Deep learning predictions of galaxy merger stage and the importance of observational realism

Abstract

Machine learning is becoming a popular tool to quantify galaxy morphologies and identify mergers. However, this technique relies on using an appropriate set of training data to be successful. By combining hydrodynamical simulations, synthetic observations and convolutional neural networks (CNNs), we quantitatively assess how realistic simulated galaxy images must be in order to reliably classify mergers. Specifically, we compare the performance of CNNs trained with two types of galaxy images, stellar maps and dust-inclusive radiatively transferred images, each with three levels of observational realism: (1) no observational effects (idealized images), (2) realistic sky and point spread function (semi-realistic images), (3) insertion into a real sky image (fully realistic images). We find that networks trained on either idealized or semi-real images have poor performance when applied to survey-realistic images. In contrast, networks trained on fully realistic images achieve 87.1% classification performance. Importantly, the level of realism in the training images is much more important than whether the images included radiative transfer, or simply used the stellar maps (87.1% compared to 79.6% accuracy, respectively). Therefore, one can avoid the large computational and storage cost of running radiative transfer with a relatively modest compromise in classification performance. Making photometry-based networks insensitive to colour incurs a very mild penalty to performance with survey-realistic data (86.0% with ronly compared to 87.1% with gri). This result demonstrates that while colour can be exploited by colour-sensitive networks, it is not necessary to achieve high accuracy and so can be avoided if desired. We provide the public release of our statistical observational realism suite, RealSim, as a companion to this paper.

Towards robust determination of non-parametric morphologies in marginal astronomical data: resolving uncertainties with cosmological hydrodynamical simulations

“…Although there is no mathematical reason that such indicators could not be applied to any 2-dimensional distribution, the limitations of realistic data may impose practical impediments. For this reason, it has become a common approach to add observational realism to simulated data, in order to fairly compare derived properties (Bottrell et al 2017a,b;Huertas-Company et al 2019;Bottrell et al 2019b;Zanisi et al 2020;Ferreira et al 2020). Our approach is therefore to use simulations of galaxies, for which we can measure 'true' values of a given morphology metric, before adding noise (e.g.…”

“…The first is through crowd-sourcing, whereby the power of the human brain continues to be tapped, through the contributions of citizen scientists (Darg et al 2010;Lintott et al 2011;Casteels et al 2014;Simmons et al 2017;Willett et al 2017). Recently, artificial intelligence is replacing humans and machine learning algorithms are increasingly being applied to the challenge of large imaging datasets, either for general morphological classification (Huertas-Company et al 2015;Domínguez Sánchez et al 2019;Cheng et al 2020;Walmsley et al 2021) or the identification of particular galaxy types/features (Bottrell et al 2019b;Pearson et al 2019;Ferreira et al 2020;Bickley et al 2021). An alternative automated approach, which has been in use for several decades, is to compute some metric of the galaxy's light distribution, a technique which is readily applicable to large datasets.…”

Galaxy mergers in eagle do not induce a significant amount of black hole growth yet do increase the rate of luminous AGN

“…However observational works must ascertain the merging state of a galaxy from only an instantaneous (often pre-coalescence) snapshot. Thus a truly fair comparison would require us to apply observational techniques to synthetic images to estimate a galaxy's current merging state (similar to Lahén et al 2018;Bottrell et al 2019;Snyder et al 2019, for example), but this is beyond the scope of this study. We can, however, qualitatively compare our results to the observational studies.…”

A deep learning approach to test the small-scale galaxy morphology and its relationship with star formation activity in hydrodynamical simulations

“…For more details about RealSim, we refer the reader to the original papers. Bottrell et al (2019) have shown that the including the correct level of realism in mock observations is crucial when using neural networks for classification tasks.…”