AstroAI Workshop 2025
Nayantara Mudur
Probing Generative Models for Inference in Cosmology
Presenter: Nayantara Mudur (Harvard University)
Title: Probing Generative Models for Inference in Cosmology
Date/Time: Tuesday, July 8th, 1:30 - 2:00 PM
Abstract: Recent years have witnessed unprecedented developments in generative models. What remains to be seen is the extent to which these developments might transform the scientific process. The ability to generate predictions for observed astrophysical fields from theory and constrain physical models from observables using these predictions forms the cornerstone of modern cosmology. In this talk, I will describe our use of diffusion models to address these two interlinked objectives – as a surrogate model to emulate dark matter density fields conditional on cosmological parameters, and as a parameter inference model that solves the inverse problem of constraining the cosmological parameters of an input field. I will show that the trained diffusion model can be used to approximate the conditional likelihood and sample the posterior on cosmological parameters using Hamiltonian Monte Carlo sampling. We obtain tight constraints on cosmological parameters and further demonstrate that this parameter inference approach can offer greater robustness to noise.