Characterizing simulated galaxy stellar mass histories

TL;DR

This paper analyzes simulated galaxy stellar mass histories using PCA and clustering to understand their properties, compare different models, and connect simulation results with observable galaxy features.

Contribution

It introduces a PCA-based framework and clustering methods to classify and compare galaxy stellar mass histories across different simulation models.

Findings

One dominant fluctuation shape in galaxy histories across models.

Classification methods often yield similar groupings of histories.

PCA and clustering effectively characterize variations in galaxy formation simulations.

Abstract

Cosmological galaxy formation simulations can now produce rich and diverse ensembles of galaxy histories. These simulated galaxy histories, taken all together, provide an answer to the question "how do galaxies form?" for the models used to construct them. We characterize such galaxy history ensembles both to understand their properties and to identify points of comparison for histories within a given galaxy formation model or between different galaxy formation models and simulations. We focus primarily on stellar mass histories of galaxies with the same final stellar mass, for six final stellar mass values and for three different simulated galaxy formation models (a semi-analytic model built upon the dark matter Millennium simulation and two models from the hydrodynamical OverWhelmingly Large Simulations project). Using principal component analysis (PCA) to classify scatter around the average stellar mass history, we find that one fluctuation dominates for all sets of histories we consider, although its shape and contribution can vary between samples. We correlate the PCA characterization with several $z=0$ galaxy properties (to connect with survey observables) and also compare it to some other galaxy history properties. We then explore separating galaxy stellar mass histories into classes, using the largest PCA contribution, k-means clustering, and simple Gaussian mixture models. For three component models, these different methods often gave similar results. These history classification methods provide a succinct and often quick way to characterize changes in the full ensemble of histories of a simulated population as physical assumptions are varied, to compare histories of different simulated populations to each other, and to assess the relation of simulated histories to fixed time observations.