Analysis of spectral energy distribution of blazar enabled by convolutional neural network

Modeling the multi-wavelength and multi-messenger emission from jets through time-dependent kinetic equations remains computationally prohibitive. I present a surrogate model designed to replace the expensive kinetic solver while preserving the underlying physical dependencies. The surrogate model takes the form of a convolutional neural network trained on radiative outputs of numerical simulations. It makes Bayesian parameter inference feasible for hundreds of spectral energy distributions (SEDs). I demonstrate its performance using data from observations of OJ 287 and 1ES 1959+650, showing how time-resolved SED fitting reveals the temporal evolution of key physical parameters and provides a systematic characterization of the emission processes driving their emission