Speakers
Description
Modern calorimeters for High Energy Physics (HEP) have very fine transverse and longitudinal segmentation to manage high incoming flux and improve particle identification capabilities. Compared to older calorimeter designs, this change alone alters the extraction of the number and energy of incident particles on the device from a simple gaussian-template clustering problem to a highly non-gaussian multidimensional clustering problem on a sparse point cloud. To face this daunting challenge, over the past 3 years we have developed new neural network architectures, refined point cloud search techniques, and a deep understanding of complex simulations of particle-material interactions in the context of calorimetry in HEP. In this paper we will detail the development and training of differentiable reconstruction algorithms for these kinds of calorimeters and demonstrate preliminary physics performance in the specific case of the high-granularity CMS Phase 2 Endcap Calorimeter. In addition to their physics performance, these algorithms are automatically portable to coprocessors, in contrast to hand-written algorithms, and therefore can take advantage of heterogeneous resources to accelerate inference. Finally, we will discuss the deployment of these algorithms using as-a-service strategies, including spillover to cloud infrastructure to handle increased demand.
| Speaker time zone | No preference |
|---|
