Most of new pressurized-water reactors, in particular SMR, feature natural convection loops for decay heat removal (DHR). In order to guarantee that these innovative systems behave the right way, experimental mockups are built at a reduced scale. Usually only partial similitude can be achieved in these mockups, hence data-driven dimensional analysis must be carried out so that the main phenomena are well captured and only side phenomena are distorted. In this thesis, the first task will be to collect data about the natural convection DHR system (Cathare simulations, existing experiments). Then the PhD student will develop an algorithm to automatically identify those non-dimensional numbers which actually drive the system, inspired by related work in the recent open literature using technique like machine learning, penalization, singular value decomposition…