PACMANN: using machine learning to detect pulsars

Neutron stars form at the death of large stars. Sometimes these neutron stars rapidly rotate, and we call them pulsars. The magnetic field of the spinning neutron stars forms beams at the magnetic poles which emit radio frequencies. Because of the spin of pulsars the radio emissions arrive on earth as predictably timed pulses.

Pulsars are observed mainly using radio telescopes. We “fold” the data across many pulses in order to get a clearer picture of the pulsar in question. After folding, humans vet each pulsar candidate by observing its properties from a series of diagnostic plots. We primarily look at four plots. Persistence: how the pulsar persists over an observation, Subband: how well the pulsar emits over different radio frequencies, Pulse Profile: how the pulse itself actually appears, and DM: dispersion measure, which is a proxy for distance from Earth. This is a somewhat manual process, making the discovery of new pulsars slower than we would like.

We have developed a program: PulsAr Classifier Machine learning Algorithm with Neural Networks or PACMANN. PACMANN is a hierarchical machine learning algorithm that is trained on the same data structures seen by human rankers. PACMANN has been successful at identifying pulsars in a pool of candidates, making it a useful tool in the automation of pulsar identification. As a next step, we will be searching archival data from the GBNCC Survey performed with the Green Bank Telescope for new pulsar candidates.