SED modeling with neural networks trained on radiative simulations

AstroGenesis uses neural networks trained on large sets of numerical simulations generated with physically motivated radiative codes. The simulations define how each theoretical model behaves across its parameter space. Once trained, the networks reproduce the simulation results with high accuracy and are much faster to evaluate, which makes parameter estimation, model comparison and the exploration of physical scenarios practical within one session.

Model outputs are not treated as final answers. The agent reports parameter constraints and shows the degeneracies that limit what the data can say about the underlying physical processes.

A homogeneous one-zone synchrotron self-Compton (SSC) model was fitted to 141 points of PKS 2155-304 (z = 0.116, EBL absorption included) from Fermi-LAT, NuSTAR, Swift XRT and UVOT, and SMARTS, all taken between 9 and 23 August 2013. The best fit gives a Doppler factor of 23.0 ± 5.2, a magnetic field of about 29 mG, a region size of about 1.6×10¹⁷ cm and an electron index of 2.12 ± 0.22. The answer comes with a plot of the data, best fit and posterior samples, a PDF report and a best-fit CSV.

Theoretical models can be integrated when they are well tested, physically motivated and formulated for a single source class, so that they are applied and interpreted consistently. The current models cover blazars.