A machine learning kilonova classifier

One of the most significant challenges in conducting successful MMA campaigns is candidate vetting. Large field-of-view telescopes such as Pan-STARRS, ATLAS, ZTF, and GOTO were able to rapidly identify kilonova (KN) candidates during the IGWN’s third and fourth observing runs (still ongoing). However, shock breakout supernovae, gamma-ray burst afterglows, and other rapidly-evolving transients can easily be mistaken for KN counterparts during their early light curve evolution. In the LSST era, this issue will be exacerbated due to the increased volume probed and the sparse light curve sampling. Here, we present preliminary results from a KN classifier built with Long Short Term Memory (LSTM) architecture, to distinguish between KNe and contaminants. Our simulated dataset consists of KN models and contaminant light curves sampled at ZTF’s typical cadence and in its target-of-opportunity mode. We add in contextual information from the GW skymaps through a convolutional neural network (CNN). Our classifier will output a score encoding the probability of a candidate being a KN. Such KN classifiers will be tantamount to identifying the most promising counterpart candidates during targeted GW campaigns in the future.