Variable Star Classification with a Multiple-input Neural Network

In this experiment, we created a Multiple-Input Neural Network, consisting of convolutional and multilayer neural networks. With this setup the selected highest-performing neural network was able to distinguish variable stars based on the visual characteristics of their light curves, while taking also into account additional numerical information (e.g., period, reddening-free brightness) to differentiate visually similar light curves. The network was trained and tested on Optical Gravitational Lensing Experiment-III (OGLE-III) data using all OGLE-III observation fields, phase-folded light curves, and period data. The neural network yielded accuracies of 89%-99% for most of the main classes (Cepheids, delta Scutis, eclipsing binaries, RR Lyrae stars, Type-II Cepheids), only the first-overtone anomalous Cepheids had an accuracy of 45%. To counteract the large confusion between the first-overtone anomalous Cepheids and the RRab stars we added the reddening-free brightness as a new input and only stars from the LMC field were retained to have a fixed distance. With this change we improved the neural network's result for the first-overtone anomalous Cepheids to almost 80%. Overall, the Multiple-input Neural Network method developed by our team is a promising alternative to existing classification methods.