Pith. sign in

REVIEW 4 major objections 7 minor 42 references

Skillful High-Resolution Ensemble Precipitation Forecasting with an Integrated Deep Learning Framework

T0 review · 4 major / 7 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a hybrid 3D SwinTransformer plus conditional latent diffusion framework produces reliable 0.05° precipitation ensembles over China that match observed precipitation better than ERA5.

desk verdict Credible hybrid deterministic+diffusion precipitation ensemble with a real operational pipeline, but the reliability claim outruns the evidence. read the letter →

arxiv 2501.02905 v1 pith:LVSNSW2B submitted 2025-01-06 cs.LG cs.AI

Shuangshuang He , Hongli Liang , Yuanting Zhang , Xingyuan Yuan This is my paper · ORCID
classification cs.LGcs.AI
keywords precipitationforecastingensemblelatentdiffusionconditionalSwinTransformerERA5CMPAextreme
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper aims to establish that high-resolution ensemble precipitation forecasting can be done reliably and cheaply by a hybrid deep-learning model that never sees the target resolution during prediction. The framework takes coarse (0.25°) atmospheric variables and returns 0.05° hourly precipitation: a 3D SwinTransformer predicts the mesoscale mean, and a conditional latent diffusion model generates a stochastic convective-scale residual; sampling the latent space eleven times yields an ensemble. The authors evaluate the ensemble against the CMPA high-resolution merged analysis over August 2021 and report a nearly uniform rank histogram and cumulative distributions closer to CMPA than ERA5, particularly for heavy precipitation. They also build a five-day real-time forecasting system using open operational forecast data and report usable Critical Success Index scores. If the claims hold, operational high-resolution probabilistic precipitation forecasting becomes feasible at a tiny fraction of the cost of running a high-resolution numerical weather model.

What carries the argument

The load-bearing idea is the additive decomposition $X_t = \bar{X}_t + X'_t$, where $\bar{X}_t$ is the mesoscale mean precipitation produced by the deterministic 3D SwinTransformer — a vision transformer that processes spatial and temporal patches — and $X'_t$ is the residual precipitation generated by a conditional latent diffusion model. The residual is computed as the difference between scaled CMPA and ERA5 precipitation, so the diffusion model effectively learns the conditional distribution of the discrepancy between a coarse reanalysis and a high-resolution merged analysis. A VAE compresses the residual to a $16\times 90\times 140$ latent space; the DiT-based denoiser — a transformer-based diffusion backbone — is conditioned on the atmospheric state and mean precipitation; DDIM sampling with 300 steps generates 11 members; probability matching combines them into an ensemble mean.

What would settle it

Extend the evaluation from the single month of August 2021 to a full twelve months spanning all seasons and compare the rank histogram and cumulative distribution functions against CMPA over that entire period. If the near-uniform rank histogram and the closer-to-CMPA-than-ERA5 distribution only appear in the summer month, the claim of an unbiased, well-calibrated ensemble system fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that the physically motivated decomposition of precipitation into a grid-resolved mean plus a sub-grid stochastic perturbation can be learned from reanalysis data, and that the learned stochastic part carries genuine forecast skill. Concretely, the deterministic model predicts mesoscale mean precipitation from consecutive atmospheric states, and a conditional diffusion model in latent space generates residual precipitation, conditioned on the same atmospheric state plus the mean prediction. At inference, repeatedly sampling Gaussian noise in the latent space yields an 11-member ensemble. Evaluation on August 2021 over China shows a nearly uniform rank histogram — with only slightly elevated tails — and cumulative distribution functions of the 11 members that track the CMPA observations more closely than ERA5 does, especially for heavy precipitation. The authors also report that a real-time system forced by open operational forecast data maintains CSI scores of roughly 0.15–0.3 at the 0.1 mm/h threshold out to five days.

Load-bearing premise

The load-bearing premise is that the real-time operational forecast data used at inference resemble the ERA5 reanalysis used in training closely enough that the model transfers without fine-tuning; the paper only rescales the precipitation variable, and the authors themselves flag fine-tuning as needed as more data accumulate.

Editorial extensions

If this is right

  • A model trained once on reanalysis can produce 120-hour, 0.05-degree ensemble forecasts from 0.25-degree inputs, which removes the usual computational barrier to high-resolution ensemble precipitation forecasting.
  • The two-component design means the deterministic backbone can be swapped (for example, for a different NWP or data-driven model) without retraining the diffusion part, since the diffusion only models the residual between the mean prediction and the high-resolution truth.
  • Repeated latent-space sampling yields 11 ensemble members whose spatial patterns agree at synoptic scales but differ in convective details, giving operational users a probabilistic rather than a single deterministic output.
  • Probability-matching over the members produces an ensemble mean that represents heavy precipitation better than the raw ERA5 field, which matters for flood and disaster response.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural next test is to fine-tune the diffusion model on residual statistics computed from real-time forecast inputs rather than ERA5; closing that gap is the paper's own stated future work and would likely raise the real-time CSI scores.
  • The decomposition into mean-plus-residual is a general post-processing template: any coarse deterministic precipitation forecast could be 'sharpened' by adding a calibrated stochastic residual learned against a high-resolution merged analysis, provided such an analysis exists for the region.
  • Because the residual is defined as the difference between CMPA and ERA5, part of what the diffusion learns may be systematic bias correction rather than pure physical stochasticity; splitting the residual into a deterministic correction term and a truly random component would sharpen the physical interpretation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. This paper proposes a deep learning framework for hourly 0.05-degree ensemble precipitation forecasting over China. It combines a deterministic 3D SwinTransformer that predicts mesoscale precipitation from ERA5 atmospheric fields with a latent conditional diffusion model that generates residual precipitation (defined as CMPA minus ERA5) to represent sub-grid variability; 11 ensemble members are produced by repeated DDIM sampling. The authors report an ablation study of deterministic design choices, a case study, a rank histogram claimed to show unbiasedness, CDF comparisons against CMPA and ERA5 for a heavy-rain period, and a real-time 5-day forecasting experiment using ECMWF open data. The central claims are that the ensemble is nearly unbiased, that its precipitation distribution is closer to CMPA than ERA5, and that it captures extreme precipitation events better than ERA5.

Significance. The practical motivation is strong: high-resolution precipitation ensembles are important for flood and disaster response, and the hybrid deterministic-plus-diffusion architecture is a plausible way to address the common blurring problem in deep-learning precipitation forecasts. The paper includes an ablation study and a real-time implementation, which are concrete strengths. However, the headline claims of unbiasedness and distributional fidelity currently rest on a single-month, single-region evaluation, a rank histogram whose construction is under-specified, and a verification target that is also used to define the training target. If these issues are addressed with additional analysis and broader evaluation, the framework could be a useful applied contribution; at present the evidence is not yet at the level claimed in the abstract and key points.

major comments (4)
  1. [§4.2.2 / Figure 7] The rank histogram is the only quantitative support for the 'nearly unbiased' claim, but the manuscript does not state how zero precipitation values were handled, whether ties were randomized, over which grid cells and time steps the histogram was computed, or what the sampling uncertainty is. Precipitation is a mixed discrete-continuous variable with a large point mass at exactly 0 mm/h, especially at 0.05-degree resolution; without explicit randomized tie-breaking or separate treatment of dry and wet cases, the rank histogram can appear artificially flat. Please report the exact construction, including any randomized rank procedure, the spatiotemporal sample (e.g., all of August 2021 over the full domain), and confidence bands or a statistical test for uniformity.
  2. [§3.1, §4.2.2] Because the residual precipitation TP' is defined as scaled CMPA minus scaled ERA5, and the ensemble is generated as TP + TP', verifying the ensemble against CMPA is partly a test of reconstruction of the training target rather than an independent forecast evaluation. This does not invalidate the architecture, but it weakens the claim that the output distribution is inherently 'closer to CMPA than ERA5.' Please add a comparison against a baseline that does not use CMPA in its construction, such as deterministic-only forecasts, a simple bias-corrected ERA5 ensemble, or a diffusion model trained on a different target, or at least explicitly quantify how much of the CDF improvement follows from the additive construction.
  3. [§4.2–4.3] The quantitative evaluation is limited to August 2021 over a single region: the rank histogram and CDF use one month, the case study is one event, and the real-time CSI evaluation covers August 2021 only. There is no multi-season or multi-year verification, no confidence intervals on the scores, and no comparison against an operational NWP ensemble or a probabilistic baseline. Given the title and key points claim 'skillful' forecasting and 'unbiased' ensembles, this single-month evaluation is not sufficient support; please broaden the evaluation or temper the claims.
  4. [§4.3 / Conclusion] The real-time forecasting experiment uses ECMWF open data after only 'adjusting the precipitation data scale' (Conclusion), without quantifying the distribution shift between ERA5 and real-time IFS forecasts. Moreover, the 3-hour ECMWF fields are converted to hourly forecasts using Pangu-Weather extrapolations, and the error introduced by this intermediate step is not evaluated separately. Without a comparison of ERA5-driven and ECMWF-driven ensemble performance on the same dates, the reported CSI results cannot be attributed to the framework rather than to the specific input stream or the extrapolation procedure.
minor comments (7)
  1. [§3.2.3 / Eq. (5)] The text says the loss 'balances L1 and SSIM,' but Eq. (5) uses a squared L2 term; please correct the inconsistency.
  2. [§3.2.3] The acronym 'SIMM' appears in the text; it should be 'SSIM.'
  3. [Figure 7 caption] The caption contains a duplicated phrase: 'The black dashed line The black dashed line represents...' Please fix.
  4. [Data Availability] The text has a typo, 'ECWWF open real-time forecast'; it should be 'ECMWF.' Also, the Code Availability section is empty; please state whether code will be released.
  5. [Table 1] The heading 'Determinstic model' is misspelled; it should be 'Deterministic.'
  6. [§2.4] Reducing CMPA from 0.01 degrees to 0.05 degrees by average pooling may smooth extreme precipitation values; please justify this choice or discuss its effect on the reported extremes.
  7. [§3.4 / §4.2.1] The manuscript does not explain why 11 ensemble members and 300 DDIM sampling steps were chosen, nor how probability matching is applied to combine the members; please provide the rationale and a precise description.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the model is trained on CMPA-derived residuals but evaluated on a held-out 2021 test period, so the CMPA comparison is a legitimate out-of-sample forecast assessment.

full rationale

The paper defines the residual precipitation as the difference between scaled CMPA and scaled ERA5 precipitation (Section 3.1) and trains the probabilistic model to generate this residual conditioned on atmospheric state. It then evaluates the resulting ensemble against CMPA on August 2021 data that was not used in training (training spans 2018-2019, validation 2020, testing 2021). Therefore, the claim that the ensemble distribution is closer to CMPA than ERA5 is not an algebraic identity or a fitted-input prediction: the model must generalize to a held-out period, and the stochastic diffusion sampler does not guarantee exact reproduction of CMPA. The rank histogram and CDF evaluations use CMPA as the reference, which is standard practice even when CMPA is also the training target, because the test period is independent. Potential concerns about rank-histogram handling of zero-precipitation ties are evaluation-methodology issues rather than circularity. No load-bearing self-citation or imported uniqueness argument appears. The framework's decomposition into a deterministic mean plus residual is a modeling choice, not a circular derivation of the paper's central claims.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The framework rests on domain assumptions about decomposition, the accuracy of CMPA as ground truth, the sufficiency of coarse inputs for conditioning fine-scale generation, and transfer from ERA5 to ECMWF. Free parameters are limited to loss weights, normalization scale, and a few architectural or inference choices. No new physical entities are introduced; the residual precipitation field is a derived training target, not a new object.

free parameters (4)
  • loss weights lambda1, lambda2 = 0.5, 1.5
    Chosen by hand to balance pixel-wise MSE and SSIM (Eq. 5); no validation sweep is reported, so optimality is untested.
  • dBZ normalization scale = temporal mean of spatial maxima over training years
    Equation (2) defines the normalization constant from the training data; it affects the target distribution and the residual definition.
  • VAE latent dimension = 16 channels (16 x 90 x 140)
    Architecture choice in Section 3.3.1; no ablation on latent size is reported.
  • ensemble size and DDIM sampling steps = 11 members, 300 steps
    Inference hyperparameters in Section 3.4; the paper does not show sensitivity to these choices.
assumptions (4)
  • domain assumption High-resolution precipitation can be decomposed into a mesoscale mean plus a subgrid residual (Eq. 3).
    Borrowed from NWP subgrid parametrization thinking; the residual is trained as CMPA minus interpolated ERA5 in Section 3.1.
  • domain assumption CMPA is an accurate ground truth for precipitation over China.
    CMPA is used both as the training target for the residual model and as the verification reference throughout Sections 2.2 and 4.2.
  • domain assumption Coarse 0.25 degree ERA5 atmospheric fields carry enough information to condition stochastic generation of 0.05 degree precipitation.
    The entire method maps coarse inputs to fine precipitation; if small-scale precipitation is not determined by these inputs, the diffusion model only learns climatology.
  • domain assumption ECMWF IFS forecasts and ERA5 reanalysis are sufficiently homogeneous for direct transfer without fine-tuning.
    Stated in the Conclusion as a current limitation: "we didn't apply fine-tuning beyond adjusting the precipitation data scale."

how reviews work

0 comments
Cite this review

Pith. "Pith review of Skillful High-Resolution Ensemble Precipitation Forecasting with an Integrated Deep Learning Framework." pith.science (2026). https://pith.science/paper/LVSNSW2B

@misc{pith2026250102905,
  author       = {Pith},
  title        = {Pith review of: Skillful High-Resolution Ensemble Precipitation Forecasting with an Integrated Deep Learning Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LVSNSW2B}},
  note         = {Machine review of arXiv:2501.02905}
}
abstract

High-resolution precipitation forecasts are crucial for providing accurate weather prediction and supporting effective responses to extreme weather events. Traditional numerical models struggle with stochastic subgrid-scale processes, while recent deep learning models often produce blurry results. To address these challenges, we propose a physics-inspired deep learning framework for high-resolution (0.05\textdegree{} $\times$ 0.05\textdegree{}) ensemble precipitation forecasting. Trained on ERA5 and CMPA high-resolution precipitation datasets, the framework integrates deterministic and probabilistic components. The deterministic model, based on a 3D SwinTransformer, captures average precipitation at mesoscale resolution and incorporates strategies to enhance performance, particularly for moderate to heavy rainfall. The probabilistic model employs conditional diffusion in latent space to account for uncertainties in residual precipitation at convective scales. During inference, ensemble members are generated by repeatedly sampling latent variables, enabling the model to represent precipitation uncertainty. Our model significantly enhances spatial resolution and forecast accuracy. Rank histogram shows that the ensemble system is reliable and unbiased. In a case study of heavy precipitation in southern China, the model outputs align more closely with observed precipitation distributions than ERA5, demonstrating superior capability in capturing extreme precipitation events. Additionally, 5-day real-time forecasts show good performance in terms of CSI scores.

Figures

Figures reproduced from arXiv: 2501.02905 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Overall model architeture. Xt−1/Xt includes surface variables (T2m, U10m, V10m, MSLP) and upper-air variables at 13 pressure levels (T, U, V, SP, Z) at time t − 1/t. For the de￾terministic model, the inputs are Xt−1 and Xt, while for the probabilistic model, the input is Xt. The model is trained on ERA5 data but can take forecasts from other models during inference. T P is the mean precipitation at time t, T P′ is t… view at source ↗
Figure 3
Figure 3. Conditional diffusion model architecture (Peebles & Xie, 2023). We introduce con￾ditioning inputs by using the atmospheric state and mean precipitation. ing objective can be formulated as: L = Eembedding(x),y,ϵ∼N(0,1),t ∥ϵ − ϵθ(Zt, t, cond)∥ 2 (6) where x is the condition, which is the atmospheric state, y is the residual precip￾itation, ϵ is random noise, t ∈ [1, 1000] is the time step of the denoising process, zt … view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: CSI of ablation studies against multiple precipitation thresholds, the experiment configuration referred [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: A case on 2021-08-01 01:00. The first row shows the ground truth, while the second row shows the predictions. The leftmost image in the second row represents the mean precip￾itation generated using ERA5 as input, the middle image (Residual Member1) represents one membe…
Figure 6
Figure 6. Figure 6: Precipitation of CMPA and all members generated by our ensemble system on 2021-08-01 01:00. ure 7, the rank histogram exhibits a nearly uniform distribution, with only slightly el￾evated tails. This indicates that in some cases, the system may slightly overestimate pre…
Figure 7
Figure 7. Figure 7: Rank Histogram. Blue bars showing the probability of the observed precipitation’s rank among the ensemble members. The black dashed line The black dashed line represents the case where the true value is equally likely to fall between any two ensemble values [PITH_FULL…
Figure 8
Figure 8. Figure 8: Cumulative PDF of precipitation. The yellow line represents ERA5 precipitation, the black dashed line represents CMPA precipitation, and the gray curves represent the 11 en￾semble members’ forecasts. 4.3 Real-time ensemble forecasting We also develop a real-time high-r…
Figure 9
Figure 9. Figure 9: The average CSI, POD, and FAR from the real-time forecast system for August 2021, over 120-hour forecasts issued at 12 UTC each day. Shaded areas show the range of values across individual ensemble members; the solid line shows average value across all members; the das…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 28 canonical work pages

  1. [1]

    , Kettleborough, J A

    allen2022model APACrefauthors Allen, M R. , Kettleborough, J A. \ Stainforth, D A. APACrefauthors \ 2022 . Model error in weather and climate forecasting Model error in weather and climate forecasting . ECMWF Predictability of Weather and Climate Seminar ECMWF predictability of weather and climate seminar \ ( \ 279--304). APACrefURL http://www.ecmwf.int/p...

    work page 2022

  2. [2]

    , Merizzi, F

    aspertiPrecipitationNowcastingGenerative2023 APACrefauthors Asperti, A. , Merizzi, F. , Paparella, A. , Pedrazzi, G. , Angelinelli, M. \ Colamonaco, S. APACrefauthors \ 2023 . Precipitation Nowcasting with Generative Diffusion Models Precipitation nowcasting with generative diffusion models \ ( \ arXiv:2308.06733). arXiv

    arXiv 2023

  3. [3]

    , Thorpe, A

    bauerQuietRevolutionNumerical2015 APACrefauthors Bauer, P. , Thorpe, A. \ Brunet, G. APACrefauthors \ 2015 . The Quiet Revolution of Numerical Weather Prediction The quiet revolution of numerical weather prediction . Nature 525 7567 47--55

    work page 2015

  4. [4]

    , Xie, L

    biAccurateMediumrangeGlobal2023 APACrefauthors Bi, K. , Xie, L. , Zhang, H. , Chen, X. , Gu, X. \ Tian, Q. APACrefauthors \ 2023 . Accurate Medium-Range Global Weather Forecasting with 3D Neural Networks Accurate medium-range global weather forecasting with 3D neural networks . Nature 619 7970 533--538

    work page 2023

  5. [5]

    , Bruinsma, W P

    bodnarFoundationModelEarth2024 APACrefauthors Bodnar, C. , Bruinsma, W P. , Lucic, A. , Stanley, M. , Vaughan, A. , Brandstetter, J. Perdikaris, P. APACrefauthors \ 2024 . A Foundation Model for the Earth System A Foundation Model for the Earth System \ ( \ arXiv:2405.13063). arXiv

    arXiv 2024

  6. [6]

    , Dacre, H F

    charlton-perez2024 APACrefauthors Charlton-Perez , A J. , Dacre, H F. , Driscoll, S. , Gray, S L. , Harvey, B. , Harvey, N J. Volont \'e , A. APACrefauthors \ 2024 . Do AI Models Produce Better Weather Forecasts than Physics-Based Models? A Quantitative Evaluation Case Study of Storm Ciar \'a n Do AI models produce better weather forecasts than physics-ba...

    work page 2024

  7. [7]

    , Zhong, X

    chenFuXiCascadeMachine2023 APACrefauthors Chen, L. , Zhong, X. , Zhang, F. , Cheng, Y. , Xu, Y. , Qi, Y. \ Li, H. APACrefauthors \ 2023 . FuXi : A Cascade Machine Learning Forecasting System for 15-Day Global Weather Forecast FuXi : A cascade machine learning forecasting system for 15-day global weather forecast . npj Climate and Atmospheric Science 6 1 190

    work page 2023

  8. [8]

    APACrefauthors \ 2001

    ebert2001 APACrefauthors Ebert, E E. APACrefauthors \ 2001 . Ability of a Poor Man's Ensemble to Predict the Probability and Distribution of Precipitation Ability of a poor man's ensemble to predict the probability and distribution of precipitation . Monthly Weather Review 129 2461--2480

    work page 2001

Show all 42 references
  1. [9]

    APACrefauthors \ 2023

    ECMWF2023 APACrefauthors ECMWF . APACrefauthors \ 2023 . IFS Documentation – CY49R1 Operational Implementation Ifs documentation – cy49r1 operational implementation \ [ ]. Available at: https://www.ecmwf.int/en/publications/ifs-documentation

    2023

  2. [10]

    , Shi, X

    gaoPreDiffPrecipitationNowcasting2023 APACrefauthors Gao, Z. , Shi, X. , Han, B. , Wang, H. , Jin, X. , Maddix, D. Wang, Y. APACrefauthors \ 2023 . PreDiff : Precipitation Nowcasting with Latent Diffusion Models PreDiff : Precipitation Nowcasting with Latent Diffusion Models \...

    2023 arXiv

  3. [11]

    , Bai, L

    gongCasCastSkillfulHighresolution2024 APACrefauthors Gong, J. , Bai, L. , Ye, P. , Xu, W. , Liu, N. , Dai, J. Ouyang, W. APACrefauthors \ 2024 . CasCast : Skillful High-resolution Precipitation Nowcasting via Cascaded Modelling CasCast : Skillful High-resolution Precipitation ...

    2024 arXiv

  4. [12]

    APACrefauthors \ 2001

    hamill2001 APACrefauthors Hamill, T M. APACrefauthors \ 2001 . Interpretation of rank histograms for verifying ensemble forecasts Interpretation of rank histograms for verifying ensemble forecasts . Monthly Weather Review 129 3 550--560 . APACrefDOI doi:10.1175/1520-0493(2001)...

    2001 doi

  5. [13]

    , Bell, B

    hersbach2020era5 APACrefauthors Hersbach, H. , Bell, B. , Berrisford, P. , Hirahara, S. , Hor \'a nyi, A. , Mu \ n oz-Sabater, J. others APACrefauthors \ 2020 . The ERA5 global reanalysis The era5 global reanalysis . Quarterly Journal of the Royal Meteorological Society 146 73...

    2020

  6. [14]

    , Jain, A

    hoDenoisingDiffusionProbabilistic2020 APACrefauthors Ho, J. , Jain, A. \ Abbeel, P. APACrefauthors \ 2020 . Denoising Diffusion Probabilistic Models Denoising Diffusion Probabilistic Models \ ( \ arXiv:2006.11239). arXiv

    2020 arXiv

  7. [15]

    , Duffy, P

    iorio2004effects APACrefauthors Iorio, J. , Duffy, P. , Govindasamy, B. , Thompson, S. , Khairoutdinov, M. \ Randall, D. APACrefauthors \ 2004 . Effects of model resolution and subgrid-scale physics on the simulation of precipitation in the continental United States Effects of...

    2004

  8. [16]

    \ Fritsch, J M

    kain1993convective APACrefauthors Kain, J S. \ Fritsch, J M. APACrefauthors \ 1993 . Convective parameterization for mesoscale models: The Kain--Fritsch scheme Convective parameterization for mesoscale models: The kain--fritsch scheme . The representation of cumulus convection...

    1993

  9. [17]

    \ Welling, M

    kingma2022 APACrefauthors Kingma, D P. \ Welling, M. APACrefauthors \ 2022 . Auto- Encoding Variational Bayes Auto- Encoding Variational Bayes \ ( \ arXiv:1312.6114). arXiv

    2022 arXiv

  10. [18]

    , Sanchez-Gonzalez, A

    lam2023learning APACrefauthors Lam, R. , Sanchez-Gonzalez, A. , Willson, M. , Wirnsberger, P. , Fortunato, M. , Alet, F. others APACrefauthors \ 2023 . Learning skillful medium-range global weather forecasting Learning skillful medium-range global weather forecasting . Science...

    2023

  11. [19]

    , Cao, J

    liang2021 APACrefauthors Liang, J. , Cao, J. , Sun, G. , Zhang, K. , Gool, L V. \ Timofte, R. APACrefauthors \ 2021 . SwinIR : Image Restoration Using Swin Transformer SwinIR : Image Restoration Using Swin Transformer \ ( \ arXiv:2108.10257). arXiv

    2021 arXiv

  12. [20]

    , Hsu, K

    liu2024 APACrefauthors Liu, C C. , Hsu, K. , Peng, M S. , Chen, D S. , Chang, P L. , Hsiao, L F. Kuo, H C. APACrefauthors \ 2024 . Evaluation of Five Global AI Models for Predicting Weather in Eastern Asia and Western Pacific Evaluation of five global AI models for predicting ...

    2024

  13. [21]

    , Fung, J C H

    liuEnhancingQuantitativePrecipitation2024 APACrefauthors Liu, H. , Fung, J C H. , Lau, A K H. \ Li, Z. APACrefauthors \ 2024 . Enhancing Quantitative Precipitation Estimation of NWP Model With Fundamental Meteorological Variables and Transformer Based Deep Learning Model Enhan...

    2024

  14. [22]

    , Lin, Y

    liu2021 APACrefauthors Liu, Z. , Lin, Y. , Cao, Y. , Hu, H. , Wei, Y. , Zhang, Z. Guo, B. APACrefauthors \ 2021 . Swin Transformer : Hierarchical Vision Transformer Using Shifted Windows Swin Transformer : Hierarchical Vision Transformer using Shifted Windows \ ( \ arXiv:2103....

    2021 arXiv

  15. [23]

    \ Palmer, W

    marshall1948distribution APACrefauthors Marshall, J. \ Palmer, W. APACrefauthors \ 1948 . The distribution of raindrops with size The distribution of raindrops with size . Journal of Meteorology 5 4 165--166

    1948

  16. [24]

    \ Dudek, M

    Molinari1992Review APACrefauthors Molinari, J. \ Dudek, M. APACrefauthors \ 1992 . Parameterization of Convective Precipitation in Mesoscale Numerical Models: A Critical Review Parameterization of convective precipitation in mesoscale numerical models: A critical review . Mont...

    1992

  17. [25]

    , Buizza, R

    molteni1996ecmwf APACrefauthors Molteni, F. , Buizza, R. , Palmer, T N. \ Petroliagis, T. APACrefauthors \ 1996 . The ECMWF ensemble prediction system: Methodology and validation The ECMWF ensemble prediction system: Methodology and validation . Quarterly Journal of the Royal ...

    1996

  18. [26]

    , Shutts, G J

    palmer2005representing APACrefauthors Palmer, T N. , Shutts, G J. , Hagedorn, R. , Doblas-Reyes, F J. , Jung, T. \ Leutbecher, M. APACrefauthors \ 2005 . Representing model uncertainty in weather and climate prediction Representing model uncertainty in weather and climate pred...

    2005

  19. [27]

    , Subramanian, S

    pathak2022 APACrefauthors Pathak, J. , Subramanian, S. , Harrington, P. , Raja, S. , Chattopadhyay, A. , Mardani, M. Anandkumar, A. APACrefauthors \ 2022 . FourCastNet : A Global Data-driven High-resolution Weather Model Using Adaptive Fourier Neural Operators FourCastNet : A ...

    2022 arXiv

  20. [28]

    \ Xie, S

    peeblesScalableDiffusionModels2023 APACrefauthors Peebles, W. \ Xie, S. APACrefauthors \ 2023 . Scalable Diffusion Models with Transformers Scalable Diffusion Models with Transformers \ ( \ arXiv:2212.09748). arXiv

    2023 arXiv

  21. [29]

    \ Thuerey, N

    rasp2021data APACrefauthors Rasp, S. \ Thuerey, N. APACrefauthors \ 2021 . Data-driven medium-range weather prediction with a resnet pretrained on climate simulations: A new model for weatherbench Data-driven medium-range weather prediction with a resnet pretrained on climate ...

    2021

  22. [30]

    , Temperton, C

    ritchie1995implementation APACrefauthors Ritchie, H. , Temperton, C. , Simmons, A J. , Hortal, M. , Davies, T. , Dent, D. \ Hamrud, M. APACrefauthors \ 1995 . Implementation of the semi-Lagrangian method in a high-resolution version of the ECMWF forecast model Implementation o...

    1995

  23. [31]

    , Blattmann, A

    rombachHighResolutionImageSynthesis2022 APACrefauthors Rombach, R. , Blattmann, A. , Lorenz, D. , Esser, P. \ Ommer, B. APACrefauthors \ 2022 . High- Resolution Image Synthesis with Latent Diffusion Models High- Resolution Image Synthesis with Latent Diffusion Models \ ( \ arX...

    2022 arXiv

  24. [32]

    , Klemp, J B

    skamarock2005description APACrefauthors Skamarock, W C. , Klemp, J B. , Dudhia, J. , Gill, D O. , Barker, D M. , Duda, M G. Powers, J G. APACrefauthors \ 2005 . A description of the advanced research WRF version 2 A description of the advanced research wrf version 2 \ . Nation...

    2005

  25. [33]

    , Meng, C

    songDenoisingDiffusionImplicit2022 APACrefauthors Song, J. , Meng, C. \ Ermon, S. APACrefauthors \ 2022 . Denoising Diffusion Implicit Models Denoising Diffusion Implicit Models \ ( \ arXiv:2010.02502). arXiv

    2022 arXiv

  26. [34]

    APACrefauthors \ 1999

    talagrand1999 APACrefauthors Talagrand, O. APACrefauthors \ 1999 . Evaluation of probabilistic prediction systems Evaluation of probabilistic prediction systems . Proceedings of the ECMWF Workshop on Predictability

    1999

  27. [35]

    , Lean, H W

    tang2013benefits APACrefauthors Tang, Y. , Lean, H W. \ Bornemann, J. APACrefauthors \ 2013 . The benefits of the Met Office variable resolution NWP model for forecasting convection The benefits of the Met Office variable resolution nwp model for forecasting convection . Meteo...

    2013

  28. [36]

    APACrefauthors \ 1989

    tiedtke1989comprehensive APACrefauthors Tiedtke, M. APACrefauthors \ 1989 . A comprehensive mass flux scheme for cumulus parameterization in large-scale models A comprehensive mass flux scheme for cumulus parameterization in large-scale models . Monthly Weather Review 117 8 1779--1800

    1989

  29. [37]

    , Fasullo, J T

    trenberth2011 APACrefauthors Trenberth, K E. , Fasullo, J T. \ Mackaro, J. APACrefauthors \ 2011 . Atmospheric moisture transports from ocean to land and global energy flows in reanalyses Atmospheric moisture transports from ocean to land and global energy flows in reanalyses ...

    2011

  30. [38]

    , Bovik, A

    ssim APACrefauthors Wang, Z. , Bovik, A. , Sheikh, H. \ Simoncelli, E. APACrefauthors \ 2004 . Image quality assessment: from error visibility to structural similarity Image quality assessment: from error visibility to structural similarity . IEEE Transactions on Image Process...

    2004

  31. [39]

    APACrefauthors \ 2011

    wilks2011rank APACrefauthors Wilks, D. APACrefauthors \ 2011 . On the reliability of the rank histogram On the reliability of the rank histogram . Monthly Weather Review 139 1 311--316

    2011

  32. [40]

    \ Xiong, A Y

    xie2011 APACrefauthors Xie, P. \ Xiong, A Y. APACrefauthors \ 2011 . A Conceptual Model for Constructing High-Resolution Gauge-Satellite Merged Precipitation Analyses: GAUGE-SATELLITE MERGED PRECIP ANALYSIS A conceptual model for constructing high-resolution gauge-satellite me...

    2011

  33. [41]

    , Wang, Q J

    zhao2021which APACrefauthors Zhao, P. , Wang, Q J. , Wu, W. \ Yang, Q. APACrefauthors \ 2021 . Which precipitation forecasts to use? Deterministic versus coarser-resolution ensemble NWP models Which precipitation forecasts to use? deterministic versus coarser-resolution ensemb...

    2021

  34. [42]

    , Sun, J

    zhou2022 APACrefauthors Zhou, K. , Sun, J. , Zheng, Y. \ Zhang, Y. APACrefauthors \ 2022 . Quantitative Precipitation Forecast Experiment Based on Basic NWP Variables Using Deep Learning Quantitative Precipitation Forecast Experiment Based on Basic NWP Variables Using Deep Lea...

    2022

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.