Real time AI for particle physics measurements (Cristinziani)

Thursday 16 Jul 2026, 16:00 → 17:00 Europe/Berlin

What collider experiments can ultimately measure is often constrained by what can be reconstructed and selected in real time, within a fixed latency of a few microseconds. 

At Belle II this is especially true for dark-sector searches, low-multiplicity tau decays, and final states with few tracks or photons, where backgrounds overwhelm the signal and the trigger must decide what to keep before any data is stored.

The talk centers on three examples at Belle II. 

The first is multi-modal GNN-based track finding, originally developed to recover sensitivity to long-lived particles, whose displaced decays break the prompt-track assumptions of conventional pattern recognition. 

The second is real-time, GNN-based clustering in the electromagnetic calorimeter as it already operates today on FPGAs, within the trigger latency budget of 1.05 microseconds. 

Finally I will look ahead to a future calorimeter trigger upgrade with substantially more inputs, where the same class of architecture is a candidate for a harder clustering problem. 

Deploying any of this remains the bottleneck: I will describe the chain from trained model to inference on heterogeneous platforms such as AMD Versal, and the worldwide efforts, including the ErUM-Data DEEP project, to simplify a process that is still genuinely difficult.