The 35th annual conference on Astronomical Data Analysis Software & Systems 2025 in Görlitz, Germany

The Compton Spectrometer and Imager (COSI) is a NASA satellite mission under development that will survey the entire sky in the 0.2–5 MeV range with a wide-field gamma-ray telescope. Its main instrument consists of a germanium detector array, surrounded on the sides and bottom by bismuth germanate (BGO) scintillator active shields (the Anticoincidence Subsystem, ACS). The ACS both suppresses and monitors background events and enables the detection of transient sources. COSI will include an onboard triggering algorithm capable of identifying Gamma-Ray Bursts (GRBs) in the ACS and transmitting data to the ground for further analysis. An automated pipeline will analyze the data to localize the GRBs and share the position with the community.
In this work, we present a short GRB localization method based on ACS data using Deep Learning (DL) techniques. Different network architectures were evaluated to estimate localization uncertainties at the 90% confidence level, including cases where the confidence region is split into multiple areas.
The first approach employs a probabilistic regressor that outputs probability distributions for the sine and cosine of the spherical angles defining the GRB position. From these distributions we derive elliptical 90% confidence regions in the tangent plane. This method is limited when the confidence region is multimodal. To address this, we developed a second approach based on a neural network classifier that predicts the GRB location on a HEALPix grid (nside=32). Training labels are smoothed over neighboring pixels to improve performance, and the Softmax output provides a probability distribution across the sky map. The 90% confidence regions are then derived from this distribution.
Future work will compare this DL-based localization approach with classical methods such as χ² fitting and Maximum Likelihood Estimation (MLE). Traditional methods explicitly fit the background, whereas the DL model implicitly learns the impact of background during training.