across the 2,288-station network, RMSE, CSI, and FSS improve at 25 mm across all ten lead times under GraphCast and HRES conditioning. (Tables R3a–R3e).

Table R3a. Station RMSE for station-based evaluation.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast7.13257.25317.43427.72628.00998.34158.60378.89559.25869.6167
HRES7.17027.43797.78968.21538.40288.85249.04749.20059.69009.9643
GraphCast+Ours6.39106.59306.86367.25027.56427.72127.84758.26198.71159.1270
HRES+Ours6.46416.90267.36087.80148.06787.84868.22418.45468.89729.7440
Pangu+Ours6.46276.80887.12207.37747.53668.51638.74678.92759.32919.3852

Table R3b. Station CSI at 25 mm. Higher is better.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.22980.19810.19150.16970.15300.13260.12230.10550.08090.0707
HRES0.22150.21830.19880.17300.15290.11950.10740.10670.08520.0807
GraphCast+Ours0.30740.29120.25970.22630.19720.17360.14380.13190.09710.0812
HRES+Ours0.31530.28770.25610.21990.19450.15450.13700.12600.10170.0864
Pangu+Ours0.29670.26080.22070.17570.16000.13080.11570.10410.08450.0680

Table R3c. Station CSI at 50 mm.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.08440.09140.08440.06880.06050.05270.04070.03380.03020.0257
HRES0.08070.07570.07400.06080.04040.03810.02670.02420.02310.0226
GraphCast+Ours0.11230.11240.09610.06740.06300.05160.05060.04250.02840.0312
HRES+Ours0.10040.09960.07960.06290.06180.04670.03720.03550.03290.0294
Pangu+Ours0.09730.09900.06710.06390.05020.03900.03730.03320.02590.0275

Table R3d. Station FSS at 25 mm. Higher is better.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.37370.33070.32140.29020.26540.23420.21790.19090.14970.1321
HRES0.36270.35840.33170.29500.26520.21350.19400.19280.15700.1493
GraphCast + Ours0.47020.45110.41230.36910.32940.29580.25140.23310.17700.1502
HRES + Ours0.47940.44680.40780.36050.32570.26760.24100.22380.18460.1591
Pangu + Ours0.45760.41370.36160.29890.27590.23130.20740.18860.15580.1273

Table R3e. Station FSS at 50 mm

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.15570.16750.15570.12870.11410.10010.07820.06540.05860.0501
HRES0.14930.14070.13780.11460.07770.07340.05200.04730.04520.0442
GraphCast + Ours0.20190.20210.17530.12630.11850.09810.09630.08150.05520.0605
HRES + Ours0.18250.18120.14750.11840.11640.08920.07170.06860.06370.0571
Pangu + Ours0.17730.18020.12580.12010.09560.07510.07190.06430.05050.0535
AuthorsOpenReview ↗

Response for Reviewer fF5r (3/3)

Tables R4a–R4d report the requested miss rates and false alarm ratios. At 25 mm/day, both decrease across all ten lead times for GraphCast and HRES.

Table R4a. Station Miss rate at 25 mm. Lower is better.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.69600.73700.73330.75490.77740.79060.79090.81950.86690.8737
HRES0.70830.69020.70870.74110.76600.81630.82840.83250.86210.8662
GraphCast+Ours0.60800.61410.65490.69290.72780.75800.79060.80960.85760.8751
HRES+Ours0.58220.58550.61750.66760.70270.75660.78360.80490.83840.8570
Pangu+Ours0.57200.61890.69260.76420.77490.81430.83460.84680.87050.8960

Table R4b. Station False alarm ratio at 25 mm. Lower is better.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.51510.55470.59570.64450.67140.73420.77260.79740.82880.8616
HRES0.52050.57500.61500.65740.69370.74510.77680.77270.81760.8310
GraphCast+Ours0.41260.45710.48790.53770.58300.61930.68540.69980.76630.8117
HRES+Ours0.43770.51520.56330.60630.64010.70260.72810.73760.78460.8207
Pangu+Ours0.50840.54740.56100.59190.64390.69360.72220.75460.80450.8358

Table R4c. Station Miss rate at 50 mm.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.90190.89450.90130.91980.92750.93240.94970.95820.95990.9659
HRES0.90770.91010.90480.92070.95150.95280.96790.96990.96960.9697
GraphCast+Ours0.87770.86930.87890.91380.91730.93140.92600.94230.96410.9585
HRES+Ours0.88980.88910.91070.92860.92380.94580.95500.95660.95770.9570
Pangu+Ours0.88680.87890.91700.91890.93330.94780.95210.95660.96500.9597

Table R4d. Station False alarm ratio at 50 mm.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.62420.59350.63270.67480.73200.80770.82460.84980.89030.9063
HRES0.60810.67620.75110.79290.80470.83470.86360.89010.91200.9186
GraphCast+Ours0.42250.55600.68260.76350.79070.82790.86240.86050.87930.8880
HRES+Ours0.47000.50660.57690.65610.75420.74870.82490.83720.87050.9149
Pangu+Ours0.59150.64780.74000.76820.83160.86590.85540.87570.90960.9205
Reviewer_mpSKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work trains a conditional diffusion model that maps the non-precipitation forecast fields of a frozen weather model (Pangu-Weather, GraphCast, or HRES) to daily precipitation. The generator is trained on ERA5 paired with IMERG and applied to forecast states without fine-tuning; the two states at t-12 and t are encoded by a multi-scale ConvGRU and injected through cross-attention. Experiments cover global deterministic and ensemble forecasts to day 7, regional 0.25° forecasts, a heavy-rain case study with stations in China, and ablation studies. The authors report improved daily precipitation accuracy and event detection over the backbones' own precipitation, better ensemble scores than ENS and GenCast, and closer agreement with station observations in the case study.

Critical strengths
  1. The setup is practical. One ERA5-trained generator serves three backbones without retraining, verification is against IMERG and gauges rather than reanalysis, and code is released.
  2. The evaluation is broad, and the Day-3 spatial comparison in Fig. 4 is convincing.
Critical weaknesses
  1. The baselines are weak. All competitors in Fig. 3 and Table 1 produce ERA5-like or physics-based precipitation, while only the proposed model is trained on IMERG, which is also the reference. The authors note this for ERA5+Ours, but it applies to every backbone. A simple IMERG-calibrated post-processing of the backbone precipitation (quantile mapping, or a small CNN on the same inputs) is needed to separate the effect of the diffusion model from the effect of the target. The U-Net in Table 2 should play this role, but it is far worse than raw GraphCast on RMSE, which is implausible for a model trained on the same pairs, and its setup is not explained.
  2. Heavy rain is not evaluated. CSI is only reported up to 4 mm/day, there is no FSS although the regional results are argued through sharper spatial patterns, and the station study is a single event. Deterministic scores use a single diffusion sample, so the small differences in Table 2 are hard to interpret.
  3. The recurrent encoder runs over only two states, so it is a gated fusion of two feature maps rather than a model of temporal evolution, and the ConvGRU vs ConvLSTM comparison says little. The authors' own test with four states (Appendix A.3) gives the same error. A plain convolution over the two states stacked as channels would be the fair comparison, and it is unclear whether the "-" rows in Table 2 are that.
Questions for authors
  1. How does the proposed model compare with an IMERG-calibrated post-processing of the backbone precipitation (quantile mapping or a small CNN on the same forecast fields) in the Fig. 3 / Table 1 setting?
  2. Why is the U-Net in Table 2 so much worse than raw GraphCast? What differs in its loss or setup?
  3. What replaces the ConvGRU in the "-" rows of Table 2?
  4. Is there any evaluation above 4 mm/day, or with FSS?
13 comments
AuthorsOpenReview ↗

Response for Reviewer mpSK (1/3)

We thank the reviewer for the constructive feedback and will incorporate the additional baseline comparisons, heavy-rainfall evaluations, temporal ablations, and methodological clarifications presented below into the revised manuscript.

W1 & Q2. U-Net Results and Baseline Training Setup.

Thank you for highlighting this issue. We acknowledge that our description did not clearly distinguish the following two settings:

  • Precipitation generation (Table 2). The U-Net in Table 2 receives the same non-precipitation atmospheric forecasts as our generator and predicts IMERG precipitation directly. It does not receive the backbone’s precipitation forecast, so it actually learn the atmospheric-to-precipitation mapping without using that estimate.
  • Precipitation correction: The U-Net receives both atmospheric forecasts and the backbone’s precipitation forecast, and is trained against IMERG to correct the existing precipitation estimate.

Without access to the precipitation estimate of Graphcast, precipitation generation is more challenging than precipitation correction, leading to higher RMSE. We further report the additional U-Net precipitation-correction baseline in the following:

Table R1. Precipitation forecasting performance of the original backbones, U-Net correction, and our framework.

BackboneMethodACCRMSEHSSCSI 0.1mmCSI 1mmCSI 4mm
GraphCastOriginal0.60916.12000.53470.50970.49280.4271
GraphCast+ U-Net correction0.63725.88310.64820.63290.56250.4873
GraphCast+ Ours0.70155.59000.68320.67360.60610.5340
HRESOriginal0.63946.01000.51640.51260.49130.4151
HRES+ U-Net correction0.64825.89120.64360.62970.55820.4805
HRES+ Ours0.69785.65000.68180.67420.60540.5221

As shown in Table R2, the U-Net precipitation-correction baseline achieves substantially lower RMSE than the original precipitation-generation U-Net. Nevertheless, our method outperforms the correction baseline across all reported metrics, despite not using the native precipitation forecast as an input. This comparison show that our gains extend beyond IMERG supervision alone and supports the value of the proposed precipitation generator.

We thank the reviewer for suggesting this comparison, which strengthens the empirical evaluation and clarifies our methodological contribution.

W1 & Q1. IMERG-Calibrated Baselines.

We thank the reviewer for suggesting IMERG-calibrated baselines. In addition to the U-Net correction baseline presented above, we have added quantile mapping (QM) and CNN-based precipitation correction. All three baselines receive the backbone’s native precipitation forecast as input and use IMERG observations for calibration or supervision. The results are reported as follows:

Table R2a. RMSE (mm/24 h; lower is better), Days 1–10.

BackboneMethodDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCastRaw6.12006.51206.83807.15907.48407.79508.09508.35808.55908.7580
GraphCastQM6.14986.53686.88717.23067.59037.93298.27248.57708.80759.0312
GraphCastCNN7.93208.23548.32348.62348.93429.28559.43269.72769.81589.7780
GraphCastOurs5.59005.92806.19906.50506.83207.13107.45407.88408.10208.3080
HRESRaw6.00006.56106.95207.32107.67208.02908.36908.65708.91409.1080
HRESQM6.10766.72017.17607.57447.95158.34158.71989.03539.31539.5275
HRESCNN8.00308.42608.47238.67908.98109.41759.57709.56009.91669.8004
HRESOurs5.85006.22306.55406.88007.19307.51807.84508.21308.47708.6520

Table R2b. Pointwise CSI at 1 mm/24 h (higher is better), Days 1–10.

BackboneMethodDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCastRaw0.50200.48400.46900.45200.43300.41100.38800.34130.32260.3067
GraphCastQM0.51710.50360.47210.46810.44140.42230.39260.35330.33820.3176
GraphCastCNN0.48320.46520.45370.44290.42360.40330.37590.33320.31810.2819
GraphCastOurs0.60600.58100.55700.52600.49000.45000.41000.38200.35270.3281
HRESRaw0.49100.47800.46400.44700.42600.40400.38100.34520.31810.2836
HRESQM0.50410.48990.47070.45210.43450.40560.39520.34960.32630.2948
HRESCNN0.47830.46380.45300.42690.41370.39810.37110.32980.30190.2707
HRESOurs0.60500.58400.56000.52800.49100.45100.41100.39040.36270.3405

Our method achieves lower RMSE and higher CSI at 1 mm/day than QM and CNN across Days 1–10 under both backbones, demonstrating that the gains extend beyond IMERG supervision alone. They also support the effectiveness of our precipitation-generation framework, which achieves these improvements without requiring the backbone’s native precipitation forecast as an input.

AuthorsOpenReview ↗

Response for Reviewer mpSK (2/3)

W2 & Q4. Evidence for heavy rainfall forecasting skill.

We thank the reviewer for suggesting stronger quantitative verification and an assessment of sampling variability. The following additional evaluations show that the improvements extend to higher rainfall thresholds and spatial forecast quality.

  • Heavy-rainfall detection: Against IMERG, our method achieves higher CSI at 25 and 50 mm/day than the corresponding native forecasts throughout Days 1–10 under both GraphCast and HRES conditioning (Tables R3a–R3b). We further assess these improvements against observations from the 2,288-station network (Tables R3c–R3d). At 25 mm/day, station CSI improves at all ten lead times under both backbones. At 50 mm/day, improvements occur at all ten lead times for HRES and seven for GraphCast.
  • Quantitative verification of spatial quality: We additionally evaluate FSS at both thresholds. Our method achieves higher FSS than the corresponding native forecasts throughout Days 1–10 under both backbones (Tables R3e–R3f).
  • Sensitivity to stochastic sampling: We evaluate eight stochastic realizations for each backbone and report the associated sample standard deviations alongside the scores in Tables R3a–R3d. All reported standard deviations are below 0.006. This makes the variation across realizations explicit and allows the reported improvements to be assessed alongside their sampling uncertainty.

Table R3a: Grid-based CSI at 25 mm/24 h (higher is better)

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.43580.38020.31450.28260.25450.23110.21030.18900.17190.1576
HRES0.40870.35020.33530.32990.30550.28120.25390.23250.21290.1980
GraphCast+Ours0.5000 ± 0.00450.4557 ± 0.00380.4125 ± 0.00500.3683 ± 0.00430.3275 ± 0.00560.3041 ± 0.00410.2751 ± 0.00370.2505 ± 0.00490.2285 ± 0.00330.2072 ± 0.0044
HRES+Ours0.4829 ± 0.00390.4487 ± 0.00480.4266 ± 0.00420.3987 ± 0.00550.3718 ± 0.00360.3430 ± 0.00470.3151 ± 0.00510.2902 ± 0.00380.2690 ± 0.00450.2478 ± 0.0032
Pangu+Ours0.4730 ± 0.00520.4117 ± 0.00400.3848 ± 0.00350.3586 ± 0.00480.3274 ± 0.00530.2895 ± 0.00420.2567 ± 0.00360.2265 ± 0.00460.2042 ± 0.00390.1837 ± 0.0034

Table R3b: Grid-based CSI at 50 mm/24 h (higher is better)

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.24080.17440.14330.12100.10000.08770.07550.06600.05780.0519
HRES0.22130.17620.15510.13890.12600.11280.10160.08680.07740.0692
GraphCast+Ours0.3432 ± 0.00440.2943 ± 0.00520.2480 ± 0.00400.2060 ± 0.00470.1756 ± 0.00360.1552 ± 0.00510.1332 ± 0.00430.1173 ± 0.00380.1058 ± 0.00450.0941 ± 0.0032
HRES+Ours0.3178 ± 0.00380.2783 ± 0.00490.2563 ± 0.00420.2364 ± 0.00530.2152 ± 0.00370.1948 ± 0.00450.1736 ± 0.00500.1555 ± 0.00390.1410 ± 0.00410.1291 ± 0.0034
Pangu+Ours0.3073 ± 0.00510.2442 ± 0.00430.2192 ± 0.00380.1968 ± 0.00460.1750 ± 0.00540.1457 ± 0.00400.1205 ± 0.00350.1006 ± 0.00420.0854 ± 0.00300.0746 ± 0.0037

Table R3c. Station CSI at 25 mm/day.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.22980.19810.19150.16970.15300.13260.12230.10550.08090.0707
HRES0.22150.21830.19880.17300.15290.11950.10740.10670.08520.0807
GraphCast + Ours0.3074 ± 0.00430.2912 ± 0.00510.2597 ± 0.00380.2263 ± 0.00470.1972 ± 0.00540.1736 ± 0.00410.1438 ± 0.00360.1319 ± 0.00450.0971 ± 0.00320.0812 ± 0.0039
HRES + Ours0.3153 ± 0.00500.2877 ± 0.00420.2561 ± 0.00480.2199 ± 0.00370.1945 ± 0.00460.1545 ± 0.00520.1370 ± 0.00390.1260 ± 0.00430.1017 ± 0.00350.0864 ± 0.0031
Pangu + Ours0.2967 ± 0.00460.2608 ± 0.00390.2207 ± 0.00530.1757 ± 0.00420.1600 ± 0.00360.1308 ± 0.00480.1157 ± 0.00340.1041 ± 0.00400.0845 ± 0.00290.0680 ± 0.0033

Table R3d. Station CSI at 50 mm/day.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.08440.09140.08440.06880.06050.05270.04070.03380.03020.0257
HRES0.08070.07570.07400.06080.04040.03810.02670.02420.02310.0226
GraphCast+Ours0.1123 ± 0.00370.1124 ± 0.00450.0961 ± 0.00320.0674 ± 0.00410.0630 ± 0.00360.0516 ± 0.00280.0506 ± 0.00390.0425 ± 0.00310.0284 ± 0.00240.0312 ± 0.0027
HRES+Ours0.1004 ± 0.00420.0996 ± 0.00350.0796 ± 0.00400.0629 ± 0.00330.0618 ± 0.00440.0467 ± 0.00290.0372 ± 0.00360.0355 ± 0.00250.0329 ± 0.00300.0294 ± 0.0022
Pangu+Ours0.0973 ± 0.00380.0990 ± 0.00430.0671 ± 0.00310.0639 ± 0.00370.0502 ± 0.00260.0390 ± 0.00340.0373 ± 0.00280.0332 ± 0.00320.0259 ± 0.00210.0275 ± 0.0025
AuthorsOpenReview ↗

Response for Reviewer mpSK (3/3)

Table R3e. Grid-based FSS 4x4 at 25 mm. Higher is better.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.73030.65860.60660.55180.49580.44150.38850.34350.30420.2695
HRES0.73870.67150.62790.58060.51740.44370.37230.30680.26350.2239
ENS Mean0.73820.64360.57640.50020.41910.33100.24350.17340.12260.0820
GraphCast+Ours0.79750.73800.66640.61110.56490.51080.45710.40430.36230.3270
HRES+Ours0.76560.71550.67560.63320.58770.53550.47940.42500.37770.3397
Pangu+Ours0.75680.70070.65620.59010.52600.47550.41400.36480.31720.2776

Table R3f. Grid-based FSS 4x4 at 50 mm.

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.53890.42840.36160.30850.25080.21560.17300.15180.12750.1065
HRES0.53420.45610.40420.36230.32570.28270.23720.20070.16770.1418
ENS Mean0.46640.33640.24860.16980.09460.04170.01490.00530.00270.0017
GraphCast+Ours0.59750.49690.44070.39700.35020.29660.24700.20790.17800.1492
HRES+Ours0.60500.54300.49940.45820.41450.36780.31590.27140.23280.2058
Pangu+Ours0.64830.57350.48560.40360.32970.26330.21420.16410.13240.1059

W3. Temporal aggregation.

Thank you for raising these questions. We would like to clarify that a two-state ConvGRU can encode temporal relationships through sequential hidden-state updates. To directly test whether this aggregation provides a practical benefit, we compare it with a channel-stacked CNN. Both configurations receive the same two atmospheric states and retain the same attention module and training protocol; the temporal aggregation module is replaced.

The ConvGRU configuration achieves higher CSI and FSS at 25 mm/day at every lead time from Day 1 to Day 10 under both GraphCast and HRES conditioning (Table R4). This comparison directly supports the benefit of recurrent aggregation over the tested concatenation-based alternative. The four-state experiment in Appendix A.3 addresses a different question, whether denser temporal inputs improve prediction, and does not replace this aggregation comparison.

Table R4. Supplementary temporal-encoder and fusion ablations, Days 1–10.

(a) CSI 4x4 at 25 mm/24 h. Higher is better.

BackboneVariantDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCastConvGRU + attention0.32230.26460.24160.22480.20700.18920.17390.15880.14520.1336
GraphCastConvGRU + Conv/Concat0.28800.23810.21800.20630.19240.17670.16310.15330.14110.1312
GraphCastchannel-stacked CNN + attention0.28270.23370.21130.20090.18680.17240.15730.14730.13750.1275
HRESConvGRU + attention0.34210.32000.30750.29280.27440.25330.23460.21670.20290.1881
HRESConvGRU + Conv/Concat0.30610.28610.27480.26390.25030.23050.21330.19860.18560.1722
HREShannel-stacked CNN + attention0.29930.27950.26870.25950.24840.22770.21090.19730.18770.1742

(b) FSS 4x4 at 25 mm/24 h. Higher is better.

BackboneVariantDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCastConvGRU + attention0.57440.47170.42390.38270.34260.30230.26250.23470.19980.1763
GraphCastConvGRU + Conv/Concat0.50170.40960.36400.33420.30090.26490.23040.21060.18070.1617
GraphCasthannel-stacked CNN + attention0.48340.39750.35050.32680.29540.26120.22510.20800.18200.1625
HRESConvGRU + attention0.59770.55400.52670.49320.45980.41450.37330.32900.29350.2598
HRESConvGRU + Conv/Concat0.51470.47560.44820.42230.39590.35540.31830.28340.25190.2228
HREShannel-stacked CNN + attention0.49750.45790.43130.40830.38690.34750.30940.27860.25170.2239

Q3. Clarification of Table 2.

In the module ablations of Table 2, the “–” entries denote basic convolution and concatenation operations for encoding and feature alignment.

Reviewer_Xa9uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The author proposed Dynamics–Precipitation Decoupling method which separating atmospheric evolution from observation-supervised precipitation. To capture extensive precipitation and variational atmospheric conditioning, the author thought generative models such as diffusion have shown potential to solve these challenges. The diffusion generator takes non-precipitation states as inputs and predict daily precipitation. Extensive experiments show improved daily precipitation accuracy.

Critical strengths

The Dynamics–Precipitation Decoupling method is a plug-and-play framework which effectively improved the precipitation accuracy compared with the corresponding WFM.

Critical weaknesses

The article title includes a medium-range weather forecast, but the actual forecast is only for 7 days, whereas a medium-range forecast should cover at least 15 days.

Questions for authors

1.The authors make forecasts up to 7 days. Could you make a longer prediction period such as 15 days? 2.The author chose to compare the model and the WFM method on daily precipitation forecasting tasks. Why not predict precipitation for each 6-hour interval? 3.It is hard to figure out how the proposed method solve the heterogeneous challenges in the introduction. 4.The author thought atmospheric conditions should be dynamic and employ convgru to capture temporal evolution. why can't we use atmospheric conditions from a longer time period, but only take xt and xt-12 as conditions ? 5.The author evaluate the daily precipitation on the {0.1,1,4}mm thresholds. These thresholds cannot reflect the model's performance in predicting daily extreme precipitation. It is recommended that the author examine the thresholds of 25mm and 50mm.

7 comments
AuthorsOpenReview ↗

Response for Reviewer Xa9u (1/2)

We thank the reviewer for the constructive comments and suggestions. We will incorporate the additional experimental results and clarifications below into the revised manuscript.

W1 & Q1. Could you make a longer prediction period such as 15 days?

Yes. Following your suggestion, We have extended the evaluation to 15 days, with RMSE, HSS, and CSI reported in Tables R1a–R1c. GraphCast+Ours improves on native GraphCast across all three metrics throughout Days 8–15, while HRES+Ours improves on native HRES throughout the available Days 8–10. These results demonstrate that the gains over the corresponding native precipitation forecasts persist beyond seven days.

Table R1a. RMSE (mm/24 h), Days 8–15.

ModelDay 8Day 9Day 10Day 11Day 12Day 13Day 14Day 15
HRES8.6578.9149.108N/AN/AN/AN/AN/A
GraphCast8.3588.5598.7588.9049.0279.1289.2059.268
GenCast7.9358.3758.4618.6928.8319.0569.1489.185
ENS Mean8.0248.4688.6898.7428.8599.1429.2059.235
Pangu-Weather+Ours8.2248.3878.5378.5908.6958.7238.7398.761
GraphCast+Ours7.8848.1028.3088.4298.5738.6898.7238.793
HRES+Ours8.2138.4778.652N/AN/AN/AN/AN/A

Table R1b. HSS, Days 8–15.

ModelDay 8Day 9Day 10Day 11Day 12Day 13Day 14Day 15
HRES0.34810.32150.2986N/AN/AN/AN/AN/A
GraphCast0.35450.32520.29960.27890.26340.24900.23820.2298
GenCast0.31830.30980.28610.26130.24710.22140.21390.2031
ENS Mean0.30780.28190.26160.24540.23290.22440.21830.2134
Pangu-Weather+Ours0.34980.30860.27440.24800.22630.20910.19720.1874
GraphCast+Ours0.40820.36900.33630.29900.27740.25780.24370.2346
HRES+Ours0.40550.36380.3272N/AN/AN/AN/AN/A

Table R1c. CSI at 4 mm/24 h, Days 8–15.

ModelDay 8Day 9Day 10Day 11Day 12Day 13Day 14Day 15
HRES0.21500.20100.1825N/AN/AN/AN/AN/A
GraphCast0.22970.21350.19170.17520.15310.14320.13530.1297
GenCast0.21480.20450.19840.18030.16160.15320.14510.1395
ENS Mean0.23710.21750.20090.18850.17820.17130.16510.1606
Pangu-Weather+Ours0.23300.20720.18640.17120.15890.15010.14320.1386
GraphCast+Ours0.26720.23870.21420.19670.18410.17290.16480.1602
HRES+Ours0.27620.24850.2267N/AN/AN/AN/AN/A

Q2. Why not predict precipitation for each 6-hour interval?

Thank you for the advice. We initially focused on daily precipitation to evaluate accumulated rainfall relevant to agricultural and water-resource planning. We agree that six-hourly prediction provides a complementary evaluation at a finer temporal scale in Table R2a-R2c.

Table R2a. RMSE Comparison mm/6 h

Model6 h12 h18 h24 h48 h72 h96 h120 h168 h240 h
GraphCast1.9391.9382.0412.1332.2492.3562.4832.5992.7922.962
GraphCast+Ours1.9071.7961.9291.9572.0712.1842.2992.4112.5972.769
HRES1.9302.0392.1632.2472.3932.5222.6522.7702.9683.150
HRES+Ours1.8911.8941.9432.0092.1732.3112.4362.5522.7472.918
Pangu+Ours1.9582.0082.1522.2592.4392.5882.7302.8463.0313.151

Table R2b. ACC Comparison

Model6 h12 h18 h24 h48 h72 h96 h120 h168 h240 h
GraphCast0.6450.6310.6120.5880.5260.4680.3970.3280.2100.103
GraphCast+Ours0.6920.6780.6390.6300.5780.5190.4500.3780.2530.129
HRES0.6610.6030.5730.5530.4890.4300.3680.3100.2050.105
HRES+Ours0.7300.6590.6300.6080.5320.4640.3990.3360.2240.117
Pangu+Ours0.6760.6390.6000.5750.4990.4230.3430.2740.1580.070

Table R2a. CSI Comparison at 25mm mm/6 h

Model6 h12 h18 h24 h48 h72 h96 h120 h168 h240 h
GraphCast0.1440.2170.1510.1410.0950.0700.0520.0400.0250.014
GraphCast+Ours0.2880.2600.2260.2040.1720.1490.1190.0950.0650.038
HRES0.1490.1080.0910.0850.0610.0460.0350.0250.0140.005
HRES+Ours0.2580.2190.2120.1990.1710.1520.1340.1170.0850.054
Pangu+Ours0.2900.2730.2460.2310.1980.1630.1300.1060.0670.035
AuthorsOpenReview ↗

Response for Reviewer Xa9u (2/2)

Q3. How the proposed method addresses precipitation heterogeneity?

Thanks for the question. We would like to clarify that precipitation heterogeneity refers to variations in both spatial structure and rainfall intensity: a broad rainband may contain a few intense local peaks, while surrounding areas receive little or no rain. Models trained with pointwise MSE loss (e.g., CNN or U-Net regressors) can average these possibilities, producing weaker peaks and more diffuse rainbands. In contrast, our diffusion generator learns a conditional distribution of possible rainfall fields and generates individual realizations from it. Each realization can therefore represent a distinct rainfall pattern, including localized peaks, instead of being restricted to an average across different possibilities.

Our results provide complementary evidence for spatial structure and heavy-rainfall event detection. Figures 4–5 show qualitative improvements in rainband structures and localized accumulation patterns, while the additional CSI evaluations at 25 and 50 mm/day in Q5 demonstrate improved detection at higher rainfall thresholds.

Q4. Why can't we use atmospheric conditions from a longer time period?

Yes. Our choice of two states is based on an empirical accuracy–efficiency trade-off. In Appendix A.3, we already examined denser sampling within the daily accumulation window: replacing the default two-state input (Xt−12,Xt)(X_{t-12}, X_t) with four states at six-hour intervals changes mean RMSE only from 6.520 to 6.518, while increasing inference overhead from 0.33 to 0.84 s.

To address your question about longer histories, we further compared three-, four-, and five-state sequences by successively adding Xt−24X_{t-24}, Xt−36X_{t-36}, and Xt−48X_{t-48}. Tables R3a–R3b show that longer histories improve some lead-time and threshold combinations but do not provide consistent CSI gains. They also increase inference time by 0.23, 0.47, and 0.65 s, respectively (Table R3c).

Table R3a. CSI at 25 mm/24 h under GraphCast conditioning.

FramesDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
20.33000.26570.25420.22830.22750.19410.17510.15050.14850.1405
30.33800.26920.24900.23160.21160.19380.17920.16590.15150.1398
40.32480.26480.24390.22640.20750.18790.17280.15970.14560.1334
50.34260.26880.25140.22760.21030.18940.18050.16500.14730.1344

Table R3b. CSI at 50 mm/24 h under GraphCast conditioning.

FramesDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
20.20320.13430.12800.10600.10560.08520.06320.06730.05580.0541
30.19240.13200.11400.10010.09030.08020.06920.06560.05720.0538
40.19880.13140.11210.09850.08930.07620.06860.06240.05360.0506
50.20080.13560.12580.10130.09670.08170.06710.06510.05220.0507

Table R3c. Additional inference time relative to two states.

StatesAdditional stateExtra time per inference
2—Baseline
3Xt−24X_{t-24}+0.23s
4Xt−36X_{t-36}+0.47s
5Xt−48X_{t-48}+0.65s

Q5. Evaluating Heavy Rainfall at 25 mm and 50 mm Thresholds.

Thank you for this suggestion. We extend the gridded evaluation to 25 and 50 mm/day. The following tables report CSI for Days 1–10. Our framework improves all backbones across all ten lead times, indicating improved heavy-rainfall event.

Table R4a. Grid-based CSI at 25 mm (higher is better)

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.43580.38020.31450.28260.25450.23110.21030.18900.17190.1576
HRES0.40870.35020.33530.32990.30550.28120.25390.23250.21290.1980
GenCast0.40120.32520.32410.29120.24410.23620.22510.19070.18740.1744
ENS Mean0.39450.30150.25150.20600.16200.12470.09210.06900.05110.0388
GraphCast+Ours0.50000.45570.41250.36830.32750.30410.27510.25050.22850.2072
HRES+Ours0.48290.44870.42660.39870.37180.34300.31510.29020.26900.2478
Pangu+Ours0.47300.41170.38480.35860.32740.28950.25670.22650.20420.1837

Table R4b. Grid-based CSI at 50 mm (higher is better)

ModelDay 1Day 2Day 3Day 4Day 5Day 6Day 7Day 8Day 9Day 10
GraphCast0.24080.17440.14330.12100.10000.08770.07550.06600.05780.0519
HRES0.22130.17620.15510.13890.12600.11280.10160.08680.07740.0692
GenCast0.23510.16420.15890.14640.13850.12740.11430.09320.08690.0542
ENS Mean0.17480.10210.06950.04520.02510.01210.00590.00290.00170.0010
GraphCast+Ours0.34320.29430.24800.20600.17560.15520.13320.11730.10580.0941
HRES+Ours0.31780.27830.25630.23640.21520.19480.17360.15550.14100.1291
Pangu+Ours0.30730.24420.21920.19680.17500.14570.12050.10060.08540.0746

The ERA5-based evaluation also improves:

ModelCSI at 25 mmCSI at 50 mm
ERA50.4520.299
ERA5+Ours0.5430.389
AuthorsOpenReview ↗

A summary of our revision

Dear Reviewer Xa9u,

Thank you again for your valuable comments. It has been three days since we posted our responses and revised manuscript, and would appreciate it if you could take some time to review the updates.

We understand your busy schedule. To facilitate your review, we summarize the main additions addressing your concerns:

  • Longer horizons and finer prediction intervals: We extended evaluation to 15 days and added six-hour precipitation forecasts through 240 h (Table 1 and Figure 13 in the revision; Tables R1a–R1c in our response).
  • Precipitation heterogeneity and temporal conditioning: We clarified how conditional generation represents heterogeneous rainfall patterns and compared two to five atmospheric states, examining the accuracy–efficiency trade-off of longer histories (response to Q3 and Figure 9 in the revision).
  • Heavy-rainfall evaluation: We added CSI at 25 and 50 mm/day, showing improved detection under GraphCast and HRES conditioning across Days 1–10 (Figure 3 in the revision; Tables R4a–R4b in our response).

The manuscript changes are highlighted in blue. Please let us know if any concerns remain. We would be grateful for your feedback on whether these additions and explanations address your questions.

Thank you for your time and consideration!

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of paper revision

Dear Reviewers,

We sincerely thank all reviewers (Xa9u, mpSK, and fF5r) for their valuable feedback. Following your suggestions, we have made the following additions to the revised manuscript. These additions do not change the paper’s contributions.

  • Extended forecast horizons and six-hour prediction. Added 15-day ensemble evaluation in Table 1, Section 5.2, and six-hour forecasts through 240 h in Figure 13. (Xa9u)

  • Additional ablations. Added longer conditioning histories in Figure 9, Section 5.4, and temporal-encoder and global-mixing fusion comparisons in Table 4. (Xa9u, mpSK, fF5r)

  • Spatial attention analysis. Added spatial and layer-wise cross-attention visualizations for South China and Pakistan in Figures 16 and 17. (fF5r)

  • IMERG-calibrated baselines and sampling variability. Added QM and clarified the original U-Net setup in Table 2, Section 5.4. Added U-Net correction comparisons and CSI sampling standard deviations in Figure 15. (mpSK)

  • Expanded heavy-rainfall verification. Added gridded and station-based verification at 25 and 50 mm/day, including CSI, FSS, miss rates, and false alarm ratios, in Figures 3, 7, and 14, with metric definitions in Appendix A.2. (Xa9u, mpSK, fF5r)

We have incorporated the suggested modifications in the revised version, which are highlighted in blue. If you are satisfied with the revisions, we kindly request your approval to consider an improved score.

Thanks for all the reviewers' time again.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.