·
HGPflow — Hypergraph Particle Reconstruction
Particle reconstruction as an assignment problem. The network predicts what fraction of each calorimeter cluster's energy came from each particle, so energy is conserved by construction.
Key result
3×
fewer parameters than the ML benchmark
On most particle-, jet- and event-level metrics HGPflow outperforms both the standard parametrized algorithm and MLPF, the deep-learning benchmark, with about 2M parameters against MLPF's 6M.
Resources