Merged or monolithic? Using machine-learning to reconstruct the dynamical history of simulated star clusters

TL;DR

This study applies machine learning to simulated star clusters to distinguish between those formed by merging and those evolved in isolation, aiming to identify merger histories in observational data.

Contribution

It demonstrates that machine learning algorithms can effectively classify star cluster formation history from simulated data with high accuracy.

Findings

Support-vector machines outperform other classifiers.

Approximately 10% misclassification rate achieved.

Cluster concentration correlates with principal component of features.

Abstract

Context. Machine-Learning (ML) solves problems by learning patterns from data, with limited or no human guidance. In Astronomy, it is mainly applied to large observational datasets, e.g. for morphological galaxy classification. Aims. We apply ML to gravitational N-body simulations of star clusters that are either formed by merging two progenitors or evolved in isolation, planning to later identify Globular Clusters (GCs) that may have a history of merging from observational data. Methods. We create mock-observations from simulated GCs, from which we measure a set of parameters (also called features in the machine-learning field). After dimensionality reduction on the feature space, the resulting datapoints are fed to various classification algorithms. Using repeated random subsampling validation we check whether the groups identified by the algorithms correspond to the underlying physical distinction between mergers and monolithically evolved simulations. Results. The three algorithms we considered (C5.0 trees, k-nearest neighbour, and support-vector machines) all achieve a test misclassification rate of about 10% without parameter tuning, with support-vector machines slightly outperforming the others. The first principal component of feature space correlates with cluster concentration. If we exclude it from the regression, the performance of the algorithms is only slightly reduced.