Radio Transient Detection with Closure Products and Machine Learning

For transient sources with timescales of 1-100 seconds, standardized imaging for all observations at each time step become impossible as large modern interferometers produce significantly large data volumes in this observation time frame. Here we propose a method based on machine learning and using interferometric closure products as input features to detect transient source candidates directly from the spatial frequency domain without imaging. We train a simple neural network classifier on a synthetic dataset of Noise/Transient/RFI events, which we construct to tackle the lack of labelled data. We also use the hyper-parameter dropout rate of the model to allow the model to approximate Bayesian inference, and select the optimal dropout rate to match the posterior prediction to the actual underlying probability distribution of the detected events. The overall F1-score of the classifier on the simulated dataset is greater than 85\%, with the signal-to-noise at 7$\sigma$. The performance of the trained neural network with Monte Carlo dropout is evaluated on semi-real data, which includes a simulated transient source and real noise. This classifier accurately identifies the presence of transient signals in the detectable signal-to-noise levels (above 4$\sigma$) with the optimal variance. Our findings suggest that a feasible radio transient classifier can be built up with only simulated data for applying to the prediction of real observation, even in the absence of annotated real samples for the purpose of training.