Xinyue Sheng and Matt Nicholl
Current time-domain surveys already discover far more transients than can be followed up spectroscopically, and this mismatch will only become more severe with LSST. But understanding the physics of transients requires spectra, particularly spectra soon after explosion, to classify them and to study their composition and dynamics. To obtain these spectra we need to prioritise the most physically interesting transients (usually outliers and those belonging to rare classes) for follow-up, using only the first few transient alerts.
We have written the NEural Engine for Discovering Luminous Events (NEEDLE) to do just this. NEEDLE performs real time analysis of transient alerts using machine learning. Although the differences between the early light curves of transients can be very subtle, NEEDLE combines the transient alerts with information on the host galaxy to select for two rare classes with strong environmental preferences: superluminous supernovae (SLSNe) preferring dwarf galaxies, and tidal disruption events (TDEs) occurring in the centres of nucleated galaxies.
NEEDLE uses:
-
photometric information contained directly in the alert packets,
-
cutouts of the detection and reference images, and
-
host galaxy magnitudes from Pan-STARRS (Panoramic Survey Telescope and Rapid Response System).
Together, these probe the shape of the transient light curve, the morphology of the host and the transient location within it, and correlations between the brightness and colours of the transient and host. To process these multi-modal data, we first use a convolutional neural network to extract features from the images, before combining these with the photometric features in a fully-connected neural network. We trained the code on bright transients from the Zwicky Transient Facility (ZTF). Despite having only a few tens of examples of the rare classes, we achieve 70-80% completeness on an unseen test set. Our work has been published in MNRAS (NEural Engine for Discovering Luminous Events (NEEDLE): identifying rare transient candidates in real time from host galaxy images - ADS).
For the past year we have been running NEEDLE on real-time alerts from ZTF, first filtering the alert stream using Lasair. We automatically and publicly report the probabilities of an alert belonging to a SLSN, TDE or normal supernova back to Lasair using the NEEDLE annotator (Annotator: NEEDLE). We highlight the best candidates to the community each week via Transient Name Server Astronotes.
Our real-time performance has matched expectations, correctly predicting 11 SLSNe and 10 TDEs so far.
While our completeness is very good, the purity of our stream remains modest due to the large class imbalance. Although the goal of NEEDLE is to find good candidates for spectroscopic classification, rather than to select pure photometric samples, we are now working to improve the purity by employing novel data augmentation strategies during training.
With LSST alerts coming soon, our next step will be to train NEEDLE using LSST six-band light curves. With the slower cadence in any one filter, compared to the ZTF alerts, the challenge will be to correctly infer the change in luminosity from the early part of the light curve. However, the availability of LSST image pairs within 30 minutes should provide a robust measure of the colour, and the inclusion of u-band photometry will be enormously helpful for selecting very blue transients like SLSNe and TDEs. The deep, high-quality host images should also improve our ability to exploit the galaxy morphology.