Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 72 of the 72 references it cites were found in a registry or an index; 16 of the 29 sources its claims cite were checked for support.

Toward machine-assisted tuning avoiding the underestimation of uncertainty in climate change projections

Open access
Hourdin, Frédéric, Ferster, Brady, Deshayes, Julie, Mignot, Juliette, Musat, Ionela, Williamson, Daniel · 2023 · Science Advances

Abstract

Documenting the uncertainty of climate change projections is a fundamental objective of the inter-comparison exercises organized to feed into the Intergovernmental Panel on Climate Change (IPCC) reports. Usually, each modeling center contributes to these exercises with one or two configurations of its climate model, corresponding to a particular choice of “free parameter” values, resulting from a long and often tedious “model tuning” phase. How much uncertainty is omitted by this selection and how might readers of IPCC reports and users of climate projections be misled by its omission? We show here how recent machine learning approaches can transform the way climate model tuning is approached, opening the way to a simultaneous acceleration of model improvement and parametric uncertainty quantification. We show how an automatic selection of model configurations defined by different values of free parameters can produce different “warming worlds,” all consistent with present-day observations of the climate system.

Overview

CoChat Summary

TL;DR

The paper applies history matching, a machine-assisted tuning approach, to the IPSL-CM6A-LR climate model and derives four coupled configurations with contrasting equilibrium climate sensitivity (ECS) that remain broadly compatible with present-day observations. It argues that CMIP-style single-configuration contributions omit parametric uncertainty, potentially underestimating uncertainty in climate change projections.

Key findings

  • History matching produced four coupled configurations with ECS values ranging from 3.7 to 5.4 K (abstract; Results reports 3.7 to 5.3 K), compared with 4.6 K for the standard IPSL-6A configuration.
  • Two of the four configurations, Exp 1 and Exp 4, showed present-day climate skill similar to IPSL-6A and could have been selected instead of IPSL-6A for CMIP6; Exp 2 underestimated cloud radiative forcing, and Exp 3 had a very weak Atlantic Meridional Overturning Circulation.
  • With only five configurations (IPSL-6A plus the four new ones), the authors cover about half of the CMIP multi-model ECS range, suggesting that CMIP ensembles may underestimate parametric uncertainty in ECS.
  • Despite the wide ECS range, the simulated climate changes at a given global temperature change were more similar than different across configurations, leading the authors to advocate separate uncertainty quantification for ECS and for climate change at a given warming level.

Methods

  • The authors translated the by-hand tuning targets of IPSL-6A into 14 scalar metrics (11 radiative, 3 rainfall) and treated 18 atmospheric free parameters as unknown, using prior ranges set by expert judgment without knowledge of IPSL-6A values.
  • History matching generated parameter vectors by random sampling, ran 2-year stand-alone atmospheric simulations forced by observed SSTs, computed metrics on the second year, and fitted Gaussian process emulators to rule out implausible parameter vectors over iterative waves (250, 200, and 180 simulations in waves 1–3).
  • ECS was estimated from abrupt4×CO2 coupled simulations using linear regression over the first 150 years, with a clim+4K atmosphere-only proxy used to select configurations with contrasted ECS before running full coupled experiments.

Limitations

  • The paper states that the exploration is “very partial,” that the authors lack expertise to ensure reasonable tuning of sea ice cover or AMOC intensity, and that structural uncertainty is generally not known a priori, so history matching cannot fully guarantee against overtuning.
  • The paper notes that about 800 2-year-long atmospheric simulations were needed to extract nine acceptable configurations, and that no historical simulations were run for Exp 1–4, so most evaluation compared piControl simulations with present-day observations.
  • Evident from the text: the study is limited to one model (IPSL-CM6A-LR), uses a small number of selected configurations, and relies on an imperfect clim+4K ECS proxy for selecting contrasted cases before coupled confirmation.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1126/sciadv.adf2758

Cites

72 of 72 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1. Illustration of the iterative procedure for history matching. Both graphs display the latitudinal variations of the zonally averaged short-wave (SW) cloud radiative effect (CRE) computed as the difference between the total and clear-sky TOA SW radiation. The CERES-EBAF L3b observations are shown in black with error bars of ±4 W/m2 (46). (A) SW CRE computed for the second year of an ensemble of 2-year forced-by-SST simulations for the three successive waves of the history matching procedure with 250, 200, and 180 simulations for waves 1, 2, and 3, respectively. The nine “BEST” simulations are shown in red and the gray line corresponds to a control simulation (CTRL) run with the parameters of the IPSL-6A configuration. (B) Ten-year average SW-CRE obtained in coupled ocean-atmosphere simulations in the multi-model ensemble of CMIP5 (yellow) and CMIP6 (orange) and in the control and four experiments retained from the history matching because of both their good behavior in present-day conditions [taken among the nine “BEST” configurations of (A)] and of their contrasted ECS (see the main text). For these last four experiments, we show the SW CRE obtained in both the coupled (thick line) and stand-alone (thin, 1-year average) atmospheric simulations that are almost superimposed.
Fig. 1. Illustration of the iterative procedure for history matching. Both graphs display the latitudinal variations of the zonally averaged short-wave (SW) cloud radiative effect (CRE) computed as the difference between the total and clear-sky TOA SW radiation. The CERES-EBAF L3b observations are shown in black with error bars of ±4 W/m2 (46). (A) SW CRE computed for the second year of an ensemble of 2-year forced-by-SST simulations for the three successive waves of the history matching procedure with 250, 200, and 180 simulations for waves 1, 2, and 3, respectively. The nine “BEST” simulations are shown in red and the gray line corresponds to a control simulation (CTRL) run with the parameters of the IPSL-6A configuration. (B) Ten-year average SW-CRE obtained in coupled ocean-atmosphere simulations in the multi-model ensemble of CMIP5 (yellow) and CMIP6 (orange) and in the control and four experiments retained from the history matching because of both their good behavior in present-day conditions [taken among the nine “BEST” configurations of (A)] and of their contrasted ECS (see the main text). For these last four experiments, we show the SW CRE obtained in both the coupled (thick line) and stand-alone (thin, 1-year average) atmospheric simulations that are almost superimposed.CC BY-NC · PubMed Central
Fig. 2. Estimation of ECS (K) for the various model configurations, including the CMIP5 (yellow) and CMIP6 (orange) multi-model ensembles. The value of the standard IPSL-6A configuration (orange) is isolated from the other CMIP6 models to serve as a reference point for the work presented here. For the CTRL (with same parameters as IPSL-6A) and the four configurations selected (Exp 1 to 4), we show both the abrupt4×CO2 ECS estimate (circles) computed from 150-year-long coupled ocean-atmosphere simulations and the clim+4K ECS proxy computed from forced by SST 10-year-long simulations (stars). This clim+4K proxy is shown as well for the five best simulations not selected to run coupled simulations (black stars).
Fig. 2. Estimation of ECS (K) for the various model configurations, including the CMIP5 (yellow) and CMIP6 (orange) multi-model ensembles. The value of the standard IPSL-6A configuration (orange) is isolated from the other CMIP6 models to serve as a reference point for the work presented here. For the CTRL (with same parameters as IPSL-6A) and the four configurations selected (Exp 1 to 4), we show both the abrupt4×CO2 ECS estimate (circles) computed from 150-year-long coupled ocean-atmosphere simulations and the clim+4K ECS proxy computed from forced by SST 10-year-long simulations (stars). This clim+4K proxy is shown as well for the five best simulations not selected to run coupled simulations (black stars).CC BY-NC · PubMed Central
Fig. 3. Metrics computed on the CTRL and Exp 1 to 4 piControl coupled simulations. The dispersion across those five configurations is compared with the multi-model spread for CMIP5 (yellow) and CMIP6 (orange) also computed from piControl simulations. It is compared as well with the multi-decadal variability in the IPSL-6A piControl simulation (one of the orange dots in the CMIP6 column), computed from 32 successive 30-year-long periods starting in 1850 (“IPSL-6A pi” column), and with the inter-member dispersion in the 33-member historical ensemble run with IPSL-6A for CMIP6 (“IPSL-6A Ens” column, brown). The period retained for the comparison of the historical simulations with observations is 1979–2005. The comparison of the IPSL pi and historical runs gives an estimate of the error, which is made by comparing with present-day observations simulations run in preindustrial rather than present-day conditions. (A) Global mean SST (°C). (B) Root mean square error computed on the mean seasonal cycle of the SST, between 65°N and 65°S to avoid sea ice dispersion, and after removing the mean bias (RMSC stands for centered root mean square). The mean bias is retrieved to minimize the impact of the global warming between 1850 and present day. (C) December-January-February-Marchmean sea ice extent. (D) Maximum intensity of the Atlantic Meridional Overturning Circulation (AMOC, Sv). (E to H) Root mean square error (RMSE) on the mean seasonal cycle of the Outgoing SW radiation [OSR, W/m2 (E)], Outgoing LW radiation [OLR, W/m2 (F)], total SW + LW Outgoing radiation [TOT RAD, W/m2 (G)], and rainfall [Precip, mm/day (H)].
Fig. 3. Metrics computed on the CTRL and Exp 1 to 4 piControl coupled simulations. The dispersion across those five configurations is compared with the multi-model spread for CMIP5 (yellow) and CMIP6 (orange) also computed from piControl simulations. It is compared as well with the multi-decadal variability in the IPSL-6A piControl simulation (one of the orange dots in the CMIP6 column), computed from 32 successive 30-year-long periods starting in 1850 (“IPSL-6A pi” column), and with the inter-member dispersion in the 33-member historical ensemble run with IPSL-6A for CMIP6 (“IPSL-6A Ens” column, brown). The period retained for the comparison of the historical simulations with observations is 1979–2005. The comparison of the IPSL pi and historical runs gives an estimate of the error, which is made by comparing with present-day observations simulations run in preindustrial rather than present-day conditions. (A) Global mean SST (°C). (B) Root mean square error computed on the mean seasonal cycle of the SST, between 65°N and 65°S to avoid sea ice dispersion, and after removing the mean bias (RMSC stands for centered root mean square). The mean bias is retrieved to minimize the impact of the global warming between 1850 and present day. (C) December-January-February-Marchmean sea ice extent. (D) Maximum intensity of the Atlantic Meridional Overturning Circulation (AMOC, Sv). (E to H) Root mean square error (RMSE) on the mean seasonal cycle of the Outgoing SW radiation [OSR, W/m2 (E)], Outgoing LW radiation [OLR, W/m2 (F)], total SW + LW Outgoing radiation [TOT RAD, W/m2 (G)], and rainfall [Precip, mm/day (H)].CC BY-NC · PubMed Central
Fig. 4. Two-meter temperature (K) change in coupled simulations CTRL (IPSL-6A configuration), Exp 1, and Exp 4. The maps correspond to differences between the abrupt4×CO2 and piControl simulations, averaged over 21 consecutive years. For the left column, this time is centered at year 30 of the simulations. For the right column, it is centered at the time when the global 2 m temperature has increased by 5.5 K, i.e., at year 33 for CTRL, year 65 for Exp 1, and year 17 for Exp 4.
Fig. 4. Two-meter temperature (K) change in coupled simulations CTRL (IPSL-6A configuration), Exp 1, and Exp 4. The maps correspond to differences between the abrupt4×CO2 and piControl simulations, averaged over 21 consecutive years. For the left column, this time is centered at year 30 of the simulations. For the right column, it is centered at the time when the global 2 m temperature has increased by 5.5 K, i.e., at year 33 for CTRL, year 65 for Exp 1, and year 17 for Exp 4.CC BY-NC · PubMed Central
Fig. 5. Surface evaporation (mm/day) and surface net solar radiation change (W/m2) in coupled simulations for a global temperature change of 5.5 K. The maps correspond to differences between the abrupt4×CO2 and piControl simulations, averaged over 21 consecutive years centered at year 33 for CTRL, year 65 for Exp 1, and year 17 for Exp 4.
Fig. 5. Surface evaporation (mm/day) and surface net solar radiation change (W/m2) in coupled simulations for a global temperature change of 5.5 K. The maps correspond to differences between the abrupt4×CO2 and piControl simulations, averaged over 21 consecutive years centered at year 33 for CTRL, year 65 for Exp 1, and year 17 for Exp 4.CC BY-NC · PubMed Central

Verification

Claims the paper makes

13 Valid · 1 Not valid · 11 Not checked · 74 Own results and interpretations

47 more claims · 3 with their cited sources checked

Create a free account to see all 99 claims and to verify the 6 references still open.

See all claims — free account
  • Probably not a substantial claim. 27
  • It may state two separate claims as one. 26

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 58 references

Research works printed but not linked: 2

  1. P. Forster T. Storelvmo K. Armour W. Collins J.-L. Dufresne D. Frame D. Lunt T. Mauritsen M. Palmer M. Watanabe M. Wild H. Zhang Climate Change 2021: The Scientific Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change V. Masson-Delmotte P. Zhai A. Pirani S. L. Connors C. Péan S. Berger N. Caud Y. Chen L. Goldfarb M. I. Gomis M. Huang K. Leitzell E. Lonnoy J. B. R. Matthews T. K. Maycock T. Waterfield O. Yelekçi R. Yu and B. Zhou Eds. (Cambridge Univ. Press 2021) chap. 7 pp. 923–1054.
  2. P. N. Edwards A Vast Machine: Computer Models Climate Data and the Politics of Global Warming (MIT Press 2010).

Cited by

See all 49

Metadata

Authors
Hourdin, F., Ferster, B., Deshayes, J., Mignot, J., Musat, I., & Williamson, D.
Affiliations
  • Centre National de la Recherche Scientifique (FR)
  • University of Exeter (GB)
Corresponding author
Hourdin, Frédéric
Venue
Science Advances
Published
2023-07-21
Volume
9
Issue
29
Pages
eadf2758-eadf2758
ISSN
2375-2548 (ISSN-L)
OpenAlex
W4384818614
Cited by
49
References
58
Citations by year
2026: 11 · 2025: 14 · 2024: 20 · 2023: 4
FWCI
5.49
Type
journal article
Language
en
Open access
Licence
cc-by-nc
Locations
  • Science Advances (publishedVersion, cc-by-nc)
  • PubMed Central (publishedVersion, cc-by-nc)
  • PubMed (publishedVersion)
  • HAL (submittedVersion, cc-by-nc)
Funders
  • Indiana Corn Marketing Council
  • Agence Nationale de la Recherche (ANR-16-CE01-0010, ANR-18-MPGA-0001)
  • Partnership for Advanced Computing in Europe AISBL
  • Centre National d’Etudes Spatiales
  • Centre National de la Recherche Scientifique
  • Grand Équipement National De Calcul Intensif (2021-A0100107732)
  • Sorbonne Université
  • Engineering and Physical Sciences Research Council (N510129, EP/Y005597/1, EP/N510129/, EP/N510129/1)
  • Biotechnology and Biological Sciences Research Council (BB/V011588/1)
  • Natural Environment Research Council (NE/N018486/1)
Topics
  • Physical Sciences › Environmental Science › Global and Planetary Change › Climate variability and models
  • Physical Sciences › Earth and Planetary Sciences › Atmospheric Science › Meteorological Phenomena and Simulations
  • Physical Sciences › Environmental Science › Global and Planetary Change › Climate Change and Geoengineering
Keywords
climate model tuning, machine learning, uncertainty quantification, climate change projections, model parameter selection, free parameters, model intercomparison, IPCC

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Toward machine-assisted tuning avoiding the underestimation of uncertainty in climate change projections. https://app.cochat.ai/papers/10.1126/sciadv.adf2758
The paper
Hourdin, F., Ferster, B., Deshayes, J., Mignot, J., Musat, I., & Williamson, D. (2023). Toward machine-assisted tuning avoiding the underestimation of uncertainty in climate change projections. Science Advances. https://doi.org/10.1126/sciadv.adf2758

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.