Search for: All records

The neutral hydrogen (HI) in galaxies provides the gas reservoir out of which stars are formed. The ability to determine the HI masses for statistically significant samples of galaxies can provide information about the connection between this gas reservoir and the star formation that drives galaxy evolution. However, there are relatively few galaxies for which HI masses are known because these measurements are significantly more difficult to make than optical observations. Artificial neural networks are a type of nonlinear technique that have been used estimate the gas masses from their optical properties (Teimoorinia et al. 2017). We present HI observations of 51 galaxies with gas and stellar properties that are rare in the Arecibo Legacy Fast ALFA Survey (ALFALFA, Haynes et al. 2018) which was used to train the Artificial Neural Network developed by Teimoorinia et al. (ANN, 2017). These sources provide a test of the Artificial Neural Network predictions of HI mass and include some rare and interesting systems including galaxies that are extremely massive in both stellar mass (log M_∗> 11.0) and HI mass (log M_HI> 10.2) with large HI line widths (w_50> 500 km/s). We find that this Artificial Neural Network systematically overestimates the gas fraction of the galaxies in our selected sample, suggesting that care must be taken when using these techniques to predict gas masses for galaxies from a broad range of optical properties. 

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 (semirealistic images), and (3) insertion into a real sky image (fully realistic images). We find that networks trained on either idealized or semireal images have poor performance when applied to survey-realistic images. In contrast, networks trained on fully realistic images achieve 87.1 per cent 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 ($87.1{{\ \rm per\ cent}}$ compared to $79.6{{\ \rm per\ cent}}$ 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 ($86.0{{\ \rm per\ cent}}$ with r-only compared to $87.1{{\ \rm per\ cent}}$ 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.