Princeton University Astrophysics Special Seminar

Astrophysics and several other scientific disciplines currently experience the confluence of three separate developments: large observational programs provide more data than ever; massive simulations probe ever more complex phenomena; and rapid advances in machine learning establish entirely new ways of dealing with and interpreting data. None of these three pillars can stand on their own to make meaningful progress in the physical sciences. I will present three areas of research that combine accurate statistical modeling, deep neural networks, and numerical simulations: realistic modeling of galaxies; simulation-based inference in physical hydrology and cosmology; and science-driven design of new surveys and observing programs. From these examples, I will derive five principles for constructing machine-learning approaches to advance the physical sciences.