Machine Learning for event reconstruction at the FCCee

by Dolores Garcia (CERN)

Wednesday 17 Jun 2026, 11:00 → 12:00 Europe/Zurich

503/1-001 - Council Chamber (CERN)

Precision measurements of Higgs couplings, electroweak parameters, and flavour observables at the Future Circular Collider (FCC-ee) place stringent demands on event reconstruction, where the achievable sensitivity scales directly with the resolution on visible final-state particles and their invariant masses. Current particle flow algorithms rely on detector-specific clustering and extensive manual tuning, limiting flexibility during the detector design phase. In this seminar, I will present HitPF, an end-to-end reconstruction algorithm that maps calorimeter hits and charged particle tracks directly to particle-level objects, bypassing intermediate clustering stages entirely. The method combines geometric algebra transformer networks with object condensation-based clustering, followed by dedicated networks for particle identification and energy regression. Benchmarked on fully simulated Z → qq̄ events at √s = 91 GeV in the FCC-ee CLD detector, HitPF outperforms the state-of-the-art PandoraPFA algorithm by 10–20% in reconstruction efficiency, achieves up to two orders of magnitude reduction in fake-particle rates for charged hadrons, and improves both visible energy and invariant mass resolution by 22%. By learning reconstruction directly from simulation, HitPF decouples performance from detector-specific tuning, enabling rapid iteration across detector concepts during the FCC-ee design phase. I will discuss the architecture, training strategy, key results, and outlook for extensions to other collision environments. 

Bio: Dolores Garcia is a CERN Fellow working at the intersection of experimental high-energy physics and machine learning. Her research focuses on end-to-end event reconstruction for future colliders using geometric deep learning. She is a leading developer of HitPF, a novel algorithm that reconstructs particles directly from raw detector hits using geometric algebra transformers and object condensation, bypassing conventional clustering stages. Benchmarked on fully simulated FCC-ee events, HitPF outperforms the state-of-the-art PandoraPFA algorithm in reconstruction efficiency, fake-rate suppression, and invariant mass resolution. She currently serves as high-level reconstruction convener for FCC. Her broader research interests lie in data-driven geometric processing of detector data and its application to the design and optimization of future collider experiments. She was awarded the FCC Week 2024 Innovation Award for her contributions to this programme.