Application of neural networks to synchro-Compton blazar emission models

Jets from supermassive black holes in the centers of active galaxies are the most powerful persistent sources of electromagnetic radiation in the Universe. To infer the physical conditions in the otherwise out-of-reach regions of extragalactic jets we usually rely on fitting of their spectral energy distribution (SED). The calculation of radiative models for the jet non-thermal emission usually relies on numerical solvers of coupled partial differential equations. In this work machine learning is used to tackle the problem of high computational complexity in order to significantly reduce the SED model evaluation time, which is needed for SED fitting with Bayesian inference methods. We compute SEDs based on the synchrotron self-Compton model for blazar emission using the radiation code ATHE${\nu}$A, and use them to train Neural Networks exploring whether these can replace the original computational expensive code. We find that a Neural Network with Gated Recurrent Unit neurons can effectively replace the ATHE${\nu}$A leptonic code for this application, while it can be efficiently coupled with MCMC and nested sampling algorithms for fitting purposes. We demonstrate this through an application to simulated data sets and with an application to observational data. We offer this tool in the community through a public repository. We present a proof-of-concept application of neural networks to blazar science. This is the first step in a list of future applications involving hadronic processes and even larger parameter spaces.