Anna Kiriliouk

Extreme value theory is a field of statistics focused on characterizing rare, extreme events—events that don’t happen often but have a major impact when they do. In climate science, extreme value theory can help to understand and predict extreme weather events, like heatwaves, floods, and heavy rainstorms. As our climate changes, these extreme events are expected to become more frequent and severe. Attribution science aims to estimate if, and to what extent, a specific extreme weather event was made more likely by human-caused climate change. This can be achieved by comparing probabilities of extreme events under two scenarios: one in which the world’s climate has been affected by human activities (such as increased greenhouse gas emissions), and one where it hasn’t. 

Hence, aa key component for attribution is the development of advanced extreme-value models capable of handling both high-dimensional data (that is, multiple climate variables simultaneously) and a changing climate.

Summary of the FNRS MIS grant project - Flexible statistical models for compound climate events

Many high-impact weather and climate events arise from a combination of multiple environmental hazards such as high temperatures, heavy precipitation, or strong winds. Characterizing these so-called compound extreme events is essential to assess and mitigate climate-induced risks, especially since their probability of occurrence is expected to increase in view of global warming.

Describing the behavior of multiple environmental variables is commonly achieved by focusing either on the center of their underlying distributions, or on their extremes. The latter is especially challenging, since the number of data points in the tail region of the sample is scarce by definition. Moreover, when interest lies simultaneously in both the center and the tail of the environmental variables, a flexible model is needed to characterize the dataset over the whole range of its support. Such an approach is especially relevant for compound events, where individual variables need not be in an extreme state, but their combination is.

The objectives of this project are to

• propose flexible multivariate dependence models that are capable of a realistic representation of both center and tail(s) of a random vector, 
• propose estimators of the parameters of such models with the goal of inferring failure probabilities for multivariate extreme events and improve their precision via data augmentation 
• project climate-induced risks of some major compound events in terms of global warming.

Philippe Naveau is a professor at the Laboratoire des Sciences du Climat et l'Environnement (CNRS).

"I’ve chosen Philippe Naveau as a mentor because his research builds bridges between the climate sciences and the statistical sciences. His work sparked my interest in both attribution science (the topic of our collaboration) and the study of compound events. 

As a statistician by training, Philippe Naveau's principal research interests are to develop high-level statistical tools from Extreme Value Theory for environmental and climate sciences, and to study algorithms that assess extreme events in complex systems. He has co-authored more than 120 articles, in statistical, climate and environmental journals. He has co-organized 20 international workshops and summer schools linked to the topic of extreme event analysis, advised 15 PhD students, and led 15 national and international projects as a PI or as a package leader. 

For example, he is the current leader of the work package “Which learning paradigms for the representation of geophysical extremes ?” within the French ANR project “Bridging Geophysics and Machine Learning”, and the main PI of the CNRS-INSU project "Extremes Learning" that aims to adapt statistical extreme value theory to climate model projections. He is also the associate editor of three journals: Annals of Applied Statistics, Extremes and Environmetrics."

More information on Philippe Naveau