Detection/attribution (D/A) studies aim to detect recent changes in climate and establish the relative contribution of different external forcings (greenhouse gases, anthropogenic and natural aerosols, stratospheric ozone) using statistical inference techniques. These methods use low-resolution observations and outputs from numerical simulations performed with general circulation models. Today, these studies must respond to many challenges. On the one hand, the current methods are mainly developed to operate on a global scale, and give less significant results on a regional scale. On the other hand, important methodological uncertainties remain, because the most studied methods are based on little explored hypotheses (for example the additivity of forcings and internal variability), with important choices in the reduction of the dimensions of the data used. Similarly, observational uncertainty and that of climate models are rarely taken into account.
The objective of the thesis is to explore the contribution of recent methods of statistical learning and deep neural networks to meet these different challenges. From a methodological point of view, the aim is to develop algorithms capable of operating at global and regional scales, taking into account the uncertainties of models and observations. We will rely on recent advances in the field of neural networks to study in particular the reduction of dimensionality, the taking into account of local spatio-temporal dependencies for attribution, and the probabilistic modeling of dependencies between observations and simulations.
The work will be based on various observations (temperature or tropospheric humidity; period 1900- 2020 or 1980-2020) and results of simulations performed with general circulation models.

 

PhD student: Constantin Bone

PhD supervisor: Guillaume Gastineau

Research laboratory: LOCEAN - Laboratoire d’océanographie et du climat : expérimentations et approches numériques