Responsabile: Mario Pasquato

Descrizione: Science is hard. Multiple times throughout its history we believed we were looking at the Universe when in fact we were really looking into a mirror. Overcoming our cognitive biases by questioning the very foundations of what we do is crucial if we want to make real progress. The notions of cause and effect underlie most of our scientific (as well as pre-scientific) understanding of the world, yet most physicists would be hard pressed to come up with clear definitions of these concepts. But worry not! Machine learning is dealing with this
conundrum for us. A few seminal works (e.g. Pearl J. 2000) from the late nineties kickstarted the field of causal inference, sitting astride statistics and computer science, by providing operational definitions for the notion of causal relation. As a result, data-driven causality took several observational sciences (such as economics of epidemiology) by storm in the last few decades. Astronomy was spared until about a year ago, when finally one of the most stubborn chicken-and-egg problems in galaxy formation (AGN-Galaxy coevolution) was attacked with causal discovery algorithms (Pasquato et al. 2023, Jin et al. 2024 submitted). With this thesis you will help us make sure that the cat stays out of the box: you will apply causal discovery techniques to cosmological simulations as well as to observational data to answer fundamental questions such as does the future cause the past or is it the other way round? On this topic there are opportunities for collaboration with Prof. Francisco Villaescusa Navarro (Flatiron institute, NY), Prof. Elie Wolfe (Perimeter Institute).

Durata: 9-12 mesi

Laurea: Magistrale

Prerequisiti: Nessuno