A deep learning approach towards identifying spiral structure in galaxies

The interest in applying machine learning to answer astrophysical problems has risen dramatically in recent years, driven in large part by a desire to classify systems for more detailed studies of their intrinsic properties. One of the long outstanding problems in astronomy has been the origin of spiral structure in galaxies. While numerous theories have been proposed over the years, none has been convincingly proven. Machine learning holds considerable promise to resolve the long standing spiral problem by providing a means to model spiral arm properties in a much more detailed and automated way than has currently been possible, allowing for more direct comparisons between observations and theory over a large sample of galaxies. In this thesis I take a deep learning approach towards building a tool for automating the identification of spiral arms in galaxies through the use of a recurrent convolutional neural network. Compared to most deep learning tasks, identifying spiral arms pose a considerable problem due to their long, winding nature. To aid the network we utilize well known physical insights to spatially transform the image to make it easier for segmenting the individual arms. Given an image of a galaxy, the network outputs a series of spiral arm masks with an associated probability that the mask represents a spiral arm. Tests on synthetic galaxies reveals that we are able to reliably detect spiral arms with a precision approaching 100% whose average signal-to-noise ratio is greater than S/N > 1/3 and whose arm contrast is greater than 10% of the underlying galaxy flux.