Particle reconstruction

Interpreting data from collider experiments relies on pattern recognition algorithms that reconstruct the particles produced in each collision. The most popular approach today, known as particle flow, combines tracking and calorimeter information to optimize reconstruction performance, proving especially useful for dense environments of the LHC. Our group plays a leading role in developing the next generation of particle flow algorithms using modern deep learning architectures. In [1], we explored particle flow using computer vision techniques such as UNet, and compared them with graph-based models. This work was extended in [2], where we proposed a set-to-set transformer model and introduced HGPflow, an approach that learns hypergraph structure while incorporating energy conservation as an inductive bias. Most recently, in [3], we further advanced the HGPflow architecture to operate on full collision events, showing its applicability not only to the LHC but also to future facilities like CLIC.

[1] https://link.springer.com/article/10.1140/epjc/s10052-021-08897-0 
[2] https://link.springer.com/article/10.1140/epjc/s10052-023-11677-7 
[3] https://arxiv.org/abs/2410.23236