pith. sign in

arxiv: 2605.17603 · v1 · pith:ORNWSTWFnew · submitted 2026-05-17 · ⚛️ physics.ao-ph · cs.LG

Longwang: Zero-Shot Global Spatiotemporal Precipitation Downscaling with a Latent Generative Prior

Yue Wang , Daniele Visioni This is my paper

Pith reviewed 2026-05-19 22:12 UTC · model grok-4.3

classification ⚛️ physics.ao-ph cs.LG
keywords precipitation downscalinglatent generative modelszero-shot learningclimate modelingspatiotemporal dataposterior samplingERA5 reanalysis
0
0 comments X

The pith

Longwang enables zero-shot downscaling of global precipitation to daily 10 km fields from monthly 100 km inputs by combining a context-conditioned latent generative prior with posterior sampling.

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

The paper introduces Longwang as a zero-shot framework that learns a generative model in latent space conditioned on context to produce high-resolution daily precipitation fields from coarse monthly global inputs. It combines this prior with a physical observation operator to draw samples consistent with the low-resolution data. The approach is tested on ERA5 reanalysis where it improves spatial detail, temporal consistency, and extreme value capture over simpler generative baselines. It also applies without retraining to historical simulations and future projections that differ substantially from the data used to build the prior. A sympathetic reader would care because fine-scale precipitation matters for impact studies yet remains unavailable from standard climate models.

Core claim

Longwang learns a context-conditioned latent generative prior and combines it with a physically informed observation operator through posterior sampling to generate daily O(10 km) precipitation fields from monthly O(100 km) inputs. On ERA5 reanalysis it outperforms standard posterior sampling with an unconditional generative prior in reconstructing fine-scale spatial patterns, preserving temporal coherence, and recovering extreme precipitation intensities. The framework further generalizes to historical climate simulations and future climate projections under substantial distribution shift.

What carries the argument

Context-conditioned latent generative prior used for posterior sampling together with a physically informed observation operator.

Load-bearing premise

A context-conditioned latent generative prior learned in an unsupervised manner can be combined with a physically informed observation operator to produce accurate posterior samples that generalize across significant distribution shifts in climate data.

What would settle it

If the downscaled daily fields on a held-out future climate projection dataset fail to recover extreme precipitation intensities or spatial patterns better than an unconditional prior when compared against any available high-resolution reference, the generalization claim would be falsified.

read the original abstract

High-resolution precipitation information is essential for climate impact assessment, yet global climate models remain too coarse to resolve key small-scale processes. Existing machine learning downscaling methods often require paired low- and high-resolution data for supervised learning, are tied to fixed regions or scale factors during inference, and can be computationally expensive to train and run in physical space. Here we introduce Longwang, a zero-shot latent generative framework for global spatiotemporal precipitation downscaling. Longwang learns a context-conditioned latent generative prior and combines it with a physically informed observation operator through posterior sampling, enabling daily O(10 km) precipitation fields to be generated from monthly O(100 km) inputs. On ERA5 reanalysis, Longwang outperforms standard posterior sampling with an unconditional generative prior in reconstructing fine-scale spatial patterns, preserving temporal coherence, and recovering extreme precipitation intensities. The framework further generalizes to historical climate simulations and future climate projections under substantial distribution shift.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces Longwang, a zero-shot latent generative framework for global spatiotemporal precipitation downscaling. It learns a context-conditioned latent generative prior and combines it with a physically informed observation operator through posterior sampling to generate daily O(10 km) precipitation fields from monthly O(100 km) inputs. On ERA5 reanalysis, it claims to outperform standard posterior sampling with an unconditional generative prior in reconstructing fine-scale spatial patterns, preserving temporal coherence, and recovering extreme precipitation intensities, while also generalizing to historical climate simulations and future projections under substantial distribution shift.

Significance. If the central claims hold, this could represent a meaningful advance in climate downscaling by enabling zero-shot, global application without paired low/high-resolution training data or region/scale-specific retraining. The combination of latent generative priors with physical observation operators and the reported robustness to distribution shifts in GCM outputs would be valuable for impact assessment, particularly if it reduces computational costs compared to supervised methods.

major comments (2)
  1. [Abstract] Abstract and results sections: The central claim of outperformance over unconditional priors on ERA5 (spatial patterns, temporal coherence, extreme intensities) and generalization to shifted GCM data is load-bearing, but the abstract provides no quantitative metrics, effect sizes, or baseline details; this limits evaluation of whether the gains are substantive or merely incremental, as noted in the soundness assessment of 4.0.
  2. [Methods] Methods: The assumption that an unsupervised/self-supervised context-conditioned latent prior encodes transferable fine-scale precipitation statistics invariant under climate-driven non-stationarities and GCM biases is central to the generalization claim, yet the observation operator supplies only integral constraints; without explicit ablations or tests on extreme tails in shifted regimes, this risks being the weakest link as highlighted by the skeptic.
minor comments (1)
  1. [Abstract] The abstract could more precisely define the coarse-graining details in the observation operator and the exact form of context conditioning to aid immediate understanding.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their positive assessment of the potential significance of our work and for the constructive comments. We address each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract and results sections: The central claim of outperformance over unconditional priors on ERA5 (spatial patterns, temporal coherence, extreme intensities) and generalization to shifted GCM data is load-bearing, but the abstract provides no quantitative metrics, effect sizes, or baseline details; this limits evaluation of whether the gains are substantive or merely incremental, as noted in the soundness assessment of 4.0.

    Authors: We agree that the abstract would benefit from explicit quantitative metrics to allow readers to assess the magnitude of improvements. In the revised manuscript we have updated the abstract to report key effect sizes from the ERA5 experiments, including improvements in spatial correlation coefficients, temporal autocorrelation preservation, and extreme-value quantile errors relative to the unconditional prior baseline. Baseline details for the unconditional generative prior are now also summarized. revision: yes

  2. Referee: [Methods] Methods: The assumption that an unsupervised/self-supervised context-conditioned latent prior encodes transferable fine-scale precipitation statistics invariant under climate-driven non-stationarities and GCM biases is central to the generalization claim, yet the observation operator supplies only integral constraints; without explicit ablations or tests on extreme tails in shifted regimes, this risks being the weakest link as highlighted by the skeptic.

    Authors: The context-conditioned prior is trained to encode fine-scale structures that depend on large-scale context, which we expect to be more invariant than unconditional statistics. We provide supporting evidence through direct generalization experiments on GCM outputs that exhibit substantial distribution shift. We nevertheless acknowledge that the paper does not contain dedicated ablations isolating extreme-tail behavior under those shifts. We will add such targeted analyses and ablations in the revision. revision: partial

Circularity Check

0 steps flagged

No significant circularity; learned prior plus physical operator yields independent empirical claims

full rationale

The paper's central derivation learns a context-conditioned latent generative prior in an unsupervised/self-supervised manner from data, then performs posterior sampling conditioned on a physically informed observation operator (coarse-graining daily 10 km to monthly 100 km). This produces fine-scale precipitation fields. The reported outperformance versus unconditional priors on ERA5 (spatial patterns, temporal coherence, extremes) and generalization to historical/future GCM data under distribution shift are evaluated empirically on held-out or shifted regimes. No step reduces a prediction to its inputs by construction, no self-definitional loops, no fitted parameters renamed as predictions, and no load-bearing self-citations or imported uniqueness theorems. The framework remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based solely on the abstract, no specific free parameters, axioms, or invented entities are detailed. The framework implicitly assumes the generative prior captures fine-scale statistics but provides no explicit ledger.

axioms (1)
  • domain assumption A context-conditioned latent generative prior can capture the statistical properties of fine-scale precipitation patterns.
    This underpins the zero-shot capability and posterior sampling approach described in the abstract.

pith-pipeline@v0.9.0 · 5688 in / 1306 out tokens · 39584 ms · 2026-05-19T22:12:10.458984+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

47 extracted references · 47 canonical work pages · 3 internal anchors

  1. [1]

    Zhang, W., Zhou, T. & Wu, P. Anthropogenic amplification of precipitation variability over the past century.Science385, 427–432 (2024). 20

    work page 2024

  2. [2]

    Schneider, T.et al.Climate goals and computing the future of clouds.Nature Climate Change7, 3–5 (2017)

    work page 2017

  3. [3]

    & Gutowski Jr, W

    Giorgi, F. & Gutowski Jr, W. J. Regional dynamical downscaling and the cordex initiative.Annual review of environment and resources40, 467–490 (2015)

    work page 2015

  4. [4]

    & Wenz, L

    Kotz, M., Levermann, A. & Wenz, L. The effect of rainfall changes on economic production.Nature601, 223–227 (2022)

    work page 2022

  5. [5]

    & Boers, N

    Hess, P. & Boers, N. Deep learning for improving numerical weather predic- tion of heavy rainfall.Journal of Advances in Modeling Earth Systems14, e2021MS002765 (2022)

    work page 2022

  6. [6]

    & Guti´ errez, J

    Ba˜ no-Medina, J., Manzanas, R. & Guti´ errez, J. M. Configuration and intercom- parison of deep learning neural models for statistical downscaling.Geoscientific Model Development13, 2109–2124 (2020)

    work page 2020

  7. [7]

    Wang, Y., Visioni, D., Kravitz, B., MacMartin, D. G. & Balwada, D. Mitigating non-stationarity in machine learning-based downscaling of climate projections. authorea(2026)

    work page 2026

  8. [8]

    & Lehrter, J

    Wang, F., Tian, D., Lowe, L., Kalin, L. & Lehrter, J. Deep learning for daily precipitation and temperature downscaling.Water Resources Research57, e2020WR029308 (2021)

    work page 2021

  9. [9]

    T., Chantry, M., Dueben, P

    Harris, L., McRae, A. T., Chantry, M., Dueben, P. D. & Palmer, T. N. A gener- ative deep learning approach to stochastic downscaling of precipitation forecasts. Journal of Advances in Modeling Earth Systems14, e2022MS003120 (2022)

    work page 2022

  10. [10]

    Mardani, M.et al.Residual corrective diffusion modeling for km-scale atmo- spheric downscaling.Communications Earth & Environment6, 124 (2025)

    work page 2025

  11. [11]

    & Boers, N

    Hess, P., Aich, M., Pan, B. & Boers, N. Fast, scale-adaptive and uncertainty- aware downscaling of earth system model fields with generative machine learning. Nature Machine Intelligence1–11 (2025)

    work page 2025

  12. [12]

    Wan, Z. Y.et al.Debias coarsely, sample conditionally: Statistical downscaling through optimal transport and probabilistic diffusion models.Advances in Neural Information Processing Systems36, 47749–47763 (2023)

    work page 2023

  13. [13]

    & Deck, K

    Bischoff, T. & Deck, K. Unpaired downscaling of fluid flows with diffusion bridges. Artificial Intelligence for the Earth Systems3, e230039 (2024)

    work page 2024

  14. [14]

    Srivastava, P.et al.Precipitation downscaling with spatiotemporal video dif- fusion.Advances in Neural Information Processing Systems37, 56374–56400 (2024). 21

    work page 2024

  15. [15]

    & Chwala, C

    Glawion, L., Polz, J., Kunstmann, H., Fersch, B. & Chwala, C. Global spatio- temporal era5 precipitation downscaling to km and sub-hourly scale using generative ai.npj Climate and Atmospheric Science8, 219 (2025)

    work page 2025

  16. [16]

    M., Ludwig, N

    Schmidt, J., Schmidt, L., Strnad, F. M., Ludwig, N. & Hennig, P. A generative framework for probabilistic, spatiotemporally coherent downscaling of climate simulation.npj Climate and Atmospheric Science8, 270 (2025)

    work page 2025

  17. [17]

    & Louppe, G

    Rozet, F. & Louppe, G. Score-based data assimilation.Advances in Neural Information Processing Systems36, 40521–40541 (2023)

    work page 2023

  18. [18]

    Fan, H.et al.Physically consistent global atmospheric data assimilation with machine learning in latent space.Science Advances12, eaea4248 (2026)

    work page 2026

  19. [19]

    Song, B.et al.Solving inverse problems with latent diffusion models via hard data consistency.arXiv preprint arXiv:2307.08123(2023)

    work page arXiv 2023

  20. [20]

    & Louppe, G

    Rozet, F., Andry, G., Lanusse, F. & Louppe, G. Learning diffusion priors from observations by expectation maximization.Advances in Neural Information Processing Systems37, 87647–87682 (2024)

    work page 2024

  21. [21]

    Kingma, D. P. & Welling, M. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114(2013)

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  22. [22]

    & Ommer, B

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P. & Ommer, B. High-resolution image synthesis with latent diffusion models.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition10684–10695 (2022)

    work page 2022

  23. [23]

    Turner, A. G. & Annamalai, H. Climate change and the south asian summer monsoon.Nature Climate Change2, 587–595 (2012)

    work page 2012

  24. [24]

    & Haug, G

    Schneider, T., Bischoff, T. & Haug, G. H. Migrations and dynamics of the intertropical convergence zone.nature513, 45–53 (2014)

    work page 2014

  25. [25]

    & Zhang, Y

    Ding, Y., Liang, P., Liu, Y. & Zhang, Y. Multiscale variability of meiyu and its prediction: A new review.Journal of Geophysical Research: Atmospheres125, e2019JD031496 (2020)

    work page 2020

  26. [26]

    Zveryaev, I. I. Seasonality in precipitation variability over europe.Journal of Geophysical Research: Atmospheres109(2004)

    work page 2004

  27. [27]

    Eyring, V.et al.Overview of the coupled model intercomparison project phase 6 (CMIP6) experimental design and organization.Geoscientific Model Development 9, 1937–1958 (2016)

    work page 1937

  28. [28]

    A.et al.A higher-resolution version of the max planck institute earth system model (MPI-ESM1

    M¨ uller, W. A.et al.A higher-resolution version of the max planck institute earth system model (MPI-ESM1. 2-HR).Journal of Advances in Modeling Earth Systems10, 1383–1413 (2018). 22

    work page 2018

  29. [29]

    Journal of Advances in Modeling Earth Systems12, e2019MS001916 (2020)

    Danabasoglu, G.et al.The community earth system model version 2 (cesm2). Journal of Advances in Modeling Earth Systems12, e2019MS001916 (2020)

    work page 2020

  30. [30]

    Dunne, J. P.et al.The gfdl earth system model version 4.1 (gfdl-esm 4.1): Overall coupled model description and simulation characteristics.Journal of Advances in Modeling Earth Systems12, e2019MS002015 (2020)

    work page 2020

  31. [31]

    0 climate model.Journal of Advances in Modeling Earth Systems11, 3691–3727 (2019)

    Held, I.et al.Structure and performance of gfdl’s cm4. 0 climate model.Journal of Advances in Modeling Earth Systems11, 3691–3727 (2019)

    work page 2019

  32. [32]

    model dev., 13, 6165–6200 (2020)

    Seland, Ø.et al.Overview of the norwegian earth system model (noresm2) and key climate response of cmip6 deck, historical, and scenario simulations, geosci. model dev., 13, 6165–6200 (2020)

    work page 2020

  33. [33]

    Lovato, T.et al.Cmip6 simulations with the cmcc earth system model (cmcc- esm2).Journal of Advances in Modeling Earth Systems14, e2021MS002814 (2022)

    work page 2022

  34. [34]

    Kravitz, B.et al.The geoengineering model intercomparison project phase 6 (GeoMIP6): Simulation design and preliminary results.Geoscientific Model Development8, 3379–3392 (2015)

    work page 2015

  35. [35]

    Visioni, D.et al.Identifying the sources of uncertainty in climate model simula- tions of solar radiation modification with the G6sulfur and G6solar geoengineering model intercomparison project (GeoMIP) simulations.Atmospheric Chemistry and Physics21, 10039–10063 (2021)

    work page 2021

  36. [36]

    Bonou, F.et al.Stratospheric sulfate aerosols impacts on west african monsoon precipitation using geomip models.Earth’s Future11(2023)

    work page 2023

  37. [37]

    Nature630, 891–898 (2024)

    Zhao, S.et al.Explainable el ni˜ no predictability from climate mode interactions. Nature630, 891–898 (2024)

    work page 2024

  38. [38]

    Xu, Z.et al.Evaluating variable-resolution cesm over china and western united states for use in water-energy nexus and impacts modeling.Journal of Geophysical Research: Atmospheres126, e2020JD034361 (2021)

    work page 2021

  39. [39]

    Hersbach, H.et al.The era5 global reanalysis.Quarterly journal of the royal meteorological society146, 1999–2049 (2020)

    work page 1999

  40. [40]

    & Balwada, D

    Lyu, J., Wang, Y., Pedersen, C., Jones, S. & Balwada, D. Multi-scale decom- position of sea surface height snapshots using machine learning.arXiv preprint arXiv:2409.17354(2024)

    work page arXiv 2024

  41. [41]

    Song, Y.et al.Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456(2020). 23

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  42. [42]

    & Nichol, A

    Dhariwal, P. & Nichol, A. Diffusion models beat gans on image synthesis. Advances in neural information processing systems34, 8780–8794 (2021)

    work page 2021

  43. [43]

    Nichol, A. Q. & Dhariwal, P. Improved denoising diffusion probabilistic models. Proceedings of Machine Learning Research139, 8162–8171 (2021)

    work page 2021

  44. [44]

    Anderson, B. D. Reverse-time diffusion equation models.Stochastic Processes and their Applications12, 313–326 (1982)

    work page 1982

  45. [45]

    Diffusion Posterior Sampling for General Noisy Inverse Problems

    Chung, H., Kim, J., Mccann, M. T., Klasky, M. L. & Ye, J. C. Diffusion posterior sampling for general noisy inverse problems.arXiv preprint arXiv:2209.14687 (2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  46. [46]

    2023 , publisher =

    Hersbach, H.et al.ERA5 hourly data on single levels from 1940 to present (2023). URL https://doi.org/10.24381/cds.adbb2d47

    work page doi:10.24381/cds.adbb2d47 1940

  47. [47]

    Cinquini, L.et al.The Earth System Grid Federation: An open infrastructure for access to distributed geospatial data.Future Generation Computer Systems 36, 400–417 (2014). 24

    work page 2014