Automated Distant Galaxy Merger Classifications from Space Telescope Images using the Illustris Simulation

TL;DR

This paper develops and tests machine learning methods, specifically random forests, to improve the observational classification of galaxy mergers over a wide redshift range using synthetic images from cosmological simulations.

Contribution

It introduces a novel application of redshift-dependent random forests for galaxy merger classification, outperforming traditional morphological indicators.

Findings

Random forests achieve about 70% completeness at 0.5 < z < 3.

Purity of merger detection increases from 10% at z=0.5 to 60% at z=3.

Estimated merger rates from RFs are roughly twice as high as theoretical expectations.

Abstract

We present image-based evolution of galaxy mergers from the Illustris cosmological simulation at 12 time-steps over 0.5 < z < 5. To do so, we created approximately one million synthetic deep Hubble Space Telescope and James Webb Space Telescope images and measured common morphological indicators. Using the merger tree, we assess methods to observationally select mergers with stellar mass ratios as low as 10:1 completing within +/- 250 Myr of the mock observation. We confirm that common one- or two-dimensional statistics select mergers so defined with low purity and completeness, leading to high statistical errors. As an alternative, we train redshift-dependent random forests (RFs) based on 5-10 inputs. Cross-validation shows the RFs yield superior, yet still imperfect, measurements of the late-stage merger fraction, and they select more mergers in bulge-dominated galaxies. When applied to CANDELS morphology catalogs, the RFs estimate a merger rate increasing to at least z = 3, albeit two times higher than expected by theory. This suggests possible mismatches in the feedback-determined morphologies, but affirms the basic understanding of galaxy merger evolution. The RFs achieve completeness of roughly 70% at 0.5 < z < 3, and purity increasing from 10% at z = 0.5 to 60% at z = 3. At earlier times, the training sets are insufficient, motivating larger simulations and smaller time sampling. By blending large surveys and large simulations, such machine learning techniques offer a promising opportunity to teach us the strengths and weaknesses of inferences about galaxy evolution.