KIPAC Seminar: Inverting the Universe: Field-Level Inference with Differentiable and Neural Hydrodynamics

Abstract: The observed Universe is not a static distribution of galaxies and gas, but the visible outcome of a coupled dynamical history. Gravity amplifies primordial fluctuations into the cosmic web; hydrodynamics generates shocks, flows, and instabilities; radiative cooling produces multiphase structure; turbulence mixes energy and metals; and feedback from stars and black holes regulates the gas that ultimately becomes observable. We can simulate these processes forward with increasing realism, but the deeper scientific challenge is the inverse problem. We need to use observations of the actual Universe to infer the initial conditions, cosmological parameters, and uncertain baryonic physics that shaped it.

In this talk, I will present differentiable hydrodynamics, diffHydro, as a framework for this inverse problem. Instead of comparing simulations to a small set of summary statistics, the method uses the actual observational information (such as stellar density, spectral absorption lines, and diffuse emission) at the field level. Gradients through the simulation connect these observed fields back to initial conditions, subgrid parameters, and learned physical closures. By integrating stochastic differentiable subgrid models and neural-network components directly inside the hydrodynamical solver, we can transform simulations into interpretable, gradient-based inference engines. The goal is not to replace simulations with a black-box, but to make simulations themselves trainable, interpretable, and directly connected to the observed Universe.