Self-supervised ConvLSTM for Fermi Large Area Telescope transient detection
A self-supervised deep learning model detects flares and gamma-ray bursts in sky maps from the Fermi-LAT telescope.
Abstract
A study of a framework for detecting transient gamma-ray phenomena (such as flares or gamma-ray bursts) based on self-supervised spatio-temporal deep learning, applied to data from the Fermi-LAT space telescope. By simulating a decade of synthetic sky observations with gtobssim, the researchers trained a ConvLSTM network to reconstruct the expected emission; deviations from this baseline are identified through residual maps and per-pixel statistical thresholds, with a spatial filter that reduces isolated false positives. Applied to real daily Fermi-LAT maps, the model flags localized, time-varying excesses consistent with highly variable sources or transient events, providing a benchmark for anomaly-detection strategies on long-duration datasets.
Authors
A. Garinei, S. Speziali, A. Vispa, A. Marini, S. Cutini, E. Piccioni, M. Marconi, F. Longo, M. Martini, F. Fallucchi, R. Giuliano, E. W. De Luca, U. Di Matteo, S. Meola
Link
Publication date: 06/05/2026 Last updated: 02/10/2026 17:20