Researchers developed a machine-learning filter that flags and removes unwanted energy depositions in the Belle II electromagnetic calorimeter before the detector groups those deposits into particle “clusters.” The calorimeter is made of about 8,700 cesium‑iodide crystals and measures energies from photons and hadrons. At the SuperKEKB collider’s very high beam intensities, ordinary beam-related noise and the messy way hadrons interact in the crystals have produced many small local energy peaks. Those peaks can be mistaken for photons or split real signals into pieces, hurting energy and position measurements and increasing computing and storage needs.

To tackle this, the team trained a Graph Neural Network (GNN) to decide, for each small local maximum of energy, whether it is a true signal or an unwanted deposition. They used the Belle II simulation chain (the basf2 software and Geant4) with recorded beam-background overlays to label examples. Based on the simulation, each local maximum was labeled as signal, beam background, a duplicate, or a “split‑off” (a secondary particle created inside the detector that can make a separate deposit). The labels let the network learn what real and unwanted peaks look like in different detector regions.

The GNN sees only the crystals that actually registered energy. Each crystal becomes a node with simple inputs such as measured energy, pulse-shape fit results, and timing. Small neighborhoods around each local maximum (roughly a 9×9 crystal area) form the graphs. The network uses message‑passing layers and a modified GravNet-style block to share information between nearby crystals and produce a classification score for each local maximum. The team trained on simulated collision samples (20,000 Υ(4S)→BB events and 10,000 e+e−→ϕ→K0LK0S γ events), used mean-squared-error loss, and stabilized training with Stochastic Weight Averaging.