Intrinsic Alignments of Galaxies: analytic and deep machine learning studies in cosmological simulations

The large-scale galaxy surveys of the 2020s and 2030s, such as the Rubin LSST and Roman missions, promise an unprecedented influx of data and provide cosmological weak lensing measurements with unprecedented precision. Achieving this requires rigorous management of systematic uncertainties through robust methods and models. This thesis investigates one key systematic: the intrinsic alignment (IA) of galaxies, where galaxy shapes align with surrounding large-scale structures, mimicking cosmic shear and thus contaminating weak lensing measurements. Understanding IA is essential not only to carry out weak lensing analyses but also to gain insights into galaxy formation and evolution, as IA reflects underlying galaxy physics, including formation history, morphology, and color. This research thesis approaches IA on two fronts: (1) analytical studies examining the dependence of IA on galaxy morphology, and (2) the development of deep learning models for IA, aimed at generating fast mock galaxy catalogs. 1) Analytical Models Approach: We found that IA signals decrease as galaxies’ dynamics become dominated by angular momentum, such as in pure disk galaxies where alignment becomes nearly undetectable. In contrast, dispersion-dominated systems—like elliptical galaxies and the bulges of two-component galaxies—exhibit stronger IA statistics. 2) Deep Learning Approach: Using novel deep learning methods, particularly graph-based geometric deep learning, we trained models on ∼18,000 galaxies from the TNG100-1 hydrodynamical simulation. The generated samples closely match the simulation data, accurately capturing correlations in both geometric and scalar galaxy properties. The theoretical findings offer valuable insights into the physics underlying IA, suggesting that IA models incorporating morphological information may enhance model robustness. Meanwhile, the deep learning approach makes significant strides in efficiently generating mock galaxy catalogs at low computational cost, hence advancing the arsenal of methods for next-generation cosmological surveys.