End-to-end multi-particle reconstruction in high occupancy imaging calorimeters with graph neural networks

Abstract

We present an end-to-end reconstruction algorithm to build particle candidates from detector hits in next-generation granular calorimeters similar to that foreseen for the high-luminosity upgrade of the CMS detector. The algorithm exploits a distance-weighted graph neural network, trained with object condensation, a graph segmentation technique. Through a single-shot approach, the reconstruction task is paired with energy regression. We describe the reconstruction performance in terms of efficiency as well as in terms of energy resolution. In addition, we show the jet reconstruction performance of our method and discuss its inference computational cost. To our knowledge, this work is the first-ever example of single-shot calorimetric reconstruction of $${\mathcal {O}}(1000)$$ O ( 1000 ) particles in high-luminosity conditions with 200 pileup.

Study of residual artificial neural network for particle identification in the CEPC high-granularity calorimeter prototype

“…Therefore, there is growing interest in Graph Neural Network (GNN) models, which offer greater flexibility in handling data from detectors with irregular geometries. In this domain, pioneering models such as DGCNN [15] and GravNet [16,17] have been developed for either classification or clustering. In this paper, we propose novel PID methods based on the ResNet architecture [13], a much deeper architecture with millions of trainable parameters tailored for high-granularity calorimeters.…”

“…The efficacy of introducing Residual Connections is studied and the final performance is compared with the benchmark BDT [18], LeNet [19], and AlexNet [14] methods. Furthermore, to ensure the generality, an updated model adaptable to any irregular detector geometry, called the Dynamic Graph Residual Networks (DGRes), allowing our ResNet-based model to be applied beyond detectors with regular structures, has been proposed and compared with other GNN models, DGCNN [15] and GravNet [16,17]. The paper is structured as follows: section 2 introduces detector geometry and Monte Carlo simulation samples which are used to study the performance of various classifiers, section 3 briefly illustrates PID based on BDT and provides insights on shower topology; section 4 demonstrates the algorithm, the performance, and the complexity of our models; section 5 concludes this research.…”

“…The integration of GravNet has resulted in the formulation of an end-to-end reconstruction algorithm tailored for the single-shot calorimetric reconstruction of approximately 1000 particles. This is particularly crucial in high-luminosity conditions with 200 pileup events [17].…”

“…For DGCNN, we utilize an existing PyTorch implementation. In the case of GravNet, we reimplement the network architecture originally for the reconstruction task, as outlined in the paper [17],…”

Isothermal annealing of radiation defects in silicon bulk material of diodes from 8” silicon wafers

“…The radiation levels will increase by about a factor of 10 [2]. In particular in the forward region at pseudorapidities above 1.5, this calls for a detector design that can operate under these conditions and offers the possibility to use fine-grained particle flow and calorimetry algorithms, such that the primary interaction particles can be separated from the pileup interactions using information from all detector subsystems [3][4][5][6].…”