"Learning from Simulations: Decoding the Large-Scale Structure of the Universe and Training AI Scientists"
Our observations have painted a simple portrait of cosmic evolution, yet fundamental questions remain unanswered: What is dark matter? What drives the accelerated expansion of the cosmos? How did it begin? In the first part of this talk, I will present machine learning frameworks that bridge numerical simulations and increasingly precise observations of the large-scale structure of the Universe. I will show how generative models serve as both emulators and tools for parameter estimation, enabling direct inference from high-dimensional data without simplified summary statistics, and how they let us reconstruct the initial conditions of the Universe and identify anomalies that might signal departures from our standard cosmological model.
In the second part, I will turn simulations from a tool for decoding our Universe into a testbed for discovery: can a language model discover physics it was never taught? Static benchmarks reward retrieving known answers, whereas science is an interactive loop of hypothesis, experiment, and revision. Our benchmark, DiscoverPhysics, evaluates that loop directly. It places agents in simulated worlds with laws that deviate from our own, where they must run experiments and submit both an explanation and an implementation of the law they inferred. I will then present Discover-Many-Worlds, which turns this benchmark into a reinforcement learning environment of automatically generated worlds with verifiable rewards for training AI scientists.
