Optimising Geometric Deep Learning for Varied Reconstruction Tasks in High Granularity Calorimeters

Abstract

In the recent years, high energy physics discoveries have been driven by the increasing of luminosity and/or detector granularity. This evolution gives access to bigger statistics and data samples, but can make it hard to process the detector outputs with current methods and algorithms. Graph Neural Networks (GNNs), have been shown to be powerful tools to address these challenges. While GNNs can deal with the non-euclidean nature of the data extracted from these detectors, their deployment presents significant difficulties, mainly from the computational complexity of the algorithms they are based on. We propose to use the known geometries of the detectors to optimise the algorithms used for GNN pipelines, which usually have quadratic algorithmic complexities. We validate the validity of our optimised approach on various tasks, such as particle identification, energy regression and instance segmentation. Finally, we present HIBOU, a generic modular architecture that provides a blueprint for how these GNN blocks are to be implemented within experiment-specific reconstruction frameworks.