KIPAC Seminar: Field-Level Inference in the Multimodal Cosmos: Scaling Scientific Discovery Across Fields with AI

Modern cosmology is entering a multi-probe era, combining galaxy clustering, weak lensing, the CMB, and more to constrain fundamental physics and astrophysics. Extracting the full information content of these datasets demands inference that goes beyond traditional summary statistics, and faces challenges similar to HEP: high-dimensional data and systematic uncertainties that must be propagated end-to-end. In this talk, I will motivate field-level inference, which directly fits the observed maps of the sky to reconstruct the initial conditions of the Universe and optimally infer cosmological parameters. I will review different approaches spanning differentiable forward modeling to generative diffusion models, highlighting applications to accelerate discovery with DESI, Simons Observatory, and Rubin LSST. I will outline a path to multi-probe analyses where separate observables are used to simultaneously increase constraining power and provide stringent internal consistency checks—establishing a robust, unified, end-to-end framework to analyze the entire cosmic sky and maximize the fundamental physics return of the next generation of cosmological surveys. I will close by showcasing how foundation models and AI agents can bridge the gap between physical domains, enabling a new era of interdisciplinary discovery across fundamental physics.