Automated Detection of Classical Novae with Neural Networks

The POINT-AGAPE collaboration surveyed M31 with the primary goal of optical detection of microlensing events, yet its data catalog is also a prime source of light curves of variable and transient objects, including classical novae (CNe). A reliable means of identification, combined with a thorough survey of the variable objects in M31, provides an excellent opportunity to locate and study an entire galactic population of CNe. This paper presents a set of 440 neural networks, working in 44 committees, designed specifically to identify fast CNe. The networks are developed using training sets consisting of simulated novae and POINT-AGAPE light curves in a novel variation on K-fold cross validation and use the binned, normalized power spectra of the light curves as input units. The networks successfully identify 9 of the 13 previously identified M31 CNe within their optimal working range (and 11 out of 13 if the network error bars are taken into account). The networks provide a catalogue of 19 new candidate fast CNe, of which four are strongly favored.