2026 workshop program
Fast GRB Localization from Raw Compton Events with Physics-Guided Implicit Neural Representations
The work
Abstract
Compton telescopes localize gamma-ray bursts (GRBs) from sparse interaction-event lists rather than conventional images. Each event constrains the incident photon direction to a circle on the celestial sphere, making GRB localization a geometric inverse problem with an expensive sky-likelihood surface. We propose a physics-guided implicit neural representation (INR) for fast GRB localization in the NASA COSI mission setting. For each GRB, we construct an unsupervised cone-residual likelihood directly from raw reconstructed angular event triplets, fit a compact INR to exact likelihood samples, evaluate it as a continuous sky surrogate, and recover the final direction through exact local refinement. The neural network is therefore used to accelerate and interpolate a physically interpretable score surface, rather than to learn source labels. On 1,250 bright simulated GRBs, the method achieves mean and median angular deviations of $0.255\degunit$ and $0.290\degunit$, respectively, compared with $0.390\degunit$ and $0.380\degunit$ for the standard LM-MLEM analysis baseline.