Pulsars or AGNs? Unmasking the Fermi-LAT Sky with a Blind 1D Deep Learning Search

by Cristian Pozo González (IAA-CSIC)

Wednesday Sep 23, 2026, 12:00 PM → 1:30 PM Europe/Madrid

04.215.0 - Seminario 215 (Facultad de Ciencias Físicas)

The Fermi Large Area Telescope (LAT) has revolutionized our view of the high-energy universe, but it has left us with a colossal mystery: roughly one-third of the sources in its latest catalog (4FGL) are astrophysical "ghosts." They emit gamma rays, yet their true identity remains entirely unknown.
Until now, traditional Machine Learning approaches have tried to classify these sources by treating their spectral features as simple, independent tabular data. This approach ignores the vital topological information hidden within their Spectral Energy Distribution (SED). Furthermore, they often take a shortcut by including spatial coordinates in the training process, which inevitably contaminates the results with location biases.
In this seminar, we will present a radically different approach. We have designed TabularResCNN, a 1D convolutional neural network (1D-CNN) that tackles the problem from a fresh perspective. It "reads" the spectral data as sequential signals to capture the intrinsic curvature and shape of the emission, and it does so completely "blind" to galactic coordinates. Powered by a hierarchical and interpretable Deep Learning architecture (supported by Grad-CAM techniques), the model successfully separates Active Galactic Nuclei (AGNs) from Pulsars with ~98% accuracy, before moving on to distinguish between Young and Millisecond Pulsars.
The result? By applying this framework to over 2,500 unassociated sources, we have unearthed more than 200 high-confidence pulsar candidates, representing a potential 60% increase in the known population. We will demonstrate the undeniable robustness of these predictions by showing how the model autonomously "rediscovered" 5 out of 5 pulsars recently confirmed by the FAST radio telescope. Finally, we will discuss how this new catalog of candidates is set to become a treasure map for future observations with next-generation facilities like SKAO, FAST, and CTAO.