Performance Modeling and Improvements on the GRB Source Localization Streaming Pipeline Aboard the Antarctic Demonstrator for the Advanced Particle-Astrophysics Telescope (ADAPT)

The Advanced Particle-astrophysics Telescope (APT) is a mission concept for a space-based gamma-ray telescope whose capabilities include prompt localization of gamma-ray bursts (GRBs) to support multi-wavelength and multi-messenger astrophysics. We describe the current state of our GRB localization pipeline aboard APT's balloon-borne prototype, the Antarctic Demonstrator for APT (ADAPT). We recently extended the pipeline using two neural networks: a classification network that filters out events arising from atmospheric background radiation, and a regression network that quantifies the uncertainty in the angular radius of each reconstructed Compton ring that constrains the GRB's source direction. With these improvements, we expect ADAPT to localize short GRBs of fluence $1$ MeV/cm$^2$ over one second to within $\sim$5$^\circ$ at least 68% of the time, for incident angles up to 80$^\circ$ from normal. This localization is achieved in $\sim$220 ms on a low-power quad-core Intel processor planned to fly with ADAPT. To produce real-time alerts that will optimally direct other, fast-slewing telescopes toward optical counterparts of short-duration GRBs, our pipeline is designed to continuously update pointing information even as data from the GRB continues to accumulate. Indeed, providing an estimated location quickly enables an instrument to begin slewing toward the source earlier, while later estimates using additional photons serve to refine its trajectory. To enable analysis to proceed concurrently with GRB photon arrival, we implement the computation as a streaming pipeline of concurrently running compute kernels that processes a stream of gamma-ray photons over time. Queues between kernels hold the output of one kernel until it can be consumed by the next. To understand the performance of this design, we model the performance of its two main compute kernels -- reconstruction and localization -- including both their service times with respect to data arrival characteristics and their queuing behaviors. Using this performance model, we calculate the pipeline's end-to-end latency to produce a localization result and identify bottlenecks that may limit its performance. We also characterize the pipeline's latency and accuracy when producing approximate localization results after seeing only part of the GRB's stream of photons. The ability to produce such intermediate results would allow the real-time ADAPT flight software to construct and report a series of increasingly accurate localizations for a GRB over time.