Speaker
Description
This work explores the application of machine learning techniques to classifying active galactic nuclei (AGN) with jets based on Very-Long-Baseline Interferometry (VLBI) observations at frequencies 1–90 GHz. Building upon previous work by Fanaroff and Riley, who classified relativistic jets in radio galaxies on kiloparsec scales, we extend this classification to parsec scales, closer to the central supermassive black hole. This approach enables detailed study of jet spatial structures and can help enhance accuracy in global positioning systems. We define two morphological classes: compact point-like sources and sources exhibiting jet structure. To construct a training dataset, synthetic AGN jet images were generated using ray-tracing of the analytical Blandford–Königl jet model, followed by uv-plane sampling and image reconstruction with the CLEAN algorithm to reproduce realistic observational effects. These simulated data were used to train a convolutional neural network (CNN), which demonstrated strong quantitative performance when evaluated on a manually labeled sample of real VLBI images. The trained model was subsequently applied to classify all (~130,000) AGN images from the Astrogeo database. These resulting classifications were later used to investigate the correlation between the morphological class, radio flux density and total radio luminosity. In addition, a public online tool for searching morphologically similar AGN images is currently under development, with the presented CNN model forming its core component.