Séminaire OSE - Observations Signal & Environment

Title: Combining machine learning and extreme value theory to characterize CMIP6 ensembles' respective performances - Application to Sea Surface Temperature in the Mediterranean Sea

Abstract: Climate models are useful tools to investigate the future of environmental variables of interest. However, they are plagued with structural uncertainty: a computational representation of the dynamics and the variability of the ocean-earth system is necessarily incomplete and truncated at some scales. The choice of prognostic equations, the numerical methods to solve them, the selection of processes to be parameterized, their physical parameterization, not to mention the initialization, all participate in the evolution of climate model outputs for a given forcing scenario. Model ensembles have become common practice to both assess this structural uncertainty and the inherent chaotic variability of the climate system. However, the question of how to compare the performance of these ensembles, let alone how to use them in practice beyond the popular duo of "ensemble mean" and “ensemble anomalies”, remains challenging.
On the case study of the Mediterranean sea surface temperature and marine heat waves, we leverage functional clustering and extreme value theory, combined with the Wasserstein distance, to produce an assessment and comparison tool for a suite of climate model outputs (namely CMIP6 ensembles) and a reference (30-year of reprocessed observations from the High Resolution L4 Sea Surface Temperature Copernicus Mediterranean dataset). This tool allows in particular the joint visualization of model performance w.r.t. the bulk and extreme behaviors of the variable of interest respectively, as well as displaying the spatial intra-model and inter-model variability. In light of this assessment, an approach based on multi-criteria decision-making is implemented and discussed, offering another operational combination of multiple model outputs and an alternative interpretation of model ensembles.

Keywords: Model validation, Model ranking, Extremes, Functional clustering, Marine heat waves.

Authors:  N. Le Carrer, P. Tandeo, F. Sevellec, C. Gaetan, P. Girardi,  P. Naveau