Abstract

Contributed Talk - Splinter AGN   (MW-1250)

Generating Training Data for Machine Learning Approaches to Quasi Periodic Blazar Light Curves

Johannes Just, Emma Kun, Amal Saadallah, Julia Becker Tjus
Ruhr University Bochum, HUN-REN CSFK Konkoly Observatory, Ruhr University Bochum, Ruhr University Bochum

Approaches to Quasi Periodic Blazar Light Curves} Blazars are among the most powerful non-transient multi-messenger sources in the sky and could be the key to understanding various open science questions; from (super massive) black hole evolution and properties to cosmic ray acceleration. Their light curves can be interpreted as time series in a machine learning context, making them suitable for time series analysis and forecasting methods. However, modern deep learning models typically require large amounts of diverse training data to learn robust representations and generalize well across different variability patterns. Since it is crucial to have sufficient training data, and there are only on the order of tens of known quasi periodic blazar light curves, synthetic training data must be simulated. This talk presents a newly developed method to generate such synthetic training data based on the assumption of a precessing blazar jet and the resulting quasi periodic Doppler boosting. With the resulting data, it might be possible to train neural networks and predict future flares subsequently. This would allow to point telescopes in the direction of a source right when it flares and also find flares that are hidden in complex data like, for example, in neutrino telescope data.