PhD thesis defense

Field-Level Inference of Cosmic Large-Scale Structure with Non-Differentiable Simulations Deep Learning for Forward Modelling and Reconstruction

by Ludvig Doeser (Stockholm University)

→ Europe/Stockholm
ALB Hörsal 4 (Albano Building 2)

ALB Hörsal 4

Albano Building 2

Description

Abstract
The origin of cosmic structure, the nature of dark matter and dark energy, and the physical processes governing the
evolution of the Universe remain central open questions in cosmology. These leave observable signatures in the largescale
structure of the Universe, making galaxy surveys a key means of investigating them. Realizing the potential in
this data requires fully exploiting increasingly detailed observations by modelling the non-linear evolution of cosmic
structure. High-fidelity numerical simulations provide our most accurate description of this regime. However, they are
conventionally compared with observations using population-level summary statistics. Field-level inference instead fits
simulations directly to observed structures, object by object, causally linking primordial initial conditions and cosmic
composition to the evolved galaxy distribution. Requiring a single model to account for structures across cosmic time
provides a stringent test of cosmological theories and can expose signatures of new physics. However, field-level inference
has largely relied on gradients, while many high-fidelity simulations are not readily differentiable.
In this thesis, I demonstrate how deep learning can bridge this gap. First, I introduce new approaches to
indirectly incorporate non-differentiable simulations into gradient-based inference. In BORG-Emulator, I integrate a fast,
differentiable neural emulator learned from N-body simulations into Bayesian field-level inference, enabling reconstruction
deep into the non-linear regime. In ROBIN-PIP, I develop a framework in which non-differentiable reference simulations
act as physics-informed priors on flexible differentiable models learned during inference. Moving beyond this indirect
use of simulations, I develop LULO for direct field-level reconstruction with non-differentiable forward models. The
framework separates neural search from physical simulation, iteratively updating the initial conditions until the simulated
structures match the data. Finally, I investigate the reliability of neural generative models for field-level posterior inference,
testing their ability to recover the full posterior geometry. Together, these developments expand the physical modelling
available to cosmological field-level inference and open new routes for extracting information from the non-linear largescale
structure of the Universe.

Plugin type
zoom
ID
649-994-16093
Join URL