Merging AI with Physical Models for the Study of the Universe

Astrophysics is entering a new era of large-scale and high-quality surveys of the sky, bringing unprecedented opportunities for new discoveries, but also outstanding challenges at all levels of the scientific analysis, as conventional techniques do not allow us to fully exploit these rich datasets. While AI can play a role in overcoming these challenges, in most cases a deep learning approach alone proves to be insufficient, due to the black box nature of neural network models and usually a lack of robust uncertainty quantification. In this talk, I will present a line of research merging Generative AI, Differentiable Programming, and explicit physical models to analyse astrophysical datasets. In doing so, this approach retains the interpretability of a physical model, and benefits from proper uncertainty quantification made possible by a Bayesian inference framework, thus making AI-based methods suitable for scientific purposes. I will showcase a showcase a range of applications of these new techniques from the pixel-level analysis of galaxy images, to reconstructing maps of the dark matter distribution in the Universe, and constraining cosmological models.