MASS ACCRETION GROUPING ANALYSIS: USING MACHINE LEARNING TO INVESTIGATE DARK MATTER HALOS
Chason, Nicholas Allen
0000-0001-6885-6022
:
2022-11-23
Abstract
We investigate various methods of grouping halo mass accretion histories from a cosmological N-body simulation and the resulting two-point spatial bias of each population. While several authors have used the time at which a halo had accumulated half of its total mass (at z=0) to separate halos into two (or more) populations - early and late forming halos - we seek to understand how utilizing the entire halo history effects halo groupings other than half-mass age. As dark matter halos accrete their mass in a noisy manner over most of their history, we explore alternate techniques, using a suite of machine learning methods, to classify halos into 3 groups per algorithm in a data-driven approach. We mostly discuss unsupervised approaches, however, we briefly explore supervised and alternative approaches. There is great diversity in the groupings of halos by their accretion histories among different methods with below a 60% match on average in terms of similar group assignments. We compare the two-point clustering statistics of these groups with the conventional age populations of equal size. We find that most models do not result in groups that are more biased than age; however, a small group of halos, found separately using K-Means and Gaussian Mixture Models, and various mass-loss methods were more biased than equal-sized age-based groups. The halos in these highly biased populations tended to be older, more concentrated, and were experiencing high tidal forces at z=0. We discuss the implications of data pre-processing and model hyper-parameters on groups obtained via machine learning methods.