pith. sign in

arxiv: 2509.26258 · v3 · submitted 2025-09-30 · ⚛️ physics.ao-ph · physics.data-an· stat.AP· stat.ML

EnScale: Temporally-consistent multivariate generative downscaling via proper scoring rules

Maybritt Schillinger , Maxim Samarin , Xinwei Shen , Reto Knutti , Nicolai Meinshausen This is my paper

Pith reviewed 2026-05-18 11:52 UTC · model grok-4.3

classification ⚛️ physics.ao-ph physics.data-anstat.APstat.ML
keywords climate downscalinggenerative modelsenergy scoreregional climate modelssuper-resolutionproper scoring rulesmultivariate emulation
0
0 comments X

The pith

EnScale emulates regional climate model outputs from global models using a generative framework optimized with proper scoring rules for efficiency and consistency.

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

EnScale is a generative machine learning framework designed to downscale coarse global circulation model outputs to high-resolution fields that match regional climate model outputs. The method trains on multiple pairs of GCM and RCM data by first adjusting large-scale mismatches and then applying a super-resolution step. This super-resolution employs novel sparse local stochastic layers to handle high-dimensional outputs efficiently. Both steps are optimized using the energy score as a proper scoring rule, enabling the model to capture the full conditional distribution. The approach reduces computational costs by roughly an order of magnitude while producing spatially and temporally consistent multivariate fields for variables such as temperature and precipitation over Central Europe.

Core claim

EnScale emulates the full GCM-to-RCM map by training on multiple pairs of GCM and corresponding RCM data. It first adjusts large-scale mismatches between GCM and coarsened RCM data, followed by a super-resolution step to generate high-resolution fields using a novel class of sparse local stochastic layers. Both steps employ generative models optimized with the energy score, a proper scoring rule. This jointly emulates multiple variables such as temperature, precipitation, solar radiation, and wind that are spatially consistent over Central Europe, with a variant EnScale-t enabling temporally consistent downscaling.

What carries the argument

Two-step generative framework of large-scale mismatch adjustment followed by super-resolution via sparse local stochastic layers, with both steps optimized using the energy score proper scoring rule.

Load-bearing premise

The paired GCM and RCM training data sufficiently represent the full conditional distribution of high-resolution fields, allowing the generative model to accurately capture spatial consistency, temporal structure, extremes, and multivariate dependencies across the target domain.

What would settle it

If downscaled outputs from EnScale applied to new GCM inputs fail to match the observed statistical properties, spatial patterns, extremes, or multivariate dependencies of corresponding RCM simulations on validation data, the emulation claim would not hold.

Figures

Figures reproduced from arXiv: 2509.26258 by Maxim Samarin, Maybritt Schillinger, Nicolai Meinshausen, Reto Knutti, Xinwei Shen.

Figure 1
Figure 1. Figure 1: Illustration of the dataset. The first row shows the GCM data from CNRM-CM5 on two example days [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Downscaling via coarse correction for EnScale. We approximate the conditional distribution of RCM data 𝑌 given GCM data 𝑋 with a two-step approach. In the second row, 𝑍 represents RCM data manually coarsened through average pooling. The map learning the conditional 𝑝𝑍|𝑋 is called the coarse model, and the map for the conditional 𝑝𝑌 |𝑍 the super-resolution model. All 𝑋, 𝑍, 𝑌 include multiple climate variabl… view at source ↗
Figure 3
Figure 3. Figure 3: Time series generation with temporal consistency in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sparse local stochastic layers from EnScale’s super-resolution model. As an example, we demonstrate modeling the distribution in an example pixel of interest (top left corner, orange); the same procedure is applied to all pixels. For each intermediate map (small arrows), light blue shaded pixels serve as inputs and orange pixels as the targets. First, a deterministic upsampling step processes each target v… view at source ↗
Figure 5
Figure 5. Figure 5: Summary of performance of EnScale compared to the benchmarks in several selected categories, shown for the interpolation test period (2030-39). Energy score (see Sec. 5.4.1), Calibration (see Sec. 5.4.3), Spatial structure (see Sec. 5.4.2), Temporal structure (see Sec. 5.4.5), Extremes (see Sec. 5.4.4), Multivariate dependencies (Sec. 5.5). The chosen metrics for the categories are outlined in more detail … view at source ↗
Figure 6
Figure 6. Figure 6: Time series examples. Solid lines show the RCM time series (green) and a single randomly chosen realization [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Exemplary samples from EnScale for all variables. The first column presents the unseen target RCM data by ALADIN63 driven by CNRM-CM5 on day 2035-05-06. Columns 2-4 show three corresponding random samples from EnScale. We chose a day with average performance, i.e., the EnScale-loss is roughly equal to the median score of all days in the interpolation test set. inference times can likely be improved for bot… view at source ↗
Figure 8
Figure 8. Figure 8: EnScale and EnScale-t show errors in calibration with MCB values between 0.09 and 0.17, outperforming CorrDiff. Also analogues and GAN reach higher MCB scores than EnScale (not shown). For reference, MCB scores 19 [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 8
Figure 8. Figure 8: Rank histograms for evaluating calibration. For each day, we calculate the spatial mean and the spatial [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Extreme quantiles for the summer season (June, July, August) compared to ALADIN63 driven by CNRM [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: As Fig. 6, omitting CorrDiff, but presenting all four target variables instead. [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Correlations between pairs of variables. The first column shows the RCM’s pairwise correlation between the [PITH_FULL_IMAGE:figures/full_fig_p025_11.png] view at source ↗
read the original abstract

The practical use of future climate projections from global circulation models (GCMs) is often limited by their coarse spatial resolution, requiring downscaling to generate high-resolution data. Regional climate models (RCMs) provide this refinement, but are computationally expensive. To address this issue, machine learning (ML) models can learn the downscaling function, mapping coarse GCM outputs to high-resolution fields. Among these, generative approaches aim to capture the full conditional distribution of RCM data given coarse-scale GCM data, which is characterized by large variability and thus challenging to model accurately. We introduce EnScale, a generative ML framework emulating the full GCM-to-RCM map by training on multiple pairs of GCM and corresponding RCM data. It first adjusts large-scale mismatches between GCM and coarsened RCM data, followed by a super-resolution step to generate high-resolution fields. To efficiently model the high-dimensional output, the super-resolution step employs a novel class of sparse local stochastic layers. Both steps employ generative models optimized with the energy score, a proper scoring rule. Compared to state-of-the-art ML downscaling approaches, our setup reduces computational cost by about one order of magnitude. EnScale jointly emulates multiple variables -- temperature, precipitation, solar radiation, and wind -- spatially consistent over Central Europe. In addition, we propose a variant EnScale-t that enables temporally consistent downscaling. We establish a comprehensive evaluation framework across various categories including calibration, spatial and temporal structure, extremes, and multivariate dependencies. Comparison with diverse benchmarks demonstrates EnScale(-t)'s competitive performance and computational efficiency, offering a promising approach for accurate and temporally consistent RCM emulation.

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

3 major / 2 minor

Summary. The manuscript introduces EnScale, a generative ML framework for emulating the GCM-to-RCM downscaling map. It trains on multiple historical GCM-RCM pairs using a two-stage process: large-scale mismatch correction followed by super-resolution via novel sparse local stochastic layers. Both stages are optimized with the energy score proper scoring rule. The method produces spatially consistent multivariate fields (temperature, precipitation, solar radiation, wind) over Central Europe, with an EnScale-t variant for temporal consistency. A comprehensive evaluation framework assesses calibration, spatial/temporal structure, extremes, and multivariate dependencies, claiming competitive performance versus benchmarks at roughly one order of magnitude lower computational cost.

Significance. If the performance claims hold under distribution shift, EnScale would offer a practical, uncertainty-aware alternative to expensive RCM simulations for high-resolution climate projections. The use of proper scoring rules for training, the sparse stochastic layers for high-dimensional outputs, and the explicit temporal-consistency variant address longstanding challenges in multivariate generative downscaling. The proposed evaluation categories could serve as a useful template for future work.

major comments (3)
  1. [Abstract and Methods (training data section)] Abstract and training-procedure description: The central claim is that EnScale emulates the full conditional p(RCM|GCM) for use in future projections. However, training occurs exclusively on historical paired data; the two-stage architecture and sparse local stochastic layers provide no explicit mechanism or guarantee for extrapolating beyond the observed support when GCM large-scale statistics shift under RCP/SSP scenarios.
  2. [Methods (super-resolution and sparse layers)] Super-resolution step and sparse local stochastic layers: The layers are introduced to capture high-dimensional variability efficiently while preserving spatial consistency. Without a precise definition of the sparsity pattern, locality radius, or how stochasticity is injected (e.g., in the relevant methods subsection or equation), it is difficult to verify that the claimed preservation of extremes and multivariate dependencies follows from the architecture rather than from post-hoc evaluation.
  3. [Results and Evaluation sections] Evaluation framework and results: The manuscript proposes a broad set of diagnostics (calibration, extremes, temporal structure). Specific quantitative evidence—such as energy-score values, extreme-value metrics, or temporal autocorrelation scores for EnScale-t versus benchmarks in the results tables—is required to substantiate that the generative outputs are competitive rather than merely plausible.
minor comments (2)
  1. [Methods] Notation for the energy score should be introduced once with a clear reference to its definition for multivariate fields.
  2. [Figures] Figure captions for multivariate and temporal diagnostics would benefit from explicit mention of which variables and lead times are shown.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed review of our manuscript. We address each major comment point by point below, providing clarifications and indicating where revisions have been made to improve the manuscript. Our goal is to enhance the rigor and transparency of the presentation of EnScale.

read point-by-point responses

  1. Referee: [Abstract and Methods (training data section)] Abstract and training-procedure description: The central claim is that EnScale emulates the full conditional p(RCM|GCM) for use in future projections. However, training occurs exclusively on historical paired data; the two-stage architecture and sparse local stochastic layers provide no explicit mechanism or guarantee for extrapolating beyond the observed support when GCM large-scale statistics shift under RCP/SSP scenarios.

    Authors: We appreciate the referee highlighting this important consideration regarding generalization. EnScale is trained on historical GCM-RCM pairs to learn an approximation to the conditional distribution p(RCM|GCM). When applied to future projections, the model is used with future GCM outputs under the standard assumption in statistical downscaling that the learned relationship generalizes to altered large-scale conditions. This is an implicit rather than explicit mechanism, and we acknowledge the referee's point that no architectural feature guarantees performance under distribution shift. In the revised manuscript, we have updated the abstract and added a dedicated paragraph in the Discussion section to explicitly state this assumption, discuss potential limitations under RCP/SSP scenarios, and suggest avenues for future work such as domain adaptation. We believe this clarifies the scope of the claims. revision: yes

  2. Referee: [Methods (super-resolution and sparse layers)] Super-resolution step and sparse local stochastic layers: The layers are introduced to capture high-dimensional variability efficiently while preserving spatial consistency. Without a precise definition of the sparsity pattern, locality radius, or how stochasticity is injected (e.g., in the relevant methods subsection or equation), it is difficult to verify that the claimed preservation of extremes and multivariate dependencies follows from the architecture rather than from post-hoc evaluation.

    Authors: We thank the referee for this observation on methodological clarity. The original manuscript describes the sparse local stochastic layers in the Methods section as a means to efficiently model high-dimensional outputs while maintaining spatial consistency. However, we agree that a more formal specification of the sparsity pattern, locality radius, and stochastic injection process would strengthen verifiability. In the revised version, we have expanded the relevant subsection to include precise definitions: the sparsity pattern is defined via a local neighborhood mask, the locality radius is set according to variable-specific correlation lengths, and stochasticity is injected through scaled Gaussian perturbations at each local patch. Updated equations and a supplementary illustration of the mask have been added to show how these choices support the preservation of extremes and dependencies directly from the architecture. revision: yes

  3. Referee: [Results and Evaluation sections] Evaluation framework and results: The manuscript proposes a broad set of diagnostics (calibration, extremes, temporal structure). Specific quantitative evidence—such as energy-score values, extreme-value metrics, or temporal autocorrelation scores for EnScale-t versus benchmarks in the results tables—is required to substantiate that the generative outputs are competitive rather than merely plausible.

    Authors: We agree that explicit quantitative metrics are necessary to support the competitiveness claims. The manuscript presents a comprehensive evaluation framework with diagnostics across calibration, spatial/temporal structure, extremes, and multivariate dependencies, along with comparisons to benchmarks in figures and tables. To address the request for specific numbers, we have revised the Results section to include dedicated tables reporting exact energy score values, extreme-value metrics (e.g., errors in high quantiles for precipitation), and temporal autocorrelation scores for EnScale-t versus the benchmarks. These additions provide the requested quantitative evidence and confirm the competitive performance while highlighting the computational advantages. revision: yes

Circularity Check

0 steps flagged

No significant circularity; EnScale derives from independent training procedure and external benchmarks

full rationale

The paper presents EnScale as a new generative ML architecture consisting of a large-scale mismatch adjustment step followed by sparse local stochastic super-resolution layers, both trained end-to-end with the energy score on paired GCM-RCM data. All load-bearing components (the two-stage map, the novel stochastic layers, the proper scoring rule objective, and the multivariate/temporal consistency claims) are defined directly from the training procedure and evaluated against held-out data and external baselines. No equation or claim reduces by construction to a fitted parameter renamed as a prediction, nor does any central premise rest on a self-citation chain whose content is itself unverified within the paper. The method remains falsifiable via the stated evaluation categories (calibration, spatial/temporal structure, extremes, multivariate dependencies) on data independent of the training pairs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the effectiveness of the introduced two-step generative architecture and the assumption that training data pairs allow learning the full conditional distribution; the main novel element is the sparse local stochastic layer class.

free parameters (1)
  • model hyperparameters and layer parameters
    The generative models are trained by fitting parameters to the paired GCM-RCM data.
axioms (1)
  • standard math The energy score is a proper scoring rule suitable for training generative models to match target conditional distributions.
    Used to optimize both the large-scale adjustment and super-resolution steps.
invented entities (1)
  • sparse local stochastic layers no independent evidence
    purpose: Efficiently model high-dimensional output in the super-resolution step while maintaining spatial consistency.
    Presented as a novel class of layers in the framework.

pith-pipeline@v0.9.0 · 5853 in / 1409 out tokens · 46075 ms · 2026-05-18T11:52:35.783173+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.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. SwAIther-Precip: Lead-Time-Aware Bias Correction Enables Kilometer-Scale Downscaling of Global AI Precipitation Forecasts over Switzerland

    physics.ao-ph 2026-05 unverdicted novelty 6.0

    SwAIther-Precip uses lead-time-conditioned U-Net bias correction followed by diffusion-based super-resolution to downscale AIFS forecasts, achieving 48% CRPS reduction and ~4 km effective resolution up to 5 days lead time.

  2. Downscaling weather forecasts from Low- to High-Resolution with Diffusion Models

    physics.ao-ph 2026-03 unverdicted novelty 5.0

    A conditional diffusion model downscales global atmospheric forecasts from 100 km to 30 km resolution while improving probabilistic skill, matching power spectra, and preserving physical relationships.

  3. Score-based generative emulation of impact-relevant Earth system model outputs

    physics.ao-ph 2025-10 conditional novelty 5.0

    A score-based generative emulator on a spherical mesh produces joint distributions of impact-relevant climate variables that closely match three parent ESMs in pre-industrial and forced regimes, with errors small rela...

Reference graph

Works this paper leans on

63 extracted references · 63 canonical work pages · cited by 3 Pith papers · 4 internal anchors

  1. [1]

    Machine learning emulation of a local-scale UK cli- mate model

    Henry Addison, Laurence Aitchison, and Peter A G Watson. “Machine learning emulation of a local-scale UK cli- mate model”. In:Neurips(2022).URL: https://www.climatechange.ai/papers/neurips2022/ 21

    work page 2022

  2. [2]

    arXiv, ://arxiv.org/abs/2506.10772, arXiv:2506.10772 [cs], doi:10.48550/arXiv.2506.10772

    Ferran Alet et al. “Skillful joint probabilistic weather forecasting from marginals”. In:arXiv preprint(June 2025). URL:http://arxiv.org/abs/2506.10772

    work page arXiv 2025

  3. [3]

    Towards Principled Methods for Training Generative Adversarial Networks

    Martin Arjovsky. “Towards Principled Methods for Training Generative Adversarial Networks”. In: (2017), pp. 1–17

    work page 2017

  4. [4]

    Configuration and intercomparison of deep learning neural models for statistical downscaling

    Jorge Baño-Medina, Rodrigo Manzanas, and Jose Manuel Gutierrez. “Configuration and intercomparison of deep learning neural models for statistical downscaling”. In:Geoscientific Model Development13.4 (Apr. 2020), pp. 2109–2124.ISSN: 19919603.DOI:10.5194/gmd-13-2109-2020

    work page doi:10.5194/gmd-13-2109-2020 2020

  5. [5]

    On the suitability of deep convolutional neural networks for continental-wide downscaling of climate change projections

    Jorge Baño-Medina, Rodrigo Manzanas, and José Manuel Gutiérrez. “On the suitability of deep convolutional neural networks for continental-wide downscaling of climate change projections”. In:Climate Dynamics57.11- 12 (2021), pp. 2941–2951.ISSN: 14320894.DOI: 10 . 1007 / s00382 - 021 - 05847 - 0.URL: https : //doi.org/10.1007/s00382-021-05847-0

    work page doi:10.1007/s00382-021-05847-0 2021

  6. [6]

    Downscaling precipitation extremes

    Rasmus E. Benestad. “Downscaling precipitation extremes”. In:Theoretical and Applied Climatology100.1-2 (Mar. 2010), pp. 1–21.ISSN: 0177-798X.DOI:10.1007/s00704-009-0158-1

    work page doi:10.1007/s00704-009-0158-1 2010

  7. [7]

    A simple hybrid statistical–dynamical downscaling method for emulating regional climate models over Western Europe. Evaluation, application, and role of added value?

    Julien Boé, Alexandre Mass, and Juliette Deman. “A simple hybrid statistical–dynamical downscaling method for emulating regional climate models over Western Europe. Evaluation, application, and role of added value?” In:Climate Dynamics61.1-2 (July 2023), pp. 271–294.ISSN: 14320894.DOI: 10.1007/s00382- 022- 06552-2

    work page doi:10.1007/s00382- 2023

  8. [8]

    Statistical Inference

    George Casella and Roger Berger.Statistical Inference. Boca Raton: Chapman and Hall/CRC, Apr. 2024, p. 536. ISBN: 9781003456285.DOI:10.1201/9781003456285

    work page doi:10.1201/9781003456285 2024

  9. [9]

    Generative machine learning methods for multivariate ensemble post-processing

    Jieyu Chen, Sebastian Lerch, and Tim Janke. “Generative machine learning methods for multivariate ensemble post-processing”. In:EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022Ml (2022).URL: https: //lens.org/189-140-183-158-58X

    work page 2022

  10. [10]

    Insights from Earth system model initial-condition large ensembles and future prospects

    C. Deser et al. “Insights from Earth system model initial-condition large ensembles and future prospects”. In: Nature Climate Change10.4 (Apr. 2020), pp. 277–286.ISSN: 1758-678X.DOI: 10.1038/s41558-020- 0731-2

    work page doi:10.1038/s41558-020- 2020

  11. [11]

    On the suitability of a convolutional neural network based RCM-emulator for fine spatio-temporal precipitation

    Antoine Doury, Samuel Somot, and Sebastien Gadat. “On the suitability of a convolutional neural network based RCM-emulator for fine spatio-temporal precipitation”. In:Climate Dynamics(Sept. 2024).ISSN: 14320894.DOI: 10.1007/s00382-024-07350-8

    work page doi:10.1007/s00382-024-07350-8 2024

  12. [12]

    Regional climate model emulator based on deep learning: concept and first evaluation of a novel hybrid downscaling approach

    Antoine Doury et al. “Regional climate model emulator based on deep learning: concept and first evaluation of a novel hybrid downscaling approach”. In:Climate Dynamics0123456789 (2022).ISSN: 14320894.DOI: 10.1007/s00382-022-06343-9.URL:https://doi.org/10.1007/s00382-022-06343-9

    work page doi:10.1007/s00382-022-06343-9.url:https://doi.org/10.1007/s00382-022-06343-9 2022

  13. [13]

    Eyring, S

    Veronika Eyring et al. “Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6) experimental design and organization”. In:Geoscientific Model Development9.5 (May 2016), pp. 1937–1958.ISSN: 1991-9603. DOI:10.5194/gmd-9-1937-2016

    work page doi:10.5194/gmd-9-1937-2016 2016

  14. [14]

    Thirty Years of Regional Climate Modeling: Where Are We and Where Are We Going next?

    Filippo Giorgi. “Thirty Years of Regional Climate Modeling: Where Are We and Where Are We Going next?” In:Journal of Geophysical Research: Atmospheres124.11 (June 2019), pp. 5696–5723.ISSN: 2169-897X.DOI: 10.1029/2018JD030094

    work page doi:10.1029/2018jd030094 2019

  15. [15]

    Regional Dynamical Downscaling and the CORDEX Initiative

    Filippo Giorgi and William J. Gutowski. “Regional Dynamical Downscaling and the CORDEX Initiative”. In:Annual Review of Environment and Resources40.1 (Nov. 2015), pp. 467–490.ISSN: 1543-5938.DOI: 10.1146/annurev-environ-102014-021217

    work page doi:10.1146/annurev-environ-102014-021217 2015

  16. [16]

    spateGAN: Spatio-Temporal Downscaling of Rainfall Fields Using a cGAN Approach

    Luca Glawion et al. “spateGAN: Spatio-Temporal Downscaling of Rainfall Fields Using a cGAN Approach”. In: Earth and Space Science10.10 (Oct. 2023).ISSN: 2333-5084.DOI:10.1029/2023EA002906

    work page doi:10.1029/2023ea002906 2023

  17. [17]

    Journal of the Royal Statistical Society: Series B (Statistical Methodology) 70(1), 209–226 (2008) http://dx.doi.org/10.1111/j.1467-9868.2007.00633.x

    Tilmann Gneiting, Fadoua Balabdaoui, and Adrian E. Raftery. “Probabilistic Forecasts, Calibration and Sharp- ness”. In:Journal of the Royal Statistical Society Series B: Statistical Methodology69.2 (Apr. 2007), pp. 243–268. ISSN: 1369-7412.DOI:10.1111/j.1467-9868.2007.00587.x

    work page doi:10.1111/j.1467-9868.2007.00587.x 2007

  18. [18]

    Strictly proper scoring rules, prediction, and estimation

    Tilmann Gneiting and Adrian E. Raftery. “Strictly proper scoring rules, prediction, and estimation”. In:Journal of the American Statistical Association102.477 (2007), pp. 359–378.ISSN: 01621459.DOI: 10 . 1198 / 016214506000001437

    work page 2007

  19. [19]

    Weather Forecasting with Ensemble Methods

    Tilmann Gneiting and Adrian E. Raftery. “Weather Forecasting with Ensemble Methods”. In:Science310.5746 (Oct. 2005), pp. 248–249.ISSN: 0036-8075.DOI:10.1126/science.1115255. 42 EnScale: Temporally-consistent multivariate generative downscaling via proper scoring rulesA PREPRINT

    work page doi:10.1126/science.1115255 2005

  20. [20]

    [HBP23] Aamal Abbas Hussain, Francesco Belardinelli, and G eorgios Piliouras

    Ian Goodfellow et al. “Generative adversarial networks”. In:Communications of the ACM63.11 (Oct. 2020), pp. 139–144.ISSN: 0001-0782.DOI:10.1145/3422622

    work page doi:10.1145/3422622 2020

  21. [21]

    Paula Harder et al.Hard-Constrained Deep Learning for Climate Downscaling. Tech. rep. 2023, pp. 1–40

    work page 2023

  22. [22]

    A Generative Deep Learning Approach to Stochastic Downscaling of Precipitation Forecasts

    Lucy Harris et al. “A Generative Deep Learning Approach to Stochastic Downscaling of Precipitation Forecasts”. In:Journal of Advances in Modeling Earth Systems14.10 (Oct. 2022).ISSN: 19422466.DOI: 10.1029/ 2022MS003120

    work page 2022

  23. [23]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , author =

    Alexander Henzi, Johanna F. Ziegel, and Tilmann Gneiting. “Isotonic distributional regression”. In:Journal of the Royal Statistical Society. Series B: Statistical Methodology83.5 (Nov. 2021), pp. 963–993.ISSN: 14679868. DOI:10.1111/rssb.12450

    work page doi:10.1111/rssb.12450 2021

  24. [24]

    On the limitations of deep learning for statistical downscaling of climate change projections: The transferability and the extrapolation issues

    Alfonso Hernanz et al. “On the limitations of deep learning for statistical downscaling of climate change projections: The transferability and the extrapolation issues”. In:Atmospheric Science Letters(2023).ISSN: 1530261X.DOI:10.1002/asl.1195

    work page doi:10.1002/asl.1195 2023

  25. [25]

    Denoising Diffusion Probabilistic Models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. “Denoising Diffusion Probabilistic Models”. In:NeurIPS 2020. Curran Associates Inc., 2020.URL:https://github.com/hojonathanho/diffusion

    work page 2020

  26. [26]

    Forced Component Estimation Statistical Method Intercomparison Project (ForceSMIP)

    Rachael N. Isphording et al. “A Standardized Benchmarking Framework to Assess Downscaled Precipitation Simulations”. In:Journal of Climate37.4 (Feb. 2024), pp. 1089–1110.ISSN: 0894-8755.DOI: 10.1175/JCLI- D-23-0317.1

    work page doi:10.1175/jcli- 2024

  27. [27]

    EURO-CORDEX: New high-resolution climate change projections for European impact research

    Daniela Jacob et al. “EURO-CORDEX: New high-resolution climate change projections for European impact research”. In:Regional Environmental Change14.2 (2014), pp. 563–578.ISSN: 1436378X.DOI: 10.1007/ s10113-013-0499-2

    work page 2014

  28. [28]

    Advancing Parsimonious Deep Learning Weather Prediction Using the HEALPix Mesh

    Matthias Karlbauer et al. “Advancing Parsimonious Deep Learning Weather Prediction Using the HEALPix Mesh”. In:Journal of Advances in Modeling Earth Systems16.8 (Aug. 2024).ISSN: 1942-2466.DOI: 10.1029/ 2023MS004021

    work page 2024

  29. [29]

    Tero Karras et al.Elucidating the Design Space of Diffusion-Based Generative Models. Tech. rep.URL: https: //github.com/NVlabs/edm

    work page

  30. [30]

    Brenner, and Stephan Hoyer

    Dmitrii Kochkov et al. “Neural general circulation models for weather and climate”. In:Nature632.8027 (Aug. 2024), pp. 1060–1066.ISSN: 0028-0836.DOI:10.1038/s41586-024-07744-y

    work page doi:10.1038/s41586-024-07744-y 2024

  31. [31]

    doi: 10.1126/science.adi2336

    Remi Lam et al. “Learning skillful medium-range global weather forecasting”. In:Science382.6677 (Dec. 2023), pp. 1416–1421.ISSN: 0036-8075.DOI:10.1126/science.adi2336

    work page doi:10.1126/science.adi2336 2023

  32. [32]

    arXiv, ://arxiv.org/abs/2412.15832, arXiv:2412.15832 [physics] version: 1, doi:10.48550/arXiv.2412.15832

    Simon Lang et al. “AIFS-CRPS: Ensemble forecasting using a model trained with a loss function based on the Continuous Ranked Probability Score”. In:arXiv preprint(Dec. 2024).URL: http://arxiv.org/abs/ 2412.15832

    work page arXiv 2024

  33. [33]

    Stochastic Super-Resolution for Downscaling Time-Evolving Atmospheric Fields With a Generative Adversarial Network

    Jussi Leinonen, Daniele Nerini, and Alexis Berne. “Stochastic Super-Resolution for Downscaling Time-Evolving Atmospheric Fields With a Generative Adversarial Network”. In:IEEE Transactions on Geoscience and Remote Sensing59.9 (2020), pp. 7211–7223.ISSN: 0196-2892.DOI:10.1109/tgrs.2020.3032790

    work page doi:10.1109/tgrs.2020.3032790 2020

  34. [34]

    Precipitation downscaling under climate change: Recent developments to bridge the gap between dynamical models and the end user

    D. Maraun et al. “Precipitation downscaling under climate change: Recent developments to bridge the gap between dynamical models and the end user”. In:Reviews of Geophysics48.3 (Sept. 2010).ISSN: 87551209. DOI:10.1029/2009RG000314

    work page doi:10.1029/2009rg000314 2010

  35. [35]

    Bias Correcting Climate Change Simulations - a Critical Review

    Douglas Maraun. “Bias Correcting Climate Change Simulations - a Critical Review”. In:Current Climate Change Reports2.4 (Dec. 2016), pp. 211–220.ISSN: 2198-6061.DOI:10.1007/s40641-016-0050-x

    work page doi:10.1007/s40641-016-0050-x 2016

  36. [36]

    Cambridge University Press, Jan

    Douglas Maraun and Martin Widmann.Statistical Downscaling and Bias Correction for Climate Research. Cambridge University Press, Jan. 2018.ISBN: 9781107066052.DOI:10.1017/9781107588783

    work page doi:10.1017/9781107588783 2018

  37. [37]

    V ALUE : A framework to validate downscaling approaches for climate change studies

    Douglas Maraun et al. “V ALUE : A framework to validate downscaling approaches for climate change studies”. In:Earth’s Future3.1 (Jan. 2015), pp. 1–14.ISSN: 2328-4277.DOI:10.1002/2014EF000259

    work page doi:10.1002/2014ef000259 2015

  38. [38]

    Residual corrective diffusion modeling for km-scale atmospheric downscaling

    Morteza Mardani et al. “Residual corrective diffusion modeling for km-scale atmospheric downscaling”. In: Communications Earth & Environment6.1 (Feb. 2025), p. 124.ISSN: 2662-4435.DOI: 10.1038/s43247- 025-02042-5

    work page doi:10.1038/s43247- 2025

  39. [39]

    Deep Learning Regional Climate Model Emulators: A Comparison of Two Downscaling Training Frameworks

    Marijn van der Meer, Sophie de Roda Husman, and Stef Lhermitte. “Deep Learning Regional Climate Model Emulators: A Comparison of Two Downscaling Training Frameworks”. In:Journal of Advances in Modeling Earth Systems15.6 (June 2023).ISSN: 1942-2466.DOI:10.1029/2022ms003593

    work page doi:10.1029/2022ms003593 2023

  40. [40]

    Downscaling of Historical Wind Fields over Switzerland using Generative Adversarial Networks

    Ophélia Miralles et al. “Downscaling of Historical Wind Fields over Switzerland using Generative Adversarial Networks”. In:Artificial Intelligence for the Earth Systems(2022), pp. 1–44.DOI: 10.1175/aies-d-22- 0018.1

    work page doi:10.1175/aies-d-22- 2022

  41. [41]

    Probabilistic Forecasting with Generative Networks via Scoring Rule Minimization

    Lorenzo Pacchiardi et al. “Probabilistic Forecasting with Generative Networks via Scoring Rule Minimization”. In:Journal of Machine Learning Research25 (2024), pp. 1–64. 43 EnScale: Temporally-consistent multivariate generative downscaling via proper scoring rulesA PREPRINT

    work page 2024

  42. [42]

    P ; Pinson and J Tastu.Discrimination ability of the Energy score. Tech. rep. 2013

    work page 2013

  43. [43]

    Andersson, Andrew El-Kadi, Do- minic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, Remi Lam, and Matthew Willson

    Ilan Price et al. “Probabilistic weather forecasting with machine learning”. In:Nature637.8044 (Jan. 2025), pp. 84–90.ISSN: 0028-0836.DOI:10.1038/s41586-024-08252-9

    work page doi:10.1038/s41586-024-08252-9 2025

  44. [44]

    Ilan Price Stephan Rasp and The Alan Turing Institute ClimateAI.Increasing the accuracy and resolution of precipitation forecasts using deep generative models. Tech. rep. 2022, p. 2022

    work page 2022

  45. [45]

    Downscaling with AI reveals the large role of internal variability in fine-scale projections of climate extremes

    Neelesh Rampal et al. “Downscaling with AI reveals the large role of internal variability in fine-scale projections of climate extremes”. In:arXiv preprint(July 2025).URL:http://arxiv.org/abs/2507.06527

    work page arXiv 2025

  46. [46]

    Enhancing Regional Climate Downscaling through Advances in Machine Learning

    Neelesh Rampal et al. “Enhancing Regional Climate Downscaling through Advances in Machine Learning”. In: Artificial Intelligence for the Earth Systems3.2 (Apr. 2024).ISSN: 2769-7525.DOI: 10.1175/AIES-D-23- 0066.1

    work page doi:10.1175/aies-d-23- 2024

  47. [47]

    Energy distance

    Maria L. Rizzo and Gábor J. Székely. “Energy distance”. In:Wiley Interdisciplinary Reviews: Computational Statistics8.1 (Jan. 2016), pp. 27–38.ISSN: 19390068.DOI:10.1002/wics.1375

    work page doi:10.1002/wics.1375 2016

  48. [48]

    Improved Techniques for Training GANs

    Tim Salimans et al. “Improved Techniques for Training GANs”. In:arXiv preprint(June 2016).URL: http: //arxiv.org/abs/1606.03498

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  49. [49]

    Spatio-Temporal Downscaling of Climate Data Using Convolu- tional and Error-Predicting Neural Networks

    Agon Serifi, Tobias Günther, and Nikolina Ban. “Spatio-Temporal Downscaling of Climate Data Using Convolu- tional and Error-Predicting Neural Networks”. In:Frontiers in Climate3.April (2021), pp. 1–15.ISSN: 26249553. DOI:10.3389/fclim.2021.656479

    work page doi:10.3389/fclim.2021.656479 2021

  50. [50]

    Engression: extrapolation through the lens of distributional regression

    Xinwei Shen and Nicolai Meinshausen. “Engression: extrapolation through the lens of distributional regression”. In:Journal of the Royal Statistical Society Series B: Statistical Methodology(Nov. 2024).ISSN: 1369-7412. DOI: 10 . 1093 / jrsssb / qkae108.URL: https : / / academic . oup . com / jrsssb / advance - article/doi/10.1093/jrsssb/qkae108/7909013

    work page doi:10.1093/jrsssb/qkae108/7909013 2024

  51. [51]

    Reverse Markov Learning: Multi-Step Generative Models for Complex Distributions

    Xinwei Shen, Nicolai Meinshausen, and Tong Zhang. “Reverse Markov Learning: Multi-Step Generative Models for Complex Distributions”. In:arXiv preprint(Feb. 2025).URL: http://arxiv.org/abs/2502.13747

    work page arXiv 2025

  52. [52]

    Generative Modeling by Estimating Gradients of the Data Distribution

    Yang Song and Stefano Ermon. “Generative Modeling by Estimating Gradients of the Data Distribution”. In: arXiv preprint(July 2019).URL:http://arxiv.org/abs/1907.05600

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  53. [53]

    Score-Based Generative Modeling through Stochastic Differential Equations

    Yang Song et al. “Score-Based Generative Modeling through Stochastic Differential Equations”. In:arXiv preprint(Nov. 2020).URL:http://arxiv.org/abs/2011.13456

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  54. [54]

    Adversarial super-resolution of climatological wind and solar data

    Karen Stengel et al. “Adversarial super-resolution of climatological wind and solar data”. In:Proceedings of the National Academy of Sciences of the United States of America117.29 (2020), pp. 16805–16815.ISSN: 10916490. DOI:10.1073/pnas.1918964117

    work page doi:10.1073/pnas.1918964117 2020

  55. [55]

    Gabor J Szekely, Gábor J Székely, and Maria L Rizzo.E-Statistics: Energy of statistical samples Energy Statistics: A Class of Statistics Based on Distances. Tech. rep.URL: https://www.researchgate.net/ publication/243786506

    work page arXiv

  56. [56]

    Taylor, R.J

    Karl E. Taylor, Ronald J. Stouffer, and Gerald A. Meehl. “An Overview of CMIP5 and the Experiment Design”. In:Bulletin of the American Meteorological Society93.4 (Apr. 2012), pp. 485–498.ISSN: 1520-0477.DOI: 10.1175/BAMS-D-11-00094.1

    work page doi:10.1175/bams-d-11-00094.1 2012

  57. [57]

    Statistical Postprocessing for Weather Forecasts: Review, Challenges, and Avenues in a Big Data World

    Stéphane Vannitsem et al. “Statistical Postprocessing for Weather Forecasts: Review, Challenges, and Avenues in a Big Data World”. In:Bulletin of the American Meteorological Society102.3 (Mar. 2021), E681–E699.ISSN: 0003-0007.DOI:10.1175/BAMS-D-19-0308.1

    work page doi:10.1175/bams-d-19-0308.1 2021

  58. [58]

    Deep Learning for Downscaling Tropical Cyclone Rainfall to Hazard-Relevant Spatial Scales

    Emily V osper et al. “Deep Learning for Downscaling Tropical Cyclone Rainfall to Hazard-Relevant Spatial Scales”. In:Journal of Geophysical Research: Atmospheres128.10 (May 2023).ISSN: 21698996.DOI: 10. 1029/2022JD038163

    work page 2023

  59. [59]

    Easy Uncertainty Quantification (EasyUQ): Generating Predictive Distributions from Single-Valued Model Output

    Eva-Maria Walz et al. “Easy Uncertainty Quantification (EasyUQ): Generating Predictive Distributions from Single-Valued Model Output”. In:SIAM Review66.1 (Feb. 2024), pp. 91–122.ISSN: 0036-1445.DOI: 10. 1137/22M1541915

    work page 2024

  60. [60]

    Regional climate risk assessment from climate models using probabilistic machine learning

    Zhong Yi Wan et al. “Statistical Downscaling via High-Dimensional Distribution Matching with Generative Models”. In:arXiv preprint(Dec. 2024).URL:http://arxiv.org/abs/2412.08079

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  61. [61]

    Downscaling Numerical Weather Models with GANs

    B. L. White, A. Singh, and A. Albert. “Downscaling Numerical Weather Models with GANs”. In:American Geophysical Union, Fall Meeting 2019(2019), pp. 1–4

    work page 2019

  62. [62]

    Downscaling general circulation model output: a review of methods and limitations

    R.L. Wilby and T.M.L. Wigley. “Downscaling general circulation model output: a review of methods and limitations”. In:Progress in Physical Geography: Earth and Environment21.4 (Dec. 1997), pp. 530–548.ISSN: 0309-1333.DOI:10.1177/030913339702100403

    work page doi:10.1177/030913339702100403 1997

  63. [63]

    Enforcing calibration in ensemble postprocessing

    Daniel S. Wilks. “Enforcing calibration in ensemble postprocessing”. In:Quarterly Journal of the Royal Meteorological Society144.710 (Jan. 2018), pp. 76–84.ISSN: 0035-9009.DOI:10.1002/qj.3185. 44

    work page doi:10.1002/qj.3185 2018