Preliminary Results of a New Deep Learning Method to Detect and Localize GRBs in the AGILE/GRID Sky Maps

TL;DR

This paper introduces a novel deep learning approach using convolutional neural networks to detect and localize gamma-ray bursts in AGILE/GRID sky maps, demonstrating high accuracy and potential for real-time astrophysical event detection.

Contribution

The paper presents a new two-model deep learning framework for GRB detection and localization in gamma-ray sky maps, trained on simulated data, with plans for broader astrophysical applications.

Findings

Detection accuracy of 95.7%

Localization error below 0.8 degrees

Deployment on AWS for real-time analysis

Abstract

AGILE is an ASI space mission launched in 2007 to study X-ray and gamma-ray phenomena in the energy range from $\sim20$ keV to $\sim10$ GeV. The AGILE Team developed a real-time analysis pipeline for the fast detection of transient sources, and the follow-up of external science alerts received through networks such as the General Coordinates Network. We developed a new Deep Learning method for detecting and localizing Gamma-Ray Bursts (GRB) in the AGILE/GRID sky maps. We trained the model using sky maps with GRBs simulated in a radius of 20 degrees from the center of the map, which is larger than 99.5 \% of the error region present in the GRBWeb catalog. We also plan to apply this method to search for counterparts of gravitational wave events, which typically have a wider localization error region. The method comprises two Deep Learning models implemented with two Convolutional Neural Networks. The first model detects and filters sky maps containing a GRB, while the second model localizes its position. We trained and tested the models using simulated data. The detection model achieves an accuracy of 95.7 \%, and the localization model has a mean error lower than 0.8 degrees. We configured a Docker container with all the required software for data simulation and deployed it using the Amazon Web Service to calculate the p-value distribution under different conditions. With the p-value distribution, we can calculate the statistical significance of a detection.