Uso de redes neurais para busca de transientes de rádio com o telescópio BINGO.

Abstract

In 2007, a very bright radio pulse was identified in archival data from the Parkes Telescope in Australia, marking the beginning of a new branch of research in astrophysics. In the early years, Fast Radio Bursts (FRBs) appeared very mysterious because the sample of events was limited, and their origins were unknown. With the improvement of instruments and data analysis techniques over the last five years, hundreds of new FRBs have been discovered. The field is now undergoing a revolution, and the understanding of FRBs has rapidly increased as new events continue to be uncovered. However, as new data are received at a high rate of candidates, it is necessary to assess whether they are FRBs or transients originating from other sources. In this work, we employ Deep Learning techniques to train deep neural networks for the classification of FRB and transient candidates. The convolutional neural networks used work with frequency and time data to generate spectrograms, also known as Waterfalls. We trained these networks using simulated FRBs and analyzed their output. We present several deep learning models with an accuracy and recall of approximately 90% on our test dataset, which is composed of simulated data and real pulses or FRB candidates. Currently, the use of machine learning algorithms for candidate classification is essential. These algorithms will also play a key role in building real-time triggers for detection in the BINGO project instruments.