Particle physicists built a new neural network called PHAT-JeT to identify short-lived particle decays inside the flood of data produced at the Large Hadron Collider (LHC). The model is designed for the real-time trigger systems that must pick which collisions to save. Those systems face extreme limits: the LHC produces about 40 million collisions per second, raw data rates near 1 petabyte per second, and trigger decisions must be made in roughly 10 microseconds.

The paper introduces two ideas that work together. First is a geometric message-passing module. Message passing means nearby particles share information so the model learns local structure on the detector plane (angles often called η and ϕ). Second is a hierarchical patch-based attention scheme. Attention is a neural-network operation that normally compares every particle pair and becomes very costly as the number of particles grows. PHAT-JeT avoids that cost by grouping particles into small patches and computing exact pairwise attention inside each patch, while using a lightweight mechanism to communicate information between patches.

At a high level this keeps the best parts of full attention—fine-grained interactions among nearby particles—without the full quadratic cost of comparing every particle to every other particle. Jets are treated as unordered collections of particles, sometimes called point clouds, and the architecture encodes geometric priors that are natural for jet data. The authors compare PHAT-JeT to other efficient attention methods. Many competing approaches reduce cost by compressing pairwise interactions or using random hashing; those moves can degrade the model’s ability to capture detailed relationships between particles. The PHAT-JeT design preserves exact intra-patch interactions while keeping global context cheaply.

Why this matters: better and faster jet tagging at the trigger level can raise the fraction of interesting collision events that are recorded for later analysis. In the resource-constrained regime relevant for field-programmable gate array (FPGA) hardware used in triggers, PHAT-JeT achieves state-of-the-art accuracy and background rejection among deployable models on four community benchmarks: HLS4ML, JetClass, Top Tagging, and Quark–Gluon. The paper also reports ablation studies and FLOP-matched comparisons that support the design choices. The authors make their code available on GitHub.