Title: A Journey Through Modern Cosmology: From Massive Simulations to AI Scientists
Abstract:
Cosmology is becoming an increasingly computational science. Modern surveys map the Universe with extraordinary precision, but turning these observations into physical understanding requires new methods, new simulations, and increasingly, new ideas. In this talk, I will describe this evolution, highlighting how large numerical simulations and AI are reshaping cosmology and, more broadly, the practice of science.
I will begin with a central question: how much information can we extract from the large-scale structure of the Universe? While traditional analyses often rely on summary statistics such as the power spectrum, a wealth of additional cosmological information resides on non-linear scales and higher-order statistics. I will introduce the Quijote simulations, the largest collection of N-body simulations ever run, and its successor, the Backlight simulations, designed to answer this question.
I will then discuss a major challenge: the small scales that contain so much information are also those most affected by uncertain astrophysical processes. This motivates the development of the CAMELS simulations, which provide a unified framework for studying cosmology and galaxy formation through machine learning, and have become the largest and most diverse collection of hydrodynamic simulations ever assembled. I will also present the DREAMS simulations, which aim to constrain the nature of dark matter by making theoretical predictions for astrophysical observables while accounting for baryonic uncertainties.
Finally, I will turn to the next frontier: coupling massive simulations with increasingly capable AI systems. I will introduce Denario, an agentic system designed to explore the space of scientific ideas at superhuman scale by formulating hypotheses, designing analyses, writing and running code, interpreting results, and iteratively improving its own research directions. I will conclude by discussing what may become possible when simulations across physical scales are combined with large populations of AI agents whose capabilities emerge from their number, coordination, and collective behavior.