Microsecond Latency Graph Neural Network Inference on Point Clouds at HAICON 2026
In his presentation, Marc Neu highlighted new approaches for applying artificial intelligence to latency-critical particle physics applications
PhD student Marc Neu attended the Helmholtz AI Conference HAICON 2026 in Munich, Germany, held from 8-11 June 2026, to present an overview of his work on real-time graph neural network inference to the wider Helmholtz AI community. His talk on "Microsecond Latency Graph Neural Network Inference on Point Clouds," given on 11 June, reflects collaborative work between the Institute for Information Processing Technologies (ITIV) and the Institute of Experimental Particle Physics (ETP) at KIT, jointly supervised by Prof. Dr.-Ing. Jürgen Becker (ITIV) and Prof. Dr. Torben Ferber (ETP). With more than 600 participants from 21 countries, the conference offered a chance to exchange ideas with researchers from across the Helmholtz Association and beyond.
Graph neural networks are well suited to the sparse, irregularly structured data produced by particle detectors, but deploying them in hardware triggers is challenging due to strict latency and throughput constraints. State-of-the-art collider triggers impose hard deadlines on the order of 1 to 10 microseconds, requiring custom machine learning accelerators based on field programmable gate arrays. In his talk, Marc gave an overview of a deployment methodology for mapping graph neural networks onto such platforms, demonstrated using a clustering algorithm for the electromagnetic calorimeter of the Belle II experiment.