Soutenance de thèse d'Andreas Filipp

Probing Subgalactic Dark Matter with Strong Gravitational Lensing and Machine Learning

LCDM, the standard model of cosmology, successfully describes the large-scale structure of our universe. However, small-scale observations present discrepancies that motivate alternative dark matter (DM) models, such as Warm Dark Matter (WDM), which match large-scale predictions while altering small-scale behavior. Strong gravitational lensing provides a purely gravitational probe sensitive to all matter regardless of its nature.

This allows to probe all over- and under-densities along the line-of-sight (LOS), and is not only limited to the main deflector.

Upcoming surveys, such as the Legacy Survey of Space and Time (LSST), are expected to discover an unprecedented number of strong lenses. Traditional lens modeling and inference frameworks, especially with the goal of constraining DM, are computationally intractable for this large volume of lenses.

To overcome these barriers, this thesis leverages machine learning methods to explore LSST's potential to probe the nature of DM, evaluate the limitations of simulation-based inference (SBI) methods, and explore refining the theoretical predictions of the DM halo mass function (HMF) based on observations of the stellar light of the lensing galaxies.

To enable scalable and fast generation of strong lens simulations and their integration in neural networks, this work contributes to the development of caustics, a GPU-accelerated, fully differentiable strong gravitational lensing simulator built in PyTorch. Using caustics, the thesis conducts a forecasting study on the sensitivity of LSST to subgalactic DM structure using Neural Ratio Estimators (NREs).

The study demonstrates that a sample of 2,500 strong lenses in this ground-based image quality can provide constraints on the population-level parameters of the HMF, enabling statistically significant exclusions of non-CDM scenarios. The study shows that both low-mass halos (down to 107  ) and LOS halos contribute to the constraining power. This means that low-mass and LOS populations need to be modeled in lens analyses to avoid artificially weakened or biased constraints.

This is in contrast to what is usually done in strong lens analysis, where LOS halos are ignored, and usually only one massive subhalo at a time is modeled, instead of the full population.

SBI methods like NREs and Sequential Neural Posterior Estimators (SNPEs) offer fast and amortized inference once trained.

This thesis investigates the robustness of these frameworks to distributional shifts and model misspecifications that can be expected to occur when moving from simulated data to observations.

Even minor, realistic deviations between training and test data, such as small differences in background source morphology or macro-lens parameters, can induce significant and unpredictable biases in the inferred DM population parameters. This highlights a critical vulnerability in current SBI applications and underscores the need for caution when applying them to real astrophysical data.

To bridge the gap between observational constraints on the subhalo mass function (SHMF) of individual galaxies and theoretical predictions, the thesis addresses the inherent galaxy-to-galaxy variance of the SHMF within the same cosmology.

The SHMF depends on the morphology, the environment, the merger history, and potentially many other quantities of the individual galaxy, causing a big variance of SHMFs between different galaxies.

For this, we use 1,024 high-resolution hydrodynamical zoom-in simulations from the DREAMS (DaRk mattEr and Astrophysics with Machine learning and Simulations) simulations suite with mock observations generated by Synthesizer, a code to emulate galaxy observables from hydrodynamic simulations.

We create a combination of a Variational Autoencoder (VAE) and a conditional Normalizing Flow (NF) to predict the galaxy-specific SHMF directly from observable galaxy morphologies and an assumed DM particle mass, whilst marginalizing over baryonic variance.

We show that the information from the galaxy images yields generally tighter and more precise posteriors on the SHMF parameters than relying solely on the DM particle mass.

This approach demonstrates that morphology-conditioned priors yield significantly tighter constraints than theoretical averages.

In summary, this thesis provides infrastructure and methodologies to pave the path to the goal of distinguishing between CDM and alternative DM models in the upcoming decade of wide-field surveys.