Uncertainty Quantification for Surface Ozone Emulators using Deep Learning

Freddie Kalaitzis; James Montgomery; Kazuyuki Miyazaki; Kelsey Doerksen; Kevin Bowman; Steven Lu; Yarin Gal; Yuliya Marchetti

REVIEW 4 major objections 5 minor 1 cited by

Uncertainty Quantification for Surface Ozone Emulators using Deep Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Uncertainty-aware deep learning estimates surface ozone model bias and maps where correcting that bias is hardest.

desk verdict Solid, modest UQ comparison for ozone emulators, but the central claim about uncertainty identifying correctable bias is unsupported by quantitative validation. read the letter →

arxiv 2508.04885 v1 pith:MDMADTYU submitted 2025-08-06 cs.LG cs.AI

Kelsey Doerksen , Yuliya Marchetti , Steven Lu , Kevin Bowman , James Montgomery , Kazuyuki Miyazaki , Yarin Gal , Freddie Kalaitzis This is my paper

classification cs.LGcs.AI

keywords surfaceozoneuncertaintyquantificationdeeplearningemulatorMOMO-ChemMCDropoutconformalizedquantileregressionU-Netbiascorrection

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a deep-learning emulator can predict where the MOMO-Chem chemical data assimilation model's surface ozone estimates are biased, and that two different uncertainty-quantification techniques—MC Dropout and conformalized quantile regression—can mark which locations are easy or hard to correct. If this is right, UQ scores become a practical station-selection tool: instead of correcting bias everywhere, an agency could target ground stations whose predictive uncertainty says the local bias is learnable. The authors train uncertainty-aware U-Nets on MOMO-Chem fields and TOAR ground measurements for North America and Europe in June 2019, and report that high-uncertainty regions coincide in both methods and with regions of high true bias. They also find that adding land-use satellite features gives small RMSE gains but no consistent UQ improvement. The strongest claim is that UQ maps can be extrapolated beyond TOAR station coverage, with the caveat that the two methods disagree in some extrapolated areas.

What carries the argument

The load-bearing mechanism is predictive uncertainty used as a proxy for bias-correction difficulty. MC Dropout estimates epistemic uncertainty through stochastic forward passes with dropout at test time; CQR produces conformalized prediction intervals from three quantile heads, and interval length is the uncertainty score. The U-Net is the shared architecture that maps multi-channel MOMO-Chem fields (28 or 51 channels) to per-pixel bias estimates. The argument works by showing that high-uncertainty pixels coincide with high bias and high RMSE in two independent UQ frameworks.

What would settle it

Take all TOAR stations in a region the paper labels high-uncertainty beyond dense station coverage—for example the interior U.S. East Coast or southeastern Europe—withhold them during training, and compare predictive uncertainty against the absolute bias residual at those withheld stations. If the rank correlation is near zero or negative, the extrapolated UQ maps are not a reliable basis for station selection.

Watch

Extended reading notes

Core claim

The central claim is that an uncertainty-aware U-Net, trained on MOMO-Chem atmospheric fields, produces surface-ozone bias predictions whose predictive uncertainty is informative about the difficulty of bias correction. The model predicts the residual between MOMO-Chem eight-hour ozone and TOAR ground-station observations, for regional grids over North America and Europe. Two UQ mechanisms are compared: MC Dropout, which treats dropout as a Bayesian approximation and yields epistemic uncertainty, and CQR, which uses three quantile heads (0.05, 0.5, 0.95) to form a 90% prediction interval whose length is the uncertainty score. On the June 2019 test set, RMSE is roughly 10 ppb for MC Dropout a

Load-bearing premise

The load-bearing premise is that predictive uncertainty computed at grid cells with no TOAR ground station still measures how hard local ozone bias is to correct, so extrapolated maps and station rankings remain trustworthy away from any training target.

Editorial extensions

If this is right

If UQ scores mark difficulty, agencies can rank TOAR stations by uncertainty and spend bias-correction effort only on stations the emulator says are learnable.
Spatial agreement between MC Dropout and CQR high-uncertainty regions gives a criterion for trusting the station ranking where the two methods coincide.
The land-use feature experiments imply that adding more satellite inputs alone does not reliably reduce uncertainty; model improvement needs to address the noisy, high-bias regimes themselves.
Extrapolated UQ maps, if trustworthy, provide a first-pass guide to where new ground-based ozone monitoring would be most informative.
Because high uncertainty tracks high bias, the emulator can be used to flag regions where MOMO-Chem's underlying chemistry, emissions, or boundary-layer processes are least constrained.

Reading between the lines

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

A direct extension is an active-learning loop: apply bias correction at high-UQ stations first and test whether RMSE drops faster than with random or low-UQ selection; this would turn the qualitative spatial agreement into a quantitative operational rule.
The disagreement the paper observes between MC Dropout and CQR in extrapolated areas suggests a conservative principle: rely on UQ-based station rankings mainly where the two methods agree, and treat disagreement zones as places needing new observations rather than model fixes.
The same uncertainty-as-difficulty criterion should transfer to other chemical species or other data-assimilation systems, since nothing in the argument is specific to ozone photochemistry.
A quantitative version of the paper's station-selection claim would be a rank correlation between predictive interval length and station-level RMSE across all TOAR stations; the paper reports representative stations but does not compute this global rank statistic.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The paper is a clean, honest application of two established UQ methods to a real problem: using a U-Net to estimate MOMO-Chem surface ozone bias from TOAR station data. What's new is the head-to-head comparison of MC Dropout and CQR on this specific task for North America and Europe in June 2019. That's a legitimate empirical contribution, even if the methodology itself is routine. The use of external TOAR targets means there's no circularity, and the land-use ablation is a sensible addition.

The good parts: the authors report RMSE across seeds, they plot spatial maps of both UQ scores and errors, and they include two station-level time series that illustrate what high and low uncertainty look like in practice. They also note where the two UQ methods disagree, which suggests they're not cherry-picking. The paper is upfront about being an initial analysis.

The soft spots are substantial and load-bearing. The central claim — that UQ scores identify regions of high bias and can be used to select stations for bias correction — is never actually tested. Section 3 says evaluation is done by “spatially plotting the averaged scores to observe, if any, visible signatures.” There's no per-station correlation between UQ and error, no binned analysis, no coverage check for the advertised 90% CQR intervals, and no validation of the Figure 4 extrapolation into station-free areas. The two anecdotal stations don't show that UQ ranks error across the full set of stations. If uncertainty scores are uninformative at observed stations, the station-selection use case falls apart.

Also unexplained is the large RMSE gap between CQR and MC Dropout: 18.40 vs 10.76 for North America with 28 features. If CQR's point prediction is the median, higher RMSE under asymmetric error distributions is plausible, but the paper doesn't say. As is, it makes the CQR results look strangely weak. The land-use improvements are small and not significance-tested, so the claim there is more modest and largely fine.

Overall, this is a workshop-level paper with a sound setup and an evaluation that stops short of the strong conclusions in Section 5. I don't doubt the authors' ability to add the missing tests. The right response is not rejection but a serious referee asking for direct UQ-to-error validation, coverage diagnostics, and an explanation of the CQR/MCD gap. I'd want to read that revised version. For now, treat the conclusions as provisional.

Referee Report

4 major / 5 minor

Summary. The paper trains U-Net emulators to predict MOMO-Chem surface ozone bias (model-minus-TOAR residuals) over North America and Europe for June 2019, with and without land-use features. It compares two UQ methods: Monte-Carlo Dropout (epistemic uncertainty) and Conformalized Quantile Regression (prediction interval length). The authors report RMSE and UQ metrics, show spatial UQ/RMSE maps, inspect two example stations, and extrapolate UQ beyond station coverage. They conclude that both UQ methods identify spatially consistent high- and low-difficulty regions, that land-use input reduces epistemic uncertainty in North America, and that UQ can discern optimal and sub-optimal stations for MOMO-Chem bias correction.

Significance. If supported, the result would provide a practical way to select surface stations for bias correction of a chemical transport model, and a useful comparison of two UQ approaches in an Earth-science emulator. The paper uses external TOAR targets, evaluates on a held-out June 2019 test set, and averages over five random seeds, which are strengths. However, the central claim currently rests on visual map agreement, with no quantitative UQ-error concordance, no empirical coverage check for CQR, and no statistical test for the land-use comparison. These gaps are fixable and the contribution could become solid after a major revision.

major comments (4)

[Section 3 (Evaluation) and Section 4/Fig. 2] The central claim that UQ scores identify easy/difficult stations and regions is supported only by visual comparison of maps. There is no station-level quantitative test. I would require: (i) per-station Spearman correlation between mean absolute error and UQ score (CQR interval length, MCD epistemic variance) on the held-out June 2019 set; (ii) binned reliability of UQ versus error; (iii) empirical coverage of the nominal 90% CQR intervals. Without these, the Conclusion's statement that both techniques show overlap 'that match with regions of high bias' and the station-selection use case are unsupported.
[Table 1] CQR RMSE is much larger than MCD RMSE (NA 28-feats: 18.40 vs 10.76; EU 28-feats: 16.50 vs 10.52), and this discrepancy is not discussed. If CQR point predictions are far worse, its interval lengths may be dominated by poor fit rather than by intrinsic difficulty, which confounds the UQ maps in Figs. 2-4. Please explain the gap (e.g., quantile loss optimization, calibration split size), report CQR empirical coverage, and verify that interval length adds information beyond the point-error magnitude.
[Section 4, land-use results] The statement that 'inclusion of GEE data improves epistemic uncertainty in North America, but does not otherwise show significant improvement' is made without a significance test. The Table 1 seed intervals overlap (e.g., NA MCD 10.76±0.05 vs 10.88±0.29; EU CQR 16.50±0.06 vs 15.93±1.23). Please report paired tests across seeds or effect sizes with confidence intervals; 'significant' is otherwise unsupported.
[Section 4, Spatial Extrapolation / Fig. 4] UQ maps are extrapolated beyond TOAR station coverage, and consistency between methods is used as evidence. This is not validation: away from training targets, both methods could share spatial smoothness artifacts. Please evaluate extrapolation by holding out spatial regions/stations, or by comparing extrapolated UQ against withheld station errors in those regions. The current analysis does not support the practical claim that UQ remains meaningful off-station.

minor comments (5)

[Section 5] Typos: 'approximaton' and 'surfaze' should be corrected.
[Figure 2] Colourbar scaling varies per metric, so visual 'overlap' between maps may be affected by arbitrary scaling. Consider using a common normalization or rank-based maps for comparability.
[Table 1] Please clarify how Max/Min/Avg interval and epistemic values are computed over time and across seeds, and report the units.
[Section 4, station selection] The selected max-UQ station had the highest UQ in only 2/5 seeds for CQR, and the min-UQ station in 1/5 seeds. Please report all seeds or a stability metric, since this bears directly on the station-selection use case.
[Section 3, CQR setup] Please specify the calibration split (size, temporal ordering) and the conformal score used, for reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: predictions target external TOAR-derived bias and are evaluated on held-out 2019 data; the UQ-to-bias agreement is an empirical claim, albeit only qualitatively supported.

full rationale

The paper's target variable is surface ozone bias, defined as the residual between MOMO-Chem output and external TOAR station measurements (Section 2). Models are trained on 2005-2018 June data and evaluated on a held-out June 2019 test set with RMSE against that same external target. The UQ scores (MC-Dropout epistemic uncertainty and CQR interval length) are computed from predictive distributions and are not constructed from test-set error or from the bias maps used for comparison. The central claim that high-UQ regions overlap with high-bias regions is therefore an empirical observation, not an identity. The only self-references are citations to MOMO-Chem [2] (the external model being emulated) and the authors' own airPy package; neither substitutes for the derivation. The lack of a quantitative per-station UQ-error correlation or empirical coverage check is a real validation gap, but it concerns the strength of evidence, not circularity. No equation or fitted parameter is defined in terms of the claimed prediction, so the derivation chain is not circular.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The load-bearing assumptions are domain-level: station representativeness, sufficient input information, validity of MC Dropout uncertainty, and spatial generalization beyond station coverage. Model hyperparameters are the main free parameters.

free parameters (3)

MC Dropout rate = 0.1
Chosen for uncertainty estimation during inference; uncertainty magnitudes depend on it, but no sensitivity analysis is reported.
CQR target coverage = 90%
Prediction interval nominal coverage is set to 90%, but empirical coverage is never checked against this target.
Training hyperparameters = learning rate 1e-3, 200 epochs, Adam
Hand-chosen and fixed; no ablation or tuning sensitivity reported.

assumptions (4)

domain assumption TOAR 8-hour daytime surface ozone measurements are treated as unbiased ground truth and representative of 11.1 km grid cells.
The target bias is computed as MOMO-Chem minus TOAR, so any station representativeness error directly contaminates the labels. This is stated in Section 2 but not analyzed.
domain assumption MOMO-Chem chemical and physical fields contain sufficient information to predict MOMO-Chem's own bias at unobserved locations.
The entire emulator premise is that the 28 or 51 input channels encode the causes of bias. No causal or ablation evidence is given.
domain assumption MC Dropout with dropout rate 0.1 provides a meaningful epistemic uncertainty estimate.
The paper relies on the standard interpretation of MC Dropout as a Bayesian approximation; this is cited but not verified for this model.
domain assumption The U-Net trained with station-derived targets generalizes to grid cells without TOAR coverage.
Spatial extrapolation in Figure 4 assumes the learned bias relationship transfers to unmonitored cells, which is never validated with held-out stations.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Uncertainty Quantification for Surface Ozone Emulators using Deep Learning." pith.science (2026). https://pith.science/paper/MDMADTYU

@misc{pith2026250804885,
  author       = {Pith},
  title        = {Pith review of: Uncertainty Quantification for Surface Ozone Emulators using Deep Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MDMADTYU}},
  note         = {Machine review of arXiv:2508.04885}
}

read the original abstract

Air pollution is a global hazard, and as of 2023, 94\% of the world's population is exposed to unsafe pollution levels. Surface Ozone (O3), an important pollutant, and the drivers of its trends are difficult to model, and traditional physics-based models fall short in their practical use for scales relevant to human-health impacts. Deep Learning-based emulators have shown promise in capturing complex climate patterns, but overall lack the interpretability necessary to support critical decision making for policy changes and public health measures. We implement an uncertainty-aware U-Net architecture to predict the Multi-mOdel Multi-cOnstituent Chemical data assimilation (MOMO-Chem) model's surface ozone residuals (bias) using Bayesian and quantile regression methods. We demonstrate the capability of our techniques in regional estimation of bias in North America and Europe for June 2019. We highlight the uncertainty quantification (UQ) scores between our two UQ methodologies and discern which ground stations are optimal and sub-optimal candidates for MOMO-Chem bias correction, and evaluate the impact of land-use information in surface ozone residual modeling.

Figures

Figures reproduced from arXiv: 2508.04885 by the authors.

**Figure 2.** 51-channel MC-Dropout and CQR averaged model results for June 2019 Test set. Top: [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Time series of ground truth and CQR predictions for the ground station with, on average, the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: MC-Dropout epistemic uncertainty (top) and Conformalized Quantile Regression interval [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

Leveraging Deep Learning for Physical Model Bias of Global Air Quality Estimates
cs.LG 2025-08 unverdicted novelty 4.0 of 10

A CNN learns to correct surface ozone model bias, outperforming traditional ML and adding satellite land use as a useful input.

Reference graph

Works this paper leans on

16 extracted references · 13 canonical work pages · cited by 1 Pith paper

[1]

Toward cleaner air and better health: Current state, challenges, and priorities

Wei Huang, Hongbing Xu, Jing Wu, Minghui Ren, Yang Ke, and Jie Qiao. Toward cleaner air and better health: Current state, challenges, and priorities. Science, 385(6707):386–390, 2024. doi: 10.1126/science.adp7832

work page doi:10.1126/science.adp7832 2024
[2]

Miyazaki, K

K. Miyazaki, K. W. Bowman, K. Yumimoto, T. Walker, and K. Sudo. Evaluation of a multi- model, multi-constituent assimilation framework for tropospheric chemical reanalysis. Atmospheric Chemistry and Physics, 20(2):931–967, 2020. doi: 10.5194/acp-20-931-2020

work page doi:10.5194/acp-20-931-2020 2020
[3]

Jacob, Robert M

Isabelle Bey, Daniel J. Jacob, Robert M. Yantosca, Jennifer A. Logan, Brendan D. Field, Ar- lene M. Fiore, Qinbin Li, Honguy Y . Liu, Loretta J. Mickley, and Mar- tin G. Schultz. Global modeling of tropospheric chemistry with assimilated meteo- rology: Model description and eval- uation. Journal of Geophysical Research: Atmo- spheres, 106(D19):23073–23095...

work page doi:10.1029/2001jd000807 2001
[4]

Chaser: A global chemical model of the troposphere 1

Kengo Sudo, Masaaki Takahashi, Jun-ichi Kurokawa, and Hajime Akimoto. Chaser: A global chemical model of the troposphere 1. model description. Journal of Geophysical Research: Atmo- spheres, 107(D17):ACH 7–1–ACH 7–20, 2002. doi: https://doi.org/10.1029/2001JD001113

work page doi:10.1029/2001jd001113 2002
[5]

Watanabe, T

S. Watanabe, T. Hajima, K. Sudo, T. Nagashima, T. Takemura, H. Okajima, T. Nozawa, H. Kawase, M. Abe, T. Yokohata, T. Ise, H. Sato, E. Kato, K. Takata, S. Emori, and M. Kawamiya. Miroc-esm 2010: model description and basic results of cmip5-20c3m experiments. Geoscientific Model Development, 4(4):845–872, 2011. doi: 10.5194/gmd-4-845-2011

work page doi:10.5194/gmd-4-845-2011 2010
[6]

Doury, S

A. Doury, S. Somot, and S. et al Gadat. Regional climate model emulator based on deep learning: concept and first evaluation of a novel hybrid downscaling approach. Climate Dynamics, 60: 1751–1779, 2023. doi: 10.1007/s00382-022-06343-9

work page doi:10.1007/s00382-022-06343-9 2023
[7]

Watson-Parris, Y

D. Watson-Parris, Y . Rao, D. Olivi ´e, Ø. Seland, P. Nowack, and et al. Camps-Valls, G. Climatebench v1.0: A benchmark for data-driven climate projections. Journal of Advances in Modeling Earth Systems, (14), 2022. doi: 10.1029/2021MS002954

doi:10.1029/2021ms002954 2022
[8]

Uncertainty quantification and inter- pretability for clinical trial approval prediction

Lu Y , Chen T, Hao N, Van Rechem C, Chen J, and Fu T. Uncertainty quantification and inter- pretability for clinical trial approval prediction. Health Data Sci, 2024. doi: 10.34133/hds.0126

work page doi:10.34133/hds.0126 2024

Show all 16 references

[9]

Un- certainty aware tropical cyclone wind speed estimation from satellite data, 2024

Nils Lehmann, Nina Maria Gottschling, Stefan Depeweg, and Eric Nalisnick. Un- certainty aware tropical cyclone wind speed estimation from satellite data, 2024. URL: https://arxiv.org/abs/2404.08325

2024 arXiv
[10]

Sulla-Menashe and M.A

D. Sulla-Menashe and M.A. Friedl. Mcd12q1 modis/terra+aqua land cover type yearly l3 global 500m sin grid v006
[11]

Gridded population of the world, version 4 (gpwv4): Population density, 2018

CIESIN. Gridded population of the world, version 4 (gpwv4): Population density, 2018

2018
[12]

Schultz, Sabine Schr¨oder, Olga Lyapina, Owen R

M.G. Schultz, Sabine Schr¨oder, Olga Lyapina, Owen R. Cooper, and Ian Galbally. Tropospheric ozone assessment report: Database and metrics data of global surface ozone observations. Elementa: Science of the Anthropocene, (5:58), 2017. doi: 10.1525/elementa.244

2017 doi
[13]

Dropout as a bayesian approximation: Representing model uncertainty in deep learning

Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Maria Florina Balcan and Kilian Q. Weinberger, editors, Proceedings of The 33rd International Conference on Machine Learning, volume 48 of Proceedings of M...

2016
[14]

Cand ‘es

Yaniv Romano, Evan Patterson, and Emmanuel J. Cand ‘es. Conformalized quantile regression,
[15]

Aleatoric and epistemic uncertainty in machine learn- ing: A tutorial introduction

Eyke H¨ullermeier and Willem Waegeman. Aleatoric and epistemic uncertainty in machine learn- ing: A tutorial introduction. CoRR, abs/1910.09457, 2019. URL http://arxiv.org/abs/1910.09457. 6

1910 arXiv
[2019]

URL https://arxiv.org/abs/1905.03222

1905 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Toward cleaner air and better health: Current state, challenges, and priorities

Wei Huang, Hongbing Xu, Jing Wu, Minghui Ren, Yang Ke, and Jie Qiao. Toward cleaner air and better health: Current state, challenges, and priorities. Science, 385(6707):386–390, 2024. doi: 10.1126/science.adp7832

work page doi:10.1126/science.adp7832 2024

[2] [2]

Miyazaki, K

K. Miyazaki, K. W. Bowman, K. Yumimoto, T. Walker, and K. Sudo. Evaluation of a multi- model, multi-constituent assimilation framework for tropospheric chemical reanalysis. Atmospheric Chemistry and Physics, 20(2):931–967, 2020. doi: 10.5194/acp-20-931-2020

work page doi:10.5194/acp-20-931-2020 2020

[3] [3]

Jacob, Robert M

Isabelle Bey, Daniel J. Jacob, Robert M. Yantosca, Jennifer A. Logan, Brendan D. Field, Ar- lene M. Fiore, Qinbin Li, Honguy Y . Liu, Loretta J. Mickley, and Mar- tin G. Schultz. Global modeling of tropospheric chemistry with assimilated meteo- rology: Model description and eval- uation. Journal of Geophysical Research: Atmo- spheres, 106(D19):23073–23095...

work page doi:10.1029/2001jd000807 2001

[4] [4]

Chaser: A global chemical model of the troposphere 1

Kengo Sudo, Masaaki Takahashi, Jun-ichi Kurokawa, and Hajime Akimoto. Chaser: A global chemical model of the troposphere 1. model description. Journal of Geophysical Research: Atmo- spheres, 107(D17):ACH 7–1–ACH 7–20, 2002. doi: https://doi.org/10.1029/2001JD001113

work page doi:10.1029/2001jd001113 2002

[5] [5]

Watanabe, T

S. Watanabe, T. Hajima, K. Sudo, T. Nagashima, T. Takemura, H. Okajima, T. Nozawa, H. Kawase, M. Abe, T. Yokohata, T. Ise, H. Sato, E. Kato, K. Takata, S. Emori, and M. Kawamiya. Miroc-esm 2010: model description and basic results of cmip5-20c3m experiments. Geoscientific Model Development, 4(4):845–872, 2011. doi: 10.5194/gmd-4-845-2011

work page doi:10.5194/gmd-4-845-2011 2010

[6] [6]

Doury, S

A. Doury, S. Somot, and S. et al Gadat. Regional climate model emulator based on deep learning: concept and first evaluation of a novel hybrid downscaling approach. Climate Dynamics, 60: 1751–1779, 2023. doi: 10.1007/s00382-022-06343-9

work page doi:10.1007/s00382-022-06343-9 2023

[7] [7]

Watson-Parris, Y

D. Watson-Parris, Y . Rao, D. Olivi ´e, Ø. Seland, P. Nowack, and et al. Camps-Valls, G. Climatebench v1.0: A benchmark for data-driven climate projections. Journal of Advances in Modeling Earth Systems, (14), 2022. doi: 10.1029/2021MS002954

doi:10.1029/2021ms002954 2022

[8] [8]

Uncertainty quantification and inter- pretability for clinical trial approval prediction

Lu Y , Chen T, Hao N, Van Rechem C, Chen J, and Fu T. Uncertainty quantification and inter- pretability for clinical trial approval prediction. Health Data Sci, 2024. doi: 10.34133/hds.0126

work page doi:10.34133/hds.0126 2024