ComptonUNet: A Deep Learning Model for GRB Localization with Compton Cameras under Noisy and Low-Statistic Conditions

TL;DR

ComptonUNet is proposed, a hybrid deep learning framework that jointly processes raw data and reconstructs images for robust GRB localization and significantly outperforms existing approaches, achieving improved localization accuracy across a wide range of low-statistic and high-background scenarios.

Abstract

Gamma-ray bursts (GRBs) are among the most energetic transient phenomena in the universe and serve as powerful probes for high-energy astrophysical processes. In particular, faint GRBs originating from a distant universe may provide unique insights into the early stages of star formation. However, detecting and localizing such weak sources remains challenging owing to low photon statistics and substantial background noise. Although recent machine learning models address individual aspects of these challenges, they often struggle to balance the trade-off between statistical robustness and noise suppression. Consequently, we propose ComptonUNet, a hybrid deep learning framework that jointly processes raw data and reconstructs images for robust GRB localization. ComptonUNet was designed to operate effectively under conditions of limited photon statistics and strong background contamination by combining the statistical efficiency of direct reconstruction models with the denoising capabilities of image-based architectures. We perform realistic simulations of GRB-like events embedded in background environments representative of low-Earth orbit missions to evaluate the performance of ComptonUNet. Our results demonstrate that ComptonUNet significantly outperforms existing approaches, achieving improved localization accuracy across a wide range of low-statistic and high-background scenarios.

Figures (6)

• Figure 1: Detailed configuration of the CC-BOX (Compton Camera Box) onboard the INSPIRE satellite kataoka2024inspire. The figure illustrates the front Ce:GAGG pixel array, side and bottom BGO scintillator active shields, pinhole structure, and multilayer detector arrangement. This design enables high-sensitivity and high-angular-resolution observations over a wide energy range (30 keV–3 MeV).
• Figure 2: Overview of (a) Unet, (b) ComptonNet, and (c) ComptonUNet architectures. Unet processes reconstructed images, while ComptonNet directly estimates gamma-ray directions from raw data. ComptonUNet combines both approaches by transforming raw data into feature maps using a ComptonNet-inspired encoder, followed by Unet-style decoder to estimate the gamma-ray directions. This design improves robustness under low-statistics and high-background conditions, making it suitable for transient gamma-ray observations.
• Figure 3: Summary of quantitative performance for the three models: (a) MSE, (b) SSIM, and (c) peak offset. ComptonUNet outperforms both Unet and ComptonNet across all metrics, by combining the strengths of both architectures.
• Figure 4: Visual comparison of reconstructed images from the three models. From top to bottom: Ground truth, BP image of Compton mode (BP Compton), BP image of pinhole mode (BP Pinhole), Unet output, ComptonNet output, and ComptonUNet output. Additionally, from left to right, the images represent different GRB durations (1, 3, 10, 30, and 100 s). The ComptonUNet model demonstrates superior performance in accurately reconstructing the source morphology and peak location compared to the other models.
• Figure 5: Noise robustness evaluation of the three models under eliminated background conditions. The dashed lines and simple lines represent the performance of each model under background conditions (as shown in Figure \ref{['f2']}) and eliminated background conditions, respectively. ComptonUNet maintains its performance even in the presence of background noise, while ComptonNet shows significant degradation.