PHAT-JeT: Efficient Particle Jet Tagging with Patch Hierarchical Attention

A novel transformer model, known as the Patch Hierarchical Attention Transformer (PHAT-JeT), has been developed for real-time jet tagging at the Large Hadron Collider. This innovative approach merges a geometry-based message-passing module that captures the local structure of the detector plane with a hierarchical attention mechanism that focuses on small groups of particles while maintaining a broader context through efficient methods. This solution effectively tackles the challenges of latency and accuracy in real-time trigger systems, where conventional transformers are often too resource-intensive due to their quadratic self-attention requirements. While existing efficient models lower computational demands, they compromise classification performance. PHAT-JeT seeks to eliminate this drawback, facilitating precise jet tagging within trigger constraints.