Clasificación de eventos astronómicos transitorios con redes neuronales recurrentes

Abstract

In this master thesis, an innovative methodology based on recurrent neural networks (RNN) is introduced to address the accurate detection and classification of transient astronomical events from real observational data. Based on the inherent ability of RNNs to model data sequences and adapt to the temporal variability of events in space, the approach employs the MANTRA dataset, covering a diversity of events such as supernovae and active galaxy nuclei. Two preprocessing architectures are developed and gated recurrent units (GRUs) are implemented in a specialized recurrent neural network. Experimental results reveal a remarkable performance improvement, with up to 17% increase in accuracy, highlighting the training efficiency compared to conventional machine learning approaches. This advance contributes significantly to the automation of astronomical event classification, facilitating the early detection of astrophysical phenomena and enriching our understanding of the ever-changing universe.