Deep Learning Using a Latent Attention-based Network for the HAWC Observatory

We present the results of a novel deep learning network applied to data from the HAWC Observatory to improve the reconstruction of high-energy gamma-ray events. HAWC consists of 300 large water Cherenkov detectors, each of which is instrumented with 4 photomultiplier tubes (PMTs) which collects the Cherenkov light produced by the extensive air showers initiated by cosmic rays or high energy gamma rays. The charge and timing information of PMTs triggered by a shower are used as inputs to the network. Our network uses the attention mechanism, made popular by the Transformer architecture used in e.g. ChatGPT, which allows a set of target arrays to be updated based on the information from a set of source arrays. It reduces the computational requirements by updating a latent vector (the target) representing the shower using the PMT outputs (the source), rather than the self-attention mechanism, which would have the PMT arrays attend to themselves. We show that this setup improves the gamma-cosmic ray separation, and the point spread function of HAWC to astrophysical sources, while achieving lower latency in the reconstruction process than traditional reconstruction methods.