Home Publications ExtremesLearning (CNRS 80 PRIME Project)

Développement de méthodes de machine learning pour combiner les biais multi-modéles dans les études de détection et d’attribution d’extrêmes climatiques

Les mode?les de climat utilise?s par les expe?riences CMIP pour le GIEC sont en ge?ne?ral valide?s sur les e?ve?nements moyens et les fluctuations typiques, et des biais sont ainsi de?termine?s. Le but de ce projet sera de mettre en place un cadre nouveau pour l’e?valuation des biais mode?les pour l’e?tude spe?cifique des e?ve?nements extre?mes. Nous de?velopperons un cadre the?orique pour combiner les me?rites relatifs des diffe?rents mode?les en fonction de leur capacite? a? reproduire des extre?mes spe?cifiques: tempe?ratures caniculaires et pre?cipitations extre?mes. Les me?thodes de la the?orie de la statistique des valeurs extre?mes et de la physique statistique seront au cœur de notre analyse et seront coupler a? des techniques de machine learning de?die?s a? lagre?gation des biais et d’incertitudes. Une fois le cadre conceptuel pose?, les donne?es des expe?riences CMIP seront analyse?es en de?tails. Les algorithmes de?veloppe?s pourront e?tre utiles a? dautres types danalyses de risque.

Exemple d’attribution de précipitations extrêmes

Cette étude parue dans Journal of Climate (Naveau et Thao, 2022) montre que le biais des simulations climatiques globales peut être efficacement géré lorsque la métrique appropriée est choisie. Cette métrique conduit à une procédure de machine learning (couplage entre la théorie des valeurs extrêmes et l’optimisation de la divergence de Kullback). Cela nous permet de démontrer que des combinaisons convexes optimales de sorties CMIP peuvent améliorer la détection des temps d’émergence du changement climatique. Notre procédure d’analyse des données est appliquée au maximum annuel de précipitations des bases de données CMIP5 et CMIP6, voir figure. L’attribution du forçage anthropique apparaît clairement dans les précipitations extrêmes du début du XXIe siècle.

Highlights of the PhD work of Paula Gonzalez

• P. Gonzalez, P. Naveau, S. Thao, and J. Worms. A statistical method to model non-stationarity in precipitation records changes. Submitted to Geophysical Research Letters, 2023.
• Invited paper at the 2022 Joint Statistical Meetings in Washington DC, USA, “Statistical Frameworks for Studying Climate Extremes in a Detection and Attribution Context”, August 2022
• PhD ESTIMR Seminar at LSCE on “Record analysis assumptions in a changing world”, Gif-sur-Yvette, May 2022
• Poster at the Third edition of VALPRED (Validation of prediction), “Record analysis of climate models in a non-stationary context”, Centre Paul Langevin, Aussois, October 2021

Related articles dealing climate attribution, extremes modelling and/or machine learning

• P. Gonzalez, P. Naveau, S. Thao, and J. Worms. A statistical method to model non-stationarity in precipitation records changes. Submitted to Geophysical Research Letters, 2023.
• Juliette Legrand, Philippe Naveau, and Marco Oesting. Evaluation of binary classifiers for asymptotically dependent and independent extremes. arXiv preprint arXiv :2112.13738, 2023. Submitted to Journal American Statistical Association, 2023
• Philomène Le Legall, Anne-Catherine Favre, Philippe Naveau, and Alexandre Tuel. Non-parametric multimodel regional frequency analysis applied to climate change detection and attribution. arXiv preprint arXiv :2111.00798, 2023.
• Touqeer Ahmad, Carlo Gaetan, and Philippe Naveau. Modelling of discrete extremes through extended versions of discrete generalized pareto distribution. arXiv preprint arXiv :2210.15253, 2023.
• Lafon, N., E. Fablet and P. Naveau. Uncertainty quantification when learning dynamical models and solvers with variational methods, in revision of Journal of Advances in Modeling Earth Systems.
• Lafon N, R.Fablet and P. Naveau. A VAE approach to sample multivariate extremes. Submitted to ICML (2023- International Conference on Machine Learning)
• P. Naveau and S. Thao. Multi-model errors and emergence times in climate attribution studies. Journal of Climate, pages 4791–4804, 2022.
• J. Worms and P. Naveau. Record events attribution in climate studies. Environmetrics, 2022.
• M. Taillardat, A.L. Fougères, P. Naveau, and R. De Fondeville. Evaluating probabilistic forecasts of extremes using continuous ranked probability score distributions. International Journal of Forecasting, 2022.
• Davide Faranda, Salvatore Pascale, Burak Bulut. Persistent anticyclonic conditions and climate change exacerbated the exceptional 2022 European-Mediterranean drought. Environmental Research Letters, 2023 https://doi.org/10.1088/1748-9326/acbc37
• Faranda, D., Bourdin, S., Ginesta, M., Krouma, M., Noyelle, R., Pons, F., Yiou, P., and Messori, G.: A climate-change attribution retrospective of some impactful weather extremes of 2021, Weather Clim. Dynam., 3, 1311–1340, https://doi.org/10.5194/wcd-3-1311-2022, 2022.
• Pauline Rivoire, Philomène Le Gall, Anne-Catherine Favre, Philippe Naveau, and Olivia Martius. High return level estimates of daily era-5 precipitation in europe estimated using regionalized extreme value distributions. Weather and Climate Extremes, 38 :100500, 2022.
• Romain Pic, Clément Dombry, Philippe Naveau, and Maxime Taillardat. Distributional regression and its evaluation with the crps : Bounds and convergence of the minimax risk. International Journal of Forecasting, 2022.
• Philomène Le Gall, Anne-Catherine Favre, Philippe Naveau, and Clémentine Prieur. Improved regional frequency analysis of rainfall data. Weather and Climate Extremes, page 100456, 2022.
• G. Buriticá and P. Naveau. Stable sums to infer high return levels of multivariate rainfall time series. Environmetrics, 2022.
• Jan Beirlant, Gaonyalelwe Maribe, Philippe Naveau, and Andréhette Verster. Bias reduced peaks over threshold tail estimation. REVSTAT-Statistical Journal, 20(3) :277–304, Jul. 2022.

Summer schools organization related to extremes, geosciences or/and machine learning

1. Co-organizers with E. Blayo, P. Brasseur, G. Peyré, J. Zerubia. Ecole TDMA : Traitement de données massives et apprentissage : applications en géophysique, écologie et shs. In IMAG, Grenoble University, Grenoble, 5-9 juin 2023.
2. Co-organizers with O. Wintenberger and M. Thomas. Doctoral courses on extremes, risk, climat and enviromnent. In INSTITUT HENRI POINCARE, Paris, 7-17 March 2022.

Organisation of international workshops related to extremes, geosciences or/and machine learning

1. Co-organizers with J. Neslehova, M. Maathuis, L. Mhalla. International workshop on combining causal inference and extreme value theory in the study of climate extremes and their causes. In Banff International Research Station (BIRS) for Mathematical Inovation and Discovery, UBC Okanagan campus in Kelowna, B.C, Canada, 6 June- 1 July 2022.
2. Co-organizers with J.M. Bardet, S. Subba Rao, A. Veraart and R. Von Sachs. International conference on adaptive and high-dimensional spatio-temporal methods for forecasting. In Centre International de Recherche Mathématique (CIRM), Lumini, 26-30 Sept 2022.
3. Scientific committee member of the workshop. Machine learning and uncertainties in climate simulations. In Moulin-Mer, Logonna-Daoulas, Finistère, France, 6-9 June 2022.
4. Co-organizers with C. Dombry, O. Wintenberger and C. Tellier. Valpred workshop (validation of forecasting), 4th edition. In Centre Paul Langevin, Aussois, 3-6 April 2023.
5. Co-organizers with C. Dombry and O. Wintenberger. Valpred workshop (validation of forecasting), 3rd edition. In Centre Paul Langevin, Aussois, 4-7 october 2021.
6. Co-organizers with C. Dombry and O. Wintenberger. Valpred workshop (validation of forecasting), 2nd edition. In Centre Paul Langevin, Aussois, 9-12 MArch 2020.

Talks in 2023 (dealing with extremes, climate change and attribution)

[1] Invited talk at “Extremes and Time Series : A Workshop on the Occasion of Richard Davis’ 70th Birthday”. Records analysis in climate attribution problems. In New York (Columbia University), January 2023. Joint work with P. Gonzalez, J. Worms and S. Thao.

Talks in 2022 (dealing with extremes, climate change and attribution)

. [1] Talk at the “Rencontres statistiques au sommet de Rochebrune”. Evaluation of binary classifiers for environmental extremes. In Megève, France, March 2022. Joint work with J. Legrand and M. Oesting.
. [2] One-line seminar organized by J. Runge (TU, Berlin). Multivariate extreme values theory and causality theory : a short review with a focus on climate change. In Berlin (on-line), TU Berlin, December 2022. Joint work with A. Kiriliouk.
. [3] Keynote speaker at the 5th International Conference on Advances in Extreme Value Analysis and Application to Natural Hazards. Evaluation of binary classifiers for environmental extremes. In Orlando, USA, May 2022. Joint work with J. Legrand and M. Oesting.
. [4] Invited talk at the workshop ENS EDIPI. Advanced statistical methods for extremes : how to compare them. In Paris, France, May 2022. Joint work with J. Legrand and M. Oesting.
. [5] Invited talk at the IMSI workshop on Detection and Attribution of Climate Change. Climate extreme event attribution and extreme value theory. In Chicago, USA (on-line), October 2022. Joint work with P. Gonzalez, S. Thao and J. Worms.
. [6] Invited talk at the “7èmes Rencontres de Statistique – Science des données : Environnement et climat”. Climate extreme event attribution and extreme value theory. In Vannes, France, November 2022.
. [7] Invited talk at “Journée SAMA”. Evaluation of binary classifiers for environmental extremes. In LMD-ENS, Paris, France, April 2022. Joint work with J. Legrand and M. Oesting.
. [8] Invited talk at IDAG . Two simple tools to deal with multi-model errors and to compare maxima distributions. In Australia (on-line), June 2022. Joint work with P. Gonzalez.
. [9] Invited seminar at the ECOP meeting. Climate extreme event attribution and extreme value theory. In Cergy, France, December 2022.
. [10] Invited seminar at the EC Joint Research Center. Statistical review on climate extreme event attribution. In Ispra, Italy, November 2022.
. [11] Invited seminar at HEC Lausanne (UNIL). Evaluation of binary classifiers for environmental extremes. In Lausanne, Swizterland, Feb 2022. Joint work with J. Legrand and M. Oesting.
. [12] Intervenant à la table ronde des journées MIA :. L’intelligence artificielle dans le contexte des transitions énergétique et écologiques. In Grenoble, France (on-iline), December 2022.
. [13] Expert panel for the National Academy of Sciences-Engineering-Medecine, Frontiers of Extreme Event Attribution. Emerging methods and overcoming data biases. In Washington DC, USA (on-line), June 2022.

Talks in 2021 (dealing with extremes, climate change and attribution)

. [1] Invited talk at Virtual ISI World Statistics Congress 2021. Non-parametric multimodel regional frequency analysis applied to climate change detection and attribution. In On-line, November 2021.
. [2] Invited talk at the Emeritaat Jan Beirlant of Dept. of Math. of Leuven. Evaluation of binary classifiers for environmental extremes. In Leuven, Belgium, 2021.
. [3] Invited talk at the DATA IA DS3 conference. How to attribute extremes in climate studies. In Ecole X, Palaiseau, France, January 2021.
. [4] Invited talk at Statistical Estimation and Detection of Extreme Hot Spots, with Environmental and Ecological Applications. Extreme forecasts evaluation. In Kaust, SA (on-line), 2021.
. [5] Invited talk at Department of Environmental Sciences, Informatics and Statistic of Venise University. Evaluation of binary classifiers for environmental extremes. In Venise, Italy, November 2021. Joint work with J. Legrand and M. Oesting.
. [6] Invited talk at Department of Environmental Sciences, Informatics and Statistic of Venise University. Combining ensemble climate simulations to take into account multi-model error. In Venise, Italy, November 2021. Joint work with S. Thao.
. [7] Invited seminar at the UBO Math Dep. Climate extreme event attribution and multivariate extreme value theory. In Brest, France, Juin 2021.
. [8] Invited seminar at Columbia University (stat dept) . Detecting changes in multivariate extremes from climatological time series. In New York, USA (on-line), June 2021. Joint work with S. Engelke and C. Zhou.
. [9] DS3 Data Science Summer School of Polytechnique. In Paris (on-line), 2021.
. [10] 13th International Workshop on Rare-Event Simulation – RESIM 2021. In Paris (on-line), 2021.