Pith. sign in

REVIEW 2 major objections 5 minor 70 references

This paper claims that one Bayesian transport map can learn the non-Gaussian joint distribution of a spatial field across multiple resolutions from a handful of training sample pairs, and then yield the conditional distribution of fine-scal

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A new multi-fidelity Bayesian transport map method learns non-Gaussian joint distributions across spatial scales and outperforms existing emulators in downscaling climate fields from small training sets.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A clean, useful extension of BTM to multi-fidelity spatial data with a closed-form likelihood and strong empirical results; the main gap is that the Markov assumption across fidelities is never tested where it could fail. the 2 major comments →

arxiv 2509.22474 v3 pith:2R3MITPR submitted 2025-09-26 stat.ME

Generative multi-scale modeling and downscaling via spatial autoregressive transport maps

Alejandro Calle-Saldarriaga , Paul F.V. Wiemann , Matthias Katzfuss This is my paper
classification stat.ME MSC 62M3062F15
keywords multi-fidelity modelingstatistical downscalingBayesian transport mapGaussian processnon-Gaussian spatial fieldsclimate model emulationscreening effectautoregressive GP
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper introduces a generative statistical emulator for multi-fidelity spatial data, such as coarse global climate model output paired with fine regional climate model output. The method, called MF BTM, learns the joint distribution across resolutions from as few as 5 to 40 training sample pairs, capturing non-Gaussianity and nonlinear cross-fidelity dependence, and then provides closed-form conditional distributions of fine fields given coarse fields. The central device is a triangular transport map whose components are fidelity-aware autoregressive Gaussian processes, ordered so coarse locations come first and each fine location conditions on nearby coarse and same-fidelity neighbors. On simulated block-averaging and block-minima scenarios and on real GCM-to-RCM temperature fields over Europe, the paper reports the best held-out log-scores among hierarchical kriging, a nonlinear autoregressive GP, a Matérn-covariance model, and a variational autoencoder. If correct, this makes statistical downscaling and multi-fidelity emulation practical in data-scarce, high-dimensional settings.

Core claim

The paper claims that a single probabilistic model can learn the full non-Gaussian, nonstationary joint distribution of a spatial field observed at multiple resolutions from a small training ensemble, and that the conditional distribution of any higher-fidelity level given the immediately coarser level is available in closed form after training. The claim is carried by a lower-triangular transport map that maps the multi-fidelity field to independent standard normals; each map component is a Gaussian process autoregression with variance priors that decay polynomially with distance to previously ordered points, and kernels whose relevance weights decay exponentially with neighbor order, explo

What carries the argument

Fidelity-aware autoregressive transport map: a lower-triangular map T from the multi-fidelity field y to iid standard normals, with component T_{r,i}(y) = (y_{r,i} − f_{r,i}(neighbors)) / d_{r,i}. The conditioning sets use a conditional maximin ordering: coarser-fidelity locations are ordered first, and each fine location conditions on its nearest previously ordered same-fidelity neighbors plus nearest neighbors in the immediately coarser fidelity. Priors are inverse-gamma on the conditional variances with means decaying polynomially in the distance to the nearest previously ordered point, and the kernel combines a linear-in-neighbors term with a squared-exponential nonlinearity whose range

Load-bearing premise

The load-bearing assumption is that each fidelity level depends on the data only through the immediately coarser level, p(y_r | y_<r) = p(y_r | y_{r-1}), so any dependence on other coarse scales or on longer-range history must be mediated through that single level.

What would settle it

Train the model on a synthetic ensemble in which the fine field is generated as a nonlinear function of both the medium and the coarse fields, so the Markov property is false, and compare held-out log-scores against a variant that conditions on both levels; a significant drop for the Markov-restricted model would show where the assumption binds. Alternatively, apply the method to a real multi-fidelity dataset with three or more levels and test whether residuals from the fitted conditionals remain independent of the skipped coarse level.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Statistical emulation of regional climate models from global model output becomes feasible with roughly 10 to 40 training pairs, producing probabilistic fine-scale fields in closed form rather than requiring thousands of deep-learning training samples.
  • The closed-form conditional distribution allows cheap generation of many synthetic high-fidelity fields from new coarse inputs, enabling large-ensemble climate studies and uncertainty propagation at a fraction of the dynamical downscaling cost.
  • Because training factorizes across fidelities and uses sparse conditioning sets, the approach scales to very large fine grids (e.g., 78,400 locations) with modest per-fidelity computational cost.
  • The model captures nonlinear cross-fidelity relationships, as demonstrated by the block-minima simulation, so it can represent aggregation rules that are not simple linear averages.
  • The probabilistic formulation gives a strictly proper scoring rule (log-score) for model comparison, making it possible to quantify gains over simpler kriging-based and deep-learning downscaling methods.
  • The ability to sample conditionally from the fitted distribution supports downstream tasks such as assessing extreme-event probabilities and calibrating regional model parameters.
  • Because the method is Bayesian, it could be extended to incorporate hyperparameter uncertainty via MCMC or Laplace approximation when uncertainty quantification is critical.

Where Pith is reading between the lines

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

  • The Markov assumption that each fidelity depends only on the immediately coarser level is stated as 'for notational simplicity'; relaxing it to condition on multiple coarse levels could extend the method to settings where fine-scale processes are influenced by planetary-scale and regional-scale features simultaneously, which the authors note is straightforward.
  • The success on block-minima suggests the transport map implicitly learns the physical aggregation operation; this could be exploited for detecting or validating downscaling relationships in climate model outputs, or for bias-correcting coarse fields by sampling from the learned conditional.
  • Applying the method to precipitation, which has heavy tails and many zeros, would be a stiffer test because the conditional Gaussian components may struggle; the authors mention flexible-map extensions that could address this, implying a roadmap for non-Gaussian conditional structure.
  • The closed-form posterior could enable rare-event simulation by importance sampling on low-fidelity fields, potentially making it easier to study extreme heat or heavy rainfall without running expensive regional models.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper extends the Bayesian transport map (BTM) framework of Katzfuss and Schäfer (2023) to multi-fidelity spatial fields. The joint distribution over R fields is factored autoregressively, and a Markov assumption is imposed so that each higher-fidelity field depends only on the immediately lower-fidelity field. A conditional maximin ordering is used to define small conditioning sets (same-fidelity nearest neighbors plus nearest neighbors from the previous fidelity), and fidelity-specific GP priors with inverse-gamma variances are specified. Conjugacy yields a closed-form integrated likelihood (Eq. 4), enabling empirical-Bayes hyperparameter estimation by stochastic gradient descent. The fitted transport map provides closed-form conditional distributions of fine fields given coarse fields. The method is compared with hierarchical kriging, NARGP, a Matérn model, and a VAE on two simulation studies (block averages, block minima) and on GCM-RCM climate model output, using average log-scores. MF BTM reports the best average log-scores in all settings.

Significance. If the results hold, the paper provides a practical and scalable non-Gaussian multi-fidelity emulator that is specifically designed for the small-training-ensemble regime. The strongest points are the closed-form integrated likelihood (Eq. 4), the scalable conditional maximin ordering, and the availability of code and data links. The method is genuinely generative and can sample high-fidelity fields conditional on coarse inputs. The empirical comparison is broad and includes reasonable baselines. However, the contribution is incremental relative to the single-fidelity BTM, and the empirical evidence for the general multi-fidelity claim is narrower than the abstract suggests.

major comments (2)
  1. [Section 2.2] The Markov assumption p(y_r|y_<r)=p(y_r|y_{r-1}) is load-bearing for the claim that the method learns the joint multi-fidelity distribution. The paper says relaxing it is 'straightforward,' but no derivation or experiment is given. Moreover, all numerical experiments are settings where the assumption is either vacuous or automatically satisfied: the climate experiment (Section 4) has R=2, so p(y_2|y_1) is the full conditional; the R=3 simulations (Sections 3.2 and 3.3) use deterministic coarsening (block averages and block minima), making y_1 a deterministic function of y_2 and hence p(y_3|y_1,y_2)=p(y_3|y_2) by construction. Thus there is no test of the Markov assumption in a regime where it can fail. If a real multi-fidelity process has dependence on multiple coarse scales, the conditioning sets used in Eq. (2) and (3) are misspecified. Please add a simulation with R>=3 in which the co
  2. [Figures 5 and 8] The empirical comparisons report only point estimates of average log-scores. Figure 5 uses 50 test fields and Figure 8 uses 10 held-out test samples, but no standard errors, confidence intervals, or repeated training/test splits are provided. In the small-n regime (n=5-10), the differences between methods could be within Monte Carlo noise; the abstract's claim that the approach 'substantially outperforms existing methods' is therefore not quantitatively supported. Please add error bars (e.g., bootstrap or repeated random splits) and, if feasible, assess whether the reported ranking is stable across independent training sets.
minor comments (5)
  1. [Section 4 and Figure 7] The RCM acronym is inconsistent: the text uses CRMC5 and CRCM5, and Figure 7 uses CRM5. Please standardize.
  2. [Equation (4)] The notation K_{r,i} is used both for the kernel function in (3) and for the n x n Gram matrix in (4). Use a bold symbol or an explicit definition for the matrix to avoid confusion.
  3. [Figure 3] The caption is long and the four rows are not labeled in the figure. Consider labeling panels (a)-(d) so the reader can connect the rows to the described quantities.
  4. [Section 3.2] The statement that the Matérn model is 'the second best for small n' appears to refer only to the left panel of Figure 5; the right panel does not include Matérn. Please make this explicit.
  5. [Appendix A] The VAE comparison is based on a bespoke architecture; the description gives general structure but no layer sizes, number of parameters, or detailed training schedule. Please provide enough detail for exact reproducibility.

Circularity Check

0 steps flagged

No circular reduction found; the core derivation is a self-contained hierarchical GP transport model validated on held-out data.

full rationale

The paper's central claim—that MF BTM substantially outperforms existing methods—is supported by held-out log-score evaluations on simulations and on CanESM2/CRCM5 climate data, not by a fitted parameter relabeled as a prediction. The model's joint density is the autoregressive product p(y)=∏_r p(y_r|y_<r); the Markov restriction in Section 2.2 is an explicit modeling assumption, not a definitional equivalence to the target result. Hyperparameters θ are estimated by empirical Bayes from training data and then evaluated on held-out test fields; no test quantity is an input to the fit. The paper cites prior work by the same group (Katzfuss & Schäfer 2023; Schäfer et al. 2021; Chen et al. 2025) for the BTM framework, screening-effect rates, and conditional-maximin ordering; these are published mathematical/algorithmic results with stated assumptions and are not equivalent to the present claim of empirical outperformance. The simulation studies draw from a transport-map generator related to the model class, but this is a model-match sensitivity check, not a circular reduction, and the nonlinear block-minima and real-data experiments provide independent evidence. The paper does assert 'relaxing [the Markov assumption] is straightforward' (Section 2.2) and that temporal extension is 'straightforward' (Sections 4 and 5); these are unverified scope claims and should be weighed as robustness risks, but they are not circularity. No self-definitional, fitted-input-as-prediction, uniqueness-imported, ansatz-smuggled, or renaming pattern is present.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central method rests on a small set of explicit modeling choices: a Markov structure between fidelities, conditional Gaussian map components, exponential screening of neighbors, and a polynomial variance decay. The hyperparameters that control these decays are fitted to the training data. No new physical entities are introduced.

free parameters (6)
  • θ^d1_r, θ^d2_r (variance decay parameters per fidelity) = estimated via empirical Bayes
    Control the prior mean of residual variance E[d^2_r,i] = exp(θ^d1_r) ℓ^{θ^d2_r}_{r,i}; fitted per fidelity, shared within fidelity.
  • θ^γ_r (kernel range per fidelity) = estimated via empirical Bayes
    Range parameter in covariance kernel (3), fidelity-specific.
  • θ^{σ1}_r, θ^{σ2}_r (nonlinearity variance decay per fidelity) = estimated via empirical Bayes
    Control σ^2_{r,i} = exp(θ^{σ1}_r) ℓ^{θ^{σ2}_r}_{r,i}; pushes functions toward linearity at fine scales.
  • θ^{q0}_r, θ^{q1}_r, θ^{q'0}_r, θ^{q'1}_r (relevance weights per fidelity) = estimated via empirical Bayes
    Define exponential decay of relevance matrix Q_r in kernel (3); these determine conditioning set sizes.
  • g (prior coefficient of variation factor) = 4
    Fixed to 4 to get a vague inverse-gamma prior; not fitted but chosen by hand.
  • ε (relevance threshold) = 0.01
    Determines adaptive conditioning set sizes; fixed upfront.
axioms (5)
  • domain assumption Markov assumption p(y_r|y_<r) = p(y_r|y_{r-1}) between fidelities
    Stated in Section 2.2; the conditional model uses only the immediately lower fidelity. If real processes have longer cross-fidelity memory, the conditional is misspecified.
  • domain assumption Conditional Gaussianity of each transport-map component given the conditioning set: y_{r,i} | y~c_{r,i}, f_{r,i}, d^2_{r,i} ~ N(...)
    Core modeling choice; the non-Gaussian joint distribution arises only through the composition of these Gaussian conditionals. Assumes the map component is monotone and Gaussian in the last argument.
  • domain assumption Screening effect: relevance decays exponentially with neighbor order, so truncating conditioning sets to m_r + m'_r nearest neighbors suffices
    Used to build sparse kernels and adaptive conditioning sets. Justified by citations to Schäfer et al. (2021) and Chen et al. (2025) for certain GP/PDE settings, but assumed for general processes.
  • ad hoc to paper Polynomial decay of conditional variance: E[d^2_{r,i}] = exp(θ^d1_r) ℓ^{θ^d2_r}_{r,i}
    Parametric form motivated by theory for quasi-quadratic log-likelihood processes; not derived for the specific map components.
  • standard math Normal-inverse-gamma conjugacy yields closed-form integrated likelihood
    Used to derive Eq. (4); standard result.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative multi-scale modeling and downscaling via spatial autoregressive transport maps." pith.science (2026). https://pith.science/paper/2R3MITPR

@misc{pith2026250922474,
  author       = {Pith},
  title        = {Pith review of: Generative multi-scale modeling and downscaling via spatial autoregressive transport maps},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2R3MITPR}},
  note         = {Machine review of arXiv:2509.22474}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Spatial fields in the Earth and environmental sciences are often available at multiple scales or resolutions. While coarse-scale data (e.g., from global circulation models) are often abundant, they lack the local detail provided by fine-scale data (e.g., from regional climate models), which are typically computationally expensive to generate. Statistical downscaling and multi-scale data fusion address this challenge by predicting high-resolution fields from low-resolution or related inputs. We propose a highly scalable Bayesian approach that can learn the joint non-Gaussian distribution and nonlinear dependence structure of nonstationary spatial fields across multiple scales from a small number of training samples. Our method employs scale-aware autoregressive Gaussian processes with suitably chosen regularization-inducing priors to model the conditional distribution of fine-scale fields given coarse-scale data. Exploiting conjugacy, the integrated likelihood is available in closed form, enabling efficient parameter optimization via stochastic gradient descent. Once trained, the method provides a closed-form characterization of the posterior distribution of fine-scale fields given coarse-scale inputs. In numerical comparisons, we demonstrate that our approach substantially outperforms existing methods and effectively characterizes and simulates fine-scale climate behavior based on output from coarse global circulation models.

Figures

Figures reproduced from arXiv: 2509.22474 by Alejandro Calle-Saldarriaga, Matthias Katzfuss, Paul F.V. Wiemann.

Figure 1
Figure 1. Figure 1: Our goal is to learn the non-Gaussian joint and conditional distributions of spatial fields at multiple [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The goal is to ensure that lower-resolution points anchor the ordering, so that [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: For a GP with exponential covariance on coarser-to-finer grids and with [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: For downscaling of (linear) block averages, samples from our BTM mimick those from the true data-generating process, and they get closer to test samples when conditioning on more lower-fidelity test fields. The first two rows show test samples from the block-averaging data￾generating process described in Section 3.2. The next three rows are samples from our model (trained on 50 samples). Arrows represent c… view at source ↗
Figure 5
Figure 5. Figure 5: Our multi-fidelity Bayesian transport map (MF BTM) performs best in terms of log-scores for all sample sizes n in two simulation scenarios. (Left): Block-averaging scenario from Section 3.2. (Right): (Nonlinear) block minima from Section 3.3. The competing methods are listed in Section 3.1. 3.3 Downscaling of (nonlinear) block minima Our model can learn non-linear relationships between fidelities, via the … view at source ↗
Figure 6
Figure 6. Figure 6: For downscaling of (nonlinear) block minima, samples from our BTM also mimick those from the true data-generating process. The first two rows show test samples from the block￾minima data-generating process described in Section 3.3. The next three rows are samples from our model (trained on 50 samples). Arrows represent conditioning relationships. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Our MF BTM approach can capture the joint and conditional spatial structure encoded by GCM-RCM climate models (Section 4). First row: Ensemble member of the maximum temperature anomalies for the CanESM2 (N1 = 336) and CRM5 (N2 = 78,400) pairings over Europe. Our MF BTM infers the N1 + N2 dimensional distribution of coarse and fine scaled features, including the N2- dimensional conditional (on N1 values) di… view at source ↗
Figure 8
Figure 8. Figure 8: The multi-fidelity BTM outperformed all other methods for all ensemble sizes on our GCM-RCM setup. NARGP and VAE had very similar performance, and HK struggled with low ensemble size. than in the settings considered here. For future work, it is straightforward to combine the proposed multi-fidelity approach with existing spatio-temporal (Lei and Katzfuss, in prep.) and multiple-process (Wiemann and Katzfus… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

70 extracted references · 10 linked inside Pith

  1. [1]

    Adams, R. A. and Fournier, J. J. (2003). Sobolev spaces , volume 140. Elsevier

    2003

  2. [2]

    B., and Marzouk, Y

    Baptista, R., Hosseini, B., Kovachki, N. B., and Marzouk, Y. (2020). Conditional sampling with monotone gans: from generative models to likelihood-free inference. arXiv preprint arXiv:2006.06755

    Pith/arXiv arXiv 2020

  3. [3]

    Baptista, R., Marzouk, Y., and Zahm, O. (2024). On the representation and learning of monotone triangular transport maps. Foundations of Computational Mathematics , 24(6):2063--2108

    2024

  4. [4]

    A., Peng, L., Pritchard, M

    Bhouri, M. A., Peng, L., Pritchard, M. S., and Gentine, P. (2023). Multi-fidelity climate model parameterization for better generalization and extrapolation. arXiv preprint arXiv:2309.10231

    Pith/arXiv arXiv 2023

  5. [5]

    Bird, S., Fernandez, M., Ho, M.-F., Qezlou, M., Monadi, R., Ni, Y., Chen, N., Croft, R., and Di Matteo, T. (2023). Priya: a new suite of lyman- forest simulations for cosmology. Journal of Cosmology and Astroparticle Physics , 2023(10):037

    2023

  6. [6]

    M., Kucukelbir, A., and McAuliffe, J

    Blei, D. M., Kucukelbir, A., and McAuliffe, J. D. (2017). Variational Inference: A Review for Statisticians . Journal of the American Statistical Association , 112(518):859--877

    2017

  7. [7]

    Bocquet, M. (2023). Surrogate modeling for the climate sciences dynamics with machine learning and data assimilation. Frontiers in Applied Mathematics and Statistics , 9:1133226

    2023

  8. [8]

    Burda, Y., Grosse, R., and Salakhutdinov, R. (2015). Importance weighted autoencoders. arXiv preprint arXiv:1509.00519

    Pith/arXiv arXiv 2015

  9. [9]

    N., Glaws, A., and King, R

    Buster, G., Benton, B. N., Glaws, A., and King, R. N. (2024). High-resolution meteorology with climate change impacts from global climate model data using generative machine learning. Nature Energy , pages 1--13

    2024

  10. [10]

    and Katzfuss, M

    Chakraborty, A. and Katzfuss, M. (2024). Learning non-gaussian spatial distributions via bayesian transport maps with parametric shrinkage. arXiv preprint arXiv:2409.19208

    Pith/arXiv arXiv 2024

  11. [11]

    and Hassanzadeh, P

    Chattopadhyay, A. and Hassanzadeh, P. (2023). Long-term instabilities of deep learning-based digital twins of the climate system: The cause and a solution. arXiv:2304.07029

    Pith/arXiv arXiv 2023

  12. [12]

    and Sanz-Alonso, D

    Chen, J. and Sanz-Alonso, D. (2024). Precision and cholesky factor estimation for gaussian processes. arXiv preprint arXiv:2412.08820

    Pith/arXiv arXiv 2024

  13. [13]

    Chen, Y., Owhadi, H., and Sch \"a fer, F. (2025). Sparse cholesky factorization for solving nonlinear pdes via gaussian processes. Mathematics of Computation , 94(353):1235--1280

    2025

  14. [14]

    A., De, S., Becker, S., and Doostan, A

    Cheng, N., Malik, O. A., De, S., Becker, S., and Doostan, A. (2024). Bi-fidelity variational auto-encoder for uncertainty quantification. Computer Methods in Applied Mechanics and Engineering , 421:116793

    2024

  15. [15]

    A., Karagiannis, G., and Kang, E

    Cheng, S., Konomi, B. A., Karagiannis, G., and Kang, E. L. (2023). Recursive nearest neighbor co-kriging models for big multi-fidelity spatial data sets. Environmetrics , page e2844

    2023

  16. [16]

    Chylek, P., Li, J., Dubey, M., Wang, M., and Lesins, G. (2011). Observed and model simulated 20th century arctic temperature variability: Canadian earth system model canesm2. Atmospheric Chemistry and Physics Discussions , 11(8):22893--22907

    2011

  17. [17]

    A., Glazer, R., Giuliani, G., Di Sante, F., Nogherotto, R., and Giorgi, F

    Coppola, E., Stocchi, P., Pichelli, E., Torres Alavez, J. A., Glazer, R., Giuliani, G., Di Sante, F., Nogherotto, R., and Giorgi, F. (2021). Non-hydrostatic regcm4 (regcm4-nh): model description and case studies over multiple domains. Geoscientific Model Development Discussions , 2021:1--25

    2021

  18. [18]

    L., Braverman, A., and Kalmus, P

    Ekanayaka, A., Kang, E. L., Braverman, A., and Kalmus, P. (2025). A multivariate spatial statistical model for statistical downscaling of sea surface temperature in the great barrier reef region. Journal of the Royal Statistical Society Series C: Applied Statistics , page qlaf019

    2025

  19. [19]

    Fern \'a ndez-Godino, M. G. (2023). Review of multi-fidelity models. Advances in Computational Science and Engineering , 1(4):351--400

    2023

  20. [20]

    and Bates, G

    Giorgi, F. and Bates, G. T. (1989). The climatological skill of a regional model over complex terrain. Monthly Weather Review , 117(11):2325--2347

    1989

  21. [21]

    M., Pichelli, E., Nogherotto, R., Raffaele, F., Malguzzi, P., Davolio, S., Stocchi, P., et al

    Giorgi, F., Coppola, E., Giuliani, G., Ciarlo, J. M., Pichelli, E., Nogherotto, R., Raffaele, F., Malguzzi, P., Davolio, S., Stocchi, P., et al. (2023). The fifth generation regional climate modeling system, regcm5: Description and illustrative examples at parameterized convection and convection-permitting resolutions. Journal of Geophysical Research: Atm...

    2023

  22. [22]

    and Katzfuss, M

    Gneiting, T. and Katzfuss, M. (2014). Probabilistic forecasting . Annual Review of Statistics and Its Application , 1(1):125--151

    2014

  23. [23]

    Guinness, J. (2018). Permutation and grouping methods for sharpening Gaussian process approximations . Technometrics , 60(4):415--429

    2018

  24. [24]

    and G \"o rtz, S

    Han, Z.-H. and G \"o rtz, S. (2012). Hierarchical kriging model for variable-fidelity surrogate modeling. AIAA journal , 50(9):1885--1896

    2012

  25. [25]

    Hao, S., Han, W., Jiang, T., Li, Y., Wu, H., Zhong, C., Zhou, Z., and Tang, H. (2024). Synthetic data in AI : Challenges, applications, and ethical implications. arXiv:2401.01629

    Pith/arXiv arXiv 2024

  26. [26]

    Hensman, J., Fusi, N., and Lawrence, N. D. (2013). Gaussian processes for big data. arXiv preprint arXiv:1309.6835

    Pith/arXiv arXiv 2013

  27. [27]

    Higgins, I., Matthey, L., Pal, A., Burgess, C., Glorot, X., Botvinick, M., Mohamed, S., and Lerchner, A. (2017). beta-vae: Learning basic visual concepts with a constrained variational framework. In International conference on learning representations

    2017

  28. [28]

    Ho, M.-F., Bird, S., and Shelton, C. R. (2021). Multifidelity emulation for the matter power spectrum using Gaussian processes . Monthly Notices of the Royal Astronomical Society , 509(2):2551--2565

    2021

  29. [29]

    Huang, L., Gao, Z., and Zhang, D. (2013). Research on multi-fidelity aerodynamic optimization methods. Chinese Journal of Aeronautics , 26(2):279--286

    2013

  30. [30]

    J., Scetbon, M., Pal, S., and Harchaoui, Z

    Irons, N. J., Scetbon, M., Pal, S., and Harchaoui, Z. (2022). Triangular flows for generative modeling: Statistical consistency, smoothness classes, and fast rates. In International Conference on Artificial Intelligence and Statistics , pages 10161--10195. PMLR

    2022

  31. [31]

    and Schäfer, F

    Katzfuss, M. and Schäfer, F. (2023). Scalable bayesian transport maps for high-dimensional non-gaussian spatial fields. Journal of the American Statistical Association , 0(0):1--15

    2023

  32. [32]

    E., Deser, C., Phillips, A., Mai, A., Hannay, C., Strand, G., Arblaster, J

    Kay, J. E., Deser, C., Phillips, A., Mai, A., Hannay, C., Strand, G., Arblaster, J. M., Bates, S., Danabasoglu, G., Edwards, J., et al. (2015). The community earth system model (cesm) large ensemble project: A community resource for studying climate change in the presence of internal climate variability. Bulletin of the American Meteorological Society , 9...

    2015

  33. [33]

    Kennedy, M. C. and O'Hagan, A. (2000). Predicting the output from a complex computer code when fast approximations are available. Biometrika , 87(1):1--13

    2000

  34. [34]

    Kingma, D. P. and Welling, M. (2014). Auto-encoding variational Bayes . In International Conference on Learning Representations

    2014

  35. [35]

    C., Zwiers, F

    Kirchmeier-Young, M. C., Zwiers, F. W., and Gillett, N. P. (2017). Attribution of extreme events in arctic sea ice extent. Journal of Climate , 30(2):553--571

    2017

  36. [36]

    J., Mudryk, L

    Kushner, P. J., Mudryk, L. R., Merryfield, W., Ambadan, J. T., Berg, A., Bichet, A., Brown, R., Derksen, C., D \'e ry, S. J., Dirkson, A., et al. (2018). Canadian snow and sea ice: assessment of snow, sea ice, and related climate processes in canada's earth system model and climate-prediction system. The Cryosphere , 12(4):1137--1156

    2018

  37. [37]

    and Garnier, J

    Le Gratiet, L. and Garnier, J. (2014). Recursive co-kriging model for design of computer experiments with multiple levels of fidelity. International Journal for Uncertainty Quantification , 4(5)

    2014

  38. [38]

    B., Gigu \`e re, M., Brissette, F., Turcotte, R., Braun, M., et al

    Leduc, M., Mailhot, A., Frigon, A., Martel, J.-L., Ludwig, R., Brietzke, G. B., Gigu \`e re, M., Brissette, F., Turcotte, R., Braun, M., et al. (2019). The climex project: A 50-member ensemble of climate change projections at 12-km resolution over europe and northeastern north america with the canadian regional climate model (crcm5). Journal of Applied Me...

    2019

  39. [39]

    Lophaven, S., Nielsen, H., and S ndergaard, J. (2002). Dace - a matlab kriging toolbox, version 2.0. Technical report, Technical University of Denmark

    2002

  40. [40]

    L., Braverman, A

    Ma, P., Kang, E. L., Braverman, A. J., and Nguyen, H. M. (2019). Spatial statistical downscaling for constructing high-resolution nature runs in global observing system simulation experiments. Technometrics , 61(3):322--340

    2019

  41. [41]

    A., Asher, T

    Ma, P., Karagiannis, G., Konomi, B. A., Asher, T. G., Toro, G. R., and Cox, A. T. (2022). Multifidelity computer model emulation with high-dimensional output: An application to storm surge. Journal of the Royal Statistical Society Series C: Applied Statistics , 71(4):861--883

    2022

  42. [42]

    Mardani, M., Brenowitz, N., Cohen, Y., Pathak, J., Chen, C.-Y., Liu, C.-C., Vahdat, A., Kashinath, K., Kautz, J., and Pritchard, M. (2023). Residual diffusion modeling for km-scale atmospheric downscaling

    2023

  43. [43]

    Martynov, A., Laprise, R., Sushama, L., Winger, K., S eparovi \'c , L., and Dugas, B. (2013). Reanalysis-driven climate simulation over cordex north america domain using the canadian regional climate model, version 5: model performance evaluation. Climate dynamics , 41(11):2973--3005

    2013

  44. [44]

    M., Moselhy, T., Parno, M., and Spantini, A

    Marzouk, Y. M., Moselhy, T., Parno, M., and Spantini, A. (2016). Sampling via measure transport: An introduction . In Ghanem, R., Higdon, D., and Owhadi, H., editors, Handbook of Uncertainty Quantification . Springer

    2016

  45. [45]

    Mearns, L., McGinnis, S., Korytina, D., Arritt, R., Biner, S., Bukovsky, M., Chang, H., Christensen, O., Herzmann, D., Jiao, Y., et al. (2017). The na-cordex dataset, version 1.0. ncar climate data gateway, boulder co

    2017

  46. [46]

    O., Gutowski, W., Jones, R., Leung, R., Mcginnis, S., Nunes, A., and Qian, Y

    Mearns, L. O., Gutowski, W., Jones, R., Leung, R., Mcginnis, S., Nunes, A., and Qian, Y. (2009). A regional climate change assessment program for North America . Eos , 90(36):311--312

    2009

  47. [47]

    and Laizet, S

    Mole, A. and Laizet, S. (2024). Multi-fidelity bayesian optimisation of wind farm wake steering using wake models and large eddy simulations. Flow, Turbulence and Combustion , pages 1--26

    2024

  48. [48]

    Modernizing Probable Maximum Precipitation Estimation

    NASEM (2024). Modernizing Probable Maximum Precipitation Estimation . The National Academies Press, Washington, DC

    2024

  49. [49]

    Nikolenko, S. I. (2021). Synthetic Data for Deep Learning , volume 174. Springer

    2021

  50. [50]

    Niu, R., Wu, D., Kim, K., Ma, Y.-A., Watson-Parris, D., and Yu, R. (2024). Multi-fidelity residual neural processes for scalable surrogate modeling. arXiv preprint arXiv:2402.18846

    Pith/arXiv arXiv 2024

  51. [51]

    Peherstorfer, B., Willcox, K., and Gunzburger, M. (2018). Survey of multifidelity methods in uncertainty propagation, inference, and optimization. Siam Review , 60(3):550--591

    2018

  52. [52]

    D., and Karniadakis, G

    Perdikaris, P., Raissi, M., Damianou, A., Lawrence, N. D., and Karniadakis, G. E. (2017). Nonlinear information fusion algorithms for data-efficient multi-fidelity modelling. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , 473(2198):20160751

    2017

  53. [53]

    J., Slinskey, E., Graves, S., Biyani, S., Wang, B., Cropper, S., and the UCLA Center for Climate Science Team (2024)

    Rahimi, S., Huang, L., Norris, J., Hall, A., Goldenson, N., Krantz, W., Bass, B., Thackeray, C., Lin, H., Chen, D., Dennis, E., Collins, E., Lebo, Z. J., Slinskey, E., Graves, S., Biyani, S., Wang, B., Cropper, S., and the UCLA Center for Climate Science Team (2024). An overview of the western united states dynamically downscaled dataset (wus-d3). Geoscie...

    2024

  54. [54]

    and Mohamed, S

    Rezende, D. and Mohamed, S. (2015). Variational inference with normalizing flows. In International conference on machine learning , pages 1530--1538. PMLR

    2015

  55. [55]

    Sch \" a fer, F., Katzfuss, M., and Owhadi, H. (2021). Sparse Cholesky factorization by Kullback-Leibler minimization . SIAM Journal on Scientific Computing , 43(3):A2019--A2046

    2021

  56. [56]

    J., and Owhadi, H

    Sch\" a fer, F., Sullivan, T. J., and Owhadi, H. (2021). Compression, inversion, and approximate pca of dense kernel matrices at near-linear computational complexity. Multiscale Modeling & Simulation , 19(2):688--730

    2021

  57. [57]

    R., and Wills, R

    Schneider, T., Leung, L. R., and Wills, R. C. (2024). Opinion: Optimizing climate models with process knowledge, resolution, and artificial intelligence. Atmospheric Chemistry and Physics , 24(12):7041--7062

    2024

  58. [58]

    and Craig, G

    Selz, T. and Craig, G. C. (2015). Simulation of upscale error growth with a stochastic convection scheme. Geophysical Research Letters , 42(8):3056--3062

    2015

  59. [59]

    Stein, M. L. (2002). The screening effect in kriging . Annals of Statistics , 30(1):298--323

    2002

  60. [60]

    Stein, M. L. (2011). When does the screening effect hold? Annals of Statistics , 39(6):2795--2819

    2011

  61. [61]

    M., Buekenhout, D., and Willems, P

    Tabari, H., Paz, S. M., Buekenhout, D., and Willems, P. (2021). Comparison of statistical downscaling methods for climate change impact analysis on precipitation-driven drought. Hydrology and Earth System Sciences , 25(6):3493--3517

    2021

  62. [62]

    Tague, C. (2024). Digital twins of the E arth: can they keep up?

    2024

  63. [63]

    Umakant, J., Sudhakar, K., Mujumdar, P., and Rao, C. R. (2007). Ranking based uncertainty quantification for a multifidelity design approach. Journal of aircraft , 44(2):410--419

    2007

  64. [64]

    Vecchia, A. (1988). Estimation and model identification for continuous spatial processes . Journal of the Royal Statistical Society, Series B , 50(2):297--312

    1988

  65. [65]

    and Marzouk, Y

    Wang, S. and Marzouk, Y. (2022). On minimax density estimation via measure transport. arXiv preprint arXiv:2207.10231

    Pith/arXiv arXiv 2022

  66. [66]

    Wiemann, P. F. V. and Katzfuss, M. (2023). Bayesian nonparametric generative modeling of large multivariate non-Gaussian spatial fields . Journal of Agricultural, Biological and Environmental Statistics , 28(4):597--617

    2023

  67. [67]

    and Nickisch, H

    Wilson, A. and Nickisch, H. (2015). Kernel interpolation for scalable structured gaussian processes (kiss-gp). In International conference on machine learning , pages 1775--1784. PMLR

    2015

  68. [68]

    Wu, D., Chinazzi, M., Vespignani, A., Ma, Y.-A., and Yu, R. (2022). Multi-fidelity hierarchical neural processes. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages 2029--2038

    2022

  69. [69]

    Xu, Z., Han, Y., and Yang, Z. (2019). Dynamical downscaling of regional climate: A review of methods and limitations. Science China Earth Sciences , 62(2):365--375

    2019

  70. [70]

    B., and Li, Z

    Yu, G., Wright, D. B., and Li, Z. (2020). The upper tail of precipitation in convection-permitting regional climate models and their utility in nonstationary rainfall and flood frequency analysis. Earth's Future , 8(10):e2020EF001613. e2020EF001613 2020EF001613

    2020

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.