Description du sujet de thèse

The prediction of scalar-valued or vector-valued outputs is now well understood in statistics and machine learning. However, in many applications, one would be rather interested in predicting a quantity of interest taking values in more complex spaces, such as geometric manifolds, spaces of measures, and Euclidean subspaces with non-linear constraints.

Several possibilities will be explored, focusing in particular on predicting measures. The developed methodology will be applied to real problems in computational fluid dynamics, in collaboration with the Radioprotection and Nuclear Safety Institute (IRSN). Examples of outputs are then flow images and histograms of physical quantities of interest. The industrial motivation is then a better understanding of two-phase flow in steam generators of nuclear plants.

Contexte de travail

Institut de mathématiques de Toulouse. The environment is rich, with many scientific events.