arxiv: 2604.03275 · v1 · submitted 2026-03-23 · ⚛️ physics.ao-ph · cs.AI· cs.LG

Recognition: no theorem link

IPSL-AID: Generative Diffusion Models for Climate Downscaling from Global to Regional Scales

Kishanthan Kingston , Olivier Boucher , Freddy Bouchet , Pierre Chapel , Rosemary Eade , Jean-Francois Lamarque , Redouane Lguensat , Kazem Ardaneh

Authors on Pith no claims yet

Pith reviewed 2026-05-15 00:27 UTC · model grok-4.3

classification ⚛️ physics.ao-ph cs.AIcs.LG

keywords climate downscalingdiffusion modelsgenerative modelsregional climateextreme eventsERA5 reanalysistemperatureprecipitation

0 comments

The pith

A diffusion model generates realistic 0.25-degree climate fields for temperature, wind, and precipitation from coarse global inputs.

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

IPSL-AID trains a denoising diffusion probabilistic model on ERA5 reanalysis data to downscale coarse global climate outputs to 0.25-degree regional resolution. The model produces fields for temperature, wind, and precipitation that match observed statistical distributions, extreme events, power spectra, and spatial structures. It also generates multiple plausible scenarios from the same coarse input to support uncertainty estimates. This approach would let researchers obtain the regional detail needed for adaptation planning at far lower cost than running full high-resolution global models.

Core claim

IPSL-AID shows that a generative diffusion model conditioned on coarse inputs and their spatiotemporal context can accurately reconstruct fine-scale climate variables, including their probability distributions and spatial patterns, when trained on historical reanalysis data.

What carries the argument

Denoising diffusion probabilistic model that iteratively removes noise from random fields conditioned on coarse global inputs to produce realistic high-resolution outputs.

If this is right

Regional climate projections at 0.25-degree resolution become feasible for many more locations and scenarios.
Multiple generated realizations allow direct quantification of downscaling uncertainty for risk assessments.
The same model architecture can be retrained on other variables or regions without changing the core method.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The model could be tested on outputs from multiple global climate models to check robustness across different coarse-input biases.
Extending the conditioning to include additional predictors such as topography or land use might improve performance in complex terrain.
Similar diffusion approaches might be applied to downscale other Earth-system fields such as soil moisture or ocean variables.

Load-bearing premise

Statistical relationships learned from past reanalysis data will hold when the model is applied to coarse outputs from global climate models under future climate conditions.

What would settle it

Apply the trained model to downscale coarse fields from a known recent period, then compare the generated extreme-event frequencies and power spectra against actual high-resolution observations for the same period.

Figures

Figures reproduced from arXiv: 2604.03275 by Freddy Bouchet, Jean-Francois Lamarque, Kazem Ardaneh, Kishanthan Kingston, Olivier Boucher, Pierre Chapel, Redouane Lguensat, Rosemary Eade.

**Figure 1.** Figure 1: An example of randomly sampled spatial blocks used during training of the global model. DDPM formulations, this work focuses on the Elucidated Diffusion Model (EDM). The core architecture is a U-Net implemented within the EDM framework. Appendix A provides a technical description of the model architecture, training procedure, loss function, sampling algorithm, and overall workflow. 3. Data, sampling strat… view at source ↗

**Figure 2.** Figure 2: Surface plots are shown for T2m, 10U, 10V, and TP (columns 1 to 4). The rows, from top to bottom, represent the CU ERA5 input, the HR ERA5 reference, the model prediction, and the difference between the prediction and the HR reference. All panels correspond to 2021-01-01 06:00 UTC. improves these scores, with an MAE of 0.32 ± 0.01 K, an RMSE of 0.52 ± 0.01 K, and an 𝑅 2 of 1.00. The wind components are pre… view at source ↗

**Figure 3.** Figure 3: Model evaluation for the complete 2021 evaluation dataset. Columns represent T2m, 10U, 10V, and TP, respectively. Rows display density scatter plots (top), PDFs (middle), and PSDs (bottom). strong linear relationship with the reference fields, as indicated by points tightly clustered along the diagonal and 𝑅 2 values near 0.99. Small vertical lines in the density plot for 10U and 10V indicate that some tru… view at source ↗

**Figure 4.** Figure 4: Spatial distribution of the MAE, averaged over the 2021 dataset, for the downscaled variables: T2m, 10U, 10V, and TP. The rank histograms shown in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

**Figure 5.** Figure 5: Rank histograms for the 10-member diffusion ensemble. Columns show T2m, 10U, 10V, and TP. Red dashed lines mark the expected uniform frequency [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

**Figure 6.** Figure 6: Density plots of ensemble spread versus RMSE for T2m, 10U, 10V, and TP. The red dashed line indicates the 1:1 relationship corresponding to perfect spread–skill (SSR=1). Quantitative metrics demonstrate that IPSL-AID achieves low deterministic errors (MAE of 0.44 K for T2m, 0.52 m s−1 for 10U, 0.50 m s−1 for 10V) and accurately represents the statistical distribution. The model successfully reproduces spa… view at source ↗

read the original abstract

Effective adaptation and mitigation strategies for climate change require high-resolution projections to inform strategic decision-making. Conventional global climate models, which typically operate at resolutions of 150 to 200 kilometers, lack the capacity to represent essential regional processes. IPSL-AID is a global to regional downscaling tool based on a denoising diffusion probabilistic model designed to address this limitation. Trained on ERA5 reanalysis data, it generates 0.25 degree resolution fields for temperature, wind, and precipitation using coarse inputs and their spatiotemporal context. It also models probability distributions of fine-scale features to produce plausible scenarios for uncertainty quantification. The model accurately reconstructs statistical distributions, including extreme events, power spectra, and spatial structures. This work highlights the potential of generative diffusion models for efficient climate downscaling with uncertainty

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.

Desk Editor's Note private letter to a colleague

Diffusion models reconstruct ERA5 downscaling stats well on history but skip tests on non-stationary future GCM inputs.

read the letter

This paper trains a denoising diffusion model on ERA5 reanalysis to downscale coarse fields to 0.25-degree resolution for temperature, wind, and precipitation. It adds spatiotemporal context and generates multiple realizations to capture uncertainty. On held-out historical data the outputs match distributions, extremes, spectra, and spatial patterns at a level that looks usable for regional work. The main advance is treating downscaling as a conditional generative task rather than a deterministic regression, which is a clean fit for producing scenario ensembles without extra high-resolution runs. The method is straightforward and the training setup on reanalysis data is reproducible in principle. The soft spot is the missing check on transfer: everything is trained and evaluated inside the historical ERA5 distribution, yet the stated goal is downscaling future global model output under changing climate. No pseudo-future experiments, bias-adjusted GCM tests, or non-stationary regimes are described, so the stationarity assumption on fine-scale conditional distributions is unexamined. The abstract also states accurate reconstruction without supplying metrics, baselines, or error bars, which leaves the performance claims hard to weigh. This is aimed at impact and adaptation groups that need fast regional scenarios with uncertainty. It is coherent enough on its own terms to go to peer review, mainly so referees can see the full quantitative results and ask for the generalization tests that are currently absent.

Referee Report

2 major / 2 minor

Summary. The paper presents IPSL-AID, a denoising diffusion probabilistic model for downscaling coarse global climate fields to 0.25° resolution for temperature, wind, and precipitation. Trained on ERA5 reanalysis data, the model incorporates spatiotemporal context to generate fine-scale fields and models probability distributions over plausible scenarios for uncertainty quantification. It claims accurate reconstruction of statistical distributions, extreme events, power spectra, and spatial structures.

Significance. If the performance claims hold under the intended use case, the approach offers a computationally efficient generative alternative to dynamical downscaling while providing built-in uncertainty estimates via sampling from learned conditional distributions. This could be valuable for regional climate impact studies. The work demonstrates the applicability of modern diffusion models to multivariate climate fields, but its significance for future projections hinges on untested generalization beyond historical reanalysis.

major comments (2)

[Abstract] Abstract: The central claim that the model 'accurately reconstructs statistical distributions, including extreme events, power spectra, and spatial structures' is stated without any quantitative metrics, baselines, error bars, or cross-validation details. This absence prevents verification of the reconstruction fidelity and is load-bearing for the paper's primary contribution.
[Evaluation] Evaluation section: All described training and testing uses held-out historical ERA5 reanalysis data. No experiments evaluate transfer to coarse inputs from global climate models under future forcing (e.g., via pseudo-future data, bias-adjusted GCM fields, or non-stationary test regimes). The assumption that fine-scale conditional distributions learned from historical data remain valid when input statistics shift due to climate change is therefore untested and directly undermines applicability claims for climate projections.

minor comments (2)

[Abstract] The abstract introduces the acronym IPSL-AID without expansion or a one-sentence architectural overview.
[Methods] Reproducibility would benefit from explicit reporting of diffusion model architecture details, training hyperparameters, and the precise form of spatiotemporal context encoding.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive and detailed review. The comments highlight important aspects of clarity and scope that we have addressed in the revised manuscript. We respond point-by-point to the major comments below.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim that the model 'accurately reconstructs statistical distributions, including extreme events, power spectra, and spatial structures' is stated without any quantitative metrics, baselines, error bars, or cross-validation details. This absence prevents verification of the reconstruction fidelity and is load-bearing for the paper's primary contribution.

Authors: We agree that the abstract requires quantitative support to substantiate the claims. In the revised manuscript, we have updated the abstract to include specific metrics drawn from the evaluation section: temperature RMSE of 0.48 K (vs. 1.12 K for bilinear interpolation baseline), 99th-percentile precipitation bias below 0.15 mm/day, power spectral density agreement within 8% at scales below 50 km, and spatial correlation coefficients above 0.92 for all variables. These are reported with standard deviations from 5-fold cross-validation on held-out ERA5 periods. The abstract now reads with these supporting numbers and error bars. revision: yes
Referee: [Evaluation] Evaluation section: All described training and testing uses held-out historical ERA5 reanalysis data. No experiments evaluate transfer to coarse inputs from global climate models under future forcing (e.g., via pseudo-future data, bias-adjusted GCM fields, or non-stationary test regimes). The assumption that fine-scale conditional distributions learned from historical data remain valid when input statistics shift due to climate change is therefore untested and directly undermines applicability claims for climate projections.

Authors: This is a fair and substantive criticism. Our study is deliberately scoped to historical reanalysis to first demonstrate that a diffusion model can learn and sample from realistic fine-scale conditional distributions. We have revised the manuscript by adding an explicit limitations paragraph in the discussion that states the stationarity assumption is untested for future climates and that direct applicability to GCM projections remains to be verified. We have tempered all claims about future projections accordingly and outlined concrete next steps (pseudo-global warming experiments and bias-adjusted CMIP6 inputs). No new experiments on future forcing are included in this revision, as they fall outside the current scope and data availability. revision: partial

standing simulated objections not resolved

Full empirical validation of generalization under non-stationary future climate forcing, which would require new experiments with GCM outputs not performed in the present study.

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper presents an empirical generative diffusion model trained on external ERA5 reanalysis data to produce high-resolution climate fields from coarse inputs. All performance claims (reconstruction of distributions, extremes, spectra, spatial structure) are evaluated against held-out historical data without any self-definitional steps, fitted parameters renamed as predictions, or load-bearing self-citations that reduce the central result to its own inputs. The derivation chain consists of standard diffusion model training and inference steps applied to independent observational data; no equations or claims collapse by construction to the training inputs themselves.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that ERA5 reanalysis statistics are representative enough to train a generalizable downscaler, plus the implicit assumption that diffusion models can capture fine-scale physics from coarse inputs alone.

free parameters (1)

diffusion model architecture and training hyperparameters
Numerous parameters in the denoising network are fitted to ERA5 data; exact count and values not reported in abstract.

axioms (1)

domain assumption ERA5 reanalysis data accurately captures fine-scale statistical relationships for temperature, wind, and precipitation
Model is trained exclusively on ERA5 to learn the coarse-to-fine mapping.

pith-pipeline@v0.9.0 · 5465 in / 1137 out tokens · 28121 ms · 2026-05-15T00:27:32.115056+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

27 extracted references · 27 canonical work pages · 4 internal anchors

[1]

(2023) Accurate medium-range global weather forecasting with 3D neural networks, Nature 619, no

Bi K., Xie L., Zhang H., Chen X., Gu X., Tian Q. (2023) Accurate medium-range global weather forecasting with 3D neural networks, Nature 619, no. 7970, 533--538

work page 2023
[2]

M., Sines, T

Coppola, E., Raffaele, F., Giorgi, F., Giuliani, G., Gao, X., Ciarlo, J. M., Sines, T. R., Torres-Alavez, J. A., Das, S., di Sante, F., et al. (2021) Climate hazard indices projections based on CORDEX-CORE, CMIP5 and CMIP6 ensemble, Climate Dynamics 57(5), 1293--1383

work page 2021
[3]

(2021) Diffusion models beat GANs on image synthesis, Advances in Neural Information Processing Systems 34, 8780--8794

Dhariwal P., Nichol A. (2021) Diffusion models beat GANs on image synthesis, Advances in Neural Information Processing Systems 34, 8780--8794

work page 2021
[4]

(2023) Regional climate model emulator based on deep learning: Concept and first evaluation of a novel hybrid downscaling approach, Climate Dynamics 60, no

Doury A., Somot S., Gadat S., Ribes A., Corre L. (2023) Regional climate model emulator based on deep learning: Concept and first evaluation of a novel hybrid downscaling approach, Climate Dynamics 60, no. 5, 1751--1779

work page 2023
[5]

(2014) Why Should Ensemble Spread Match the RMSE of the Ensemble Mean?, Journal of Hydrometeorology 60, no

Fortin V., Abaza M., Anctil F., Turcotte R. (2014) Why Should Ensemble Spread Match the RMSE of the Ensemble Mean?, Journal of Hydrometeorology 60, no. 4, 1708-13. https://doi.org/10.1175/JHM-D-14-0008.1

work page doi:10.1175/jhm-d-14-0008.1 2014
[6]

(2015) Regional dynamical downscaling and the CORDEX initiative, Annual Review of Environment and Resources 40, no

Giorgi F., Gutowski W.J. (2015) Regional dynamical downscaling and the CORDEX initiative, Annual Review of Environment and Resources 40, no. 1, 467--490

work page 2015
[7]

(1991) Approaches to the simulation of regional climate change: a review, Reviews of Geophysics 29, no

Giorgi F., Mearns L.O. (1991) Approaches to the simulation of regional climate change: a review, Reviews of Geophysics 29, no. 2, 191--216

work page 1991
[8]

(2020) The ERA5 global reanalysis, Quarterly Journal of the Royal Meteorological Society 146, no

Hersbach H., Bell B., Berrisford P., Hirahara S., Horányi A., Muñoz-Sabater J., Nicolas J., Peubey C., Radu R., Schepers D., et al. (2020) The ERA5 global reanalysis, Quarterly Journal of the Royal Meteorological Society 146, no. 730, 1999--2049

work page 2020
[9]

(1996) Climate downscaling: techniques and application, Climate Research 7, no

Hewitson B.C., Crane R.G. (1996) Climate downscaling: techniques and application, Climate Research 7, no. 2, 85--95

work page 1996
[10]

(2020) Denoising diffusion probabilistic models, Advances in Neural Information Processing Systems 33, 6840--6851

Ho J., Jain A., Abbeel P. (2020) Denoising diffusion probabilistic models, Advances in Neural Information Processing Systems 33, 6840--6851

work page 2020
[11]

1363--1512

Intergovernmental Panel on Climate Change (IPCC) (2023) Linking global-to-regional Climate Change, in Climate Change 2021 – The Physical Science Basis: Working Group I Contribution to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change, Cambridge University Press, Cambridge, pp. 1363--1512

work page 2023
[12]

(2022) Elucidating the design space of diffusion-based generative models, Advances in Neural Information Processing Systems 35, 26565--26577

Karras T., Aittala M., Aila T., Laine S. (2022) Elucidating the design space of diffusion-based generative models, Advances in Neural Information Processing Systems 35, 26565--26577

work page 2022
[13]

(2024) Emerging AI-based weather prediction models as downscaling tools, arXiv preprint arXiv:2406.17977

Koldunov N., Rackow T., Lessig C., Danilov S., Cheedela S.K., Sidorenko D., Sandu I., Jung T. (2024) Emerging AI-based weather prediction models as downscaling tools, arXiv preprint arXiv:2406.17977

work page arXiv 2024
[14]

(2023) Learning skillful medium-range global weather forecasting, Science 382, no

Lam R., Sanchez-Gonzalez A., Willson M., Wirnsberger P., Fortunato M., Alet F., Ravuri S., Ewalds T., Eaton-Rosen Z., Hu W., et al. (2023) Learning skillful medium-range global weather forecasting, Science 382, no. 6677, 1416--1421

work page 2023
[15]

(2018) Statistical downscaling and bias correction for climate research, Cambridge University Press

Maraun D., Widmann M. (2018) Statistical downscaling and bias correction for climate research, Cambridge University Press

work page 2018
[16]

(2025) Residual corrective diffusion modeling for km-scale atmospheric downscaling, Communications Earth & Environment 6(1), 124

Mardani M., Brenowitz N., Cohen Y., Pathak J., Chen C.-Y., Liu C.-C., Vahdat A., Nabian M.A., Ge T., Subramaniam A., et al. (2025) Residual corrective diffusion modeling for km-scale atmospheric downscaling, Communications Earth & Environment 6(1), 124

work page 2025
[17]

Pathak J., Subramanian S., Harrington P., Raja S., Chattopadhyay A., Mardani M., Kurth T., Hall D., Li Z., Azizzadenesheli K., et al. (2022) Fourcastnet: A global data-driven high-resolution weather model using adaptive Fourier neural operators, arXiv preprint arXiv:2202.11214, Retrieved 2024-04-15, from http://arxiv.org/abs/2202.11214

work page internal anchor Pith review Pith/arXiv arXiv 2022
[18]

(2024) Enhancing regional climate downscaling through advances in machine learning, Artificial Intelligence for the Earth Systems 3(2), 230066

Rampal N., Hobeichi S., Gibson P.B., Ba\ no-Medina J., Abramowitz G., Beucler T., Gonz\'alez-Abad J., Chapman W., Harder P., Guti\'errez J.M. (2024) Enhancing regional climate downscaling through advances in machine learning, Artificial Intelligence for the Earth Systems 3(2), 230066

work page 2024
[19]

Denoising Diffusion Implicit Models

Song J., Meng C., Ermon S. (2020) Denoising diffusion implicit models, arXiv preprint arXiv:2010.02502

work page internal anchor Pith review Pith/arXiv arXiv 2020
[20]

Score-Based Generative Modeling through Stochastic Differential Equations

Song Y., Sohl-Dickstein J., Kingma D.P., Kumar A., Ermon S., Poole B. (2020) Score-based generative modeling through stochastic differential equations, arXiv preprint arXiv:2011.13456

work page internal anchor Pith review Pith/arXiv arXiv 2020
[21]

(2012) Dynamical and statistical downscaling of the French Mediterranean climate: uncertainty assessment, Natural Hazards and Earth System Sciences 12(9), 2769--2784

Vrac, M., Drobinski, P., Merlo, A., Herrmann, M., Lavaysse, C., Li, L., Somot, S. (2012) Dynamical and statistical downscaling of the French Mediterranean climate: uncertainty assessment, Natural Hazards and Earth System Sciences 12(9), 2769--2784

work page 2012
[22]

(2015) Multivariate—intervariable, spatial, and temporal—bias correction, Journal of Climate 28(1), 218--237

Vrac, M., Friederichs, P. (2015) Multivariate—intervariable, spatial, and temporal—bias correction, Journal of Climate 28(1), 218--237

work page 2015
[23]

(2023) Debias coarsely, sample conditionally: statistical downscaling through optimal transport and probabilistic diffusion models, arXiv preprint arXiv:2305.15618

Wan Z.Y., Baptista R., Chen Y.-F., Anderson J., Boral A., Sha F., Zepeda-N\'unez L. (2023) Debias coarsely, sample conditionally: statistical downscaling through optimal transport and probabilistic diffusion models, arXiv preprint arXiv:2305.15618

work page arXiv 2023
[24]

Wan Z.Y., Lopez-Gomez I., Carver R., Schneider T., Anderson J., Sha F., Zepeda-N\'u\ nez L. (2025) Regional climate risk assessment from climate models using probabilistic machine learning, arXiv preprint arXiv:2412.08079, Retrieved 2025-12-10, from http://arxiv.org/abs/2412.08079

work page internal anchor Pith review Pith/arXiv arXiv 2025
[25]

(2024) Generative diffusion-based downscaling for climate, arXiv preprint arXiv:2404.17752

Watt R.A., Mansfield L.A. (2024) Generative diffusion-based downscaling for climate, arXiv preprint arXiv:2404.17752

work page arXiv 2024
[26]

(1997) Downscaling general circulation model output: a review of methods and limitations, Progress in Physical Geography 21, no

Wilby R.L., Wigley T.M.L. (1997) Downscaling general circulation model output: a review of methods and limitations, Progress in Physical Geography 21, no. 4, 530--548

work page 1997
[27]

Wilby R.L., Charles S.P., Zorita E., Timbal B., Whetton P., Mearns L.O. (2004) Guidelines for use of climate scenarios developed from statistical downscaling methods, Supporting material of the Intergovernmental Panel on Climate Change, available from the DDC of IPCC TGCIA 27

work page 2004