Machine Learning & AI
Connecting simulations to observations with machine learning and simulation-based inference
Modern cosmology and galaxy formation share a problem: the models are too complex for analytic likelihoods, and the simulations are too expensive to run everywhere we need them. Machine learning offers a way through — emulators that learn the mapping from physical parameters to observables, neural networks that extract cosmological information directly from field-level data, and simulation-based (implicit likelihood) inference that turns suites of simulations into rigorous posteriors on the parameters of cosmology and galaxy formation.
Much of this work builds on the CAMELS project — thousands of cosmological simulations with varied cosmology and feedback physics, designed from the start as a machine-learning training set (Villaescusa-Navarro et al., 2021; Villaescusa-Navarro et al., 2023). The image above shows gas surface-density maps from a sample of these simulations. With suites like these we can perform robust field-level inference of cosmological parameters (Shao et al., 2023) and, conversely, use observations to calibrate the subgrid physics of the simulations themselves (Jo et al., 2023).
On the methods side, we develop tools and generative models for the community: the LtU-ILI framework packages the full simulation-based inference pipeline for astrophysical applications (Ho et al., 2024), autoregressive networks generate halo catalogs directly (Pandey et al., 2025), and hybrid physics-informed approaches like sapphire aim to combine the interpretability of analytic models with the flexibility of machine learning (Pandya et al., 2026).