← Back to Work

The Optimal use of Segmentation for Sampling Calorimeters

Deep Neural Networks • Point Clouds • Detector Design • 2024
Graph Neural Networks for Optimizaton Calorimeter Designs

Full Paper Link

Visualization of calorimeter segmentation

Synopsis

By designing the reconstruction method and the physical detector together, our design optimizes for the best possible physical measurements. Our optimization strategy was so succesful, it's application saved $1M in detector material for the same reconsturction performance for the EPIC detector design at the Electron Ion Collider.

Research Goal: We demonstrate that for the energy estimation of isolated charged pion showers, relatively fine longitudinal segmentation is key to achieving an energy resolution better than 10% across the full phase space.

Abstract

To ensure that trends in detector design are due entirely to hardware and not to sub-optimal software, we deploy deep neural networks to perform energy reconstruction. We simulate a detector similar to the forward calorimeter system intended for the ePIC detector at the upcoming Electron Ion Collider (EIC). Our results provide a valuable benchmark for ongoing EIC detector optimizations and may inform future studies involving high-granularity calorimeters across various facilities.

Introduction

Calorimeters are critical components of particle and nuclear physics detectors. The reconstruction of energy flow requires custom algorithms because of the diversity of possible geometry and readout configurations. Classically, software is not fully optimized during the design phase, which can lead to suboptimal instrument design. This is particularly acute for the Electron Ion Collider (EIC), as its calorimeters must be finely segmented to deliver its science goals 1,2.

We propose employing machine learning (ML) to address this challenge. Deep neural networks can process low-level information within clusters of cells, often outperforming existing approaches and providing an approximation to optimal reconstruction 3-11. By using ML, hardware design can be guided by approximately optimal software utilization, and the development of software-based compensation for sampling calorimeters can be streamlined 12.

Our approach aims to capture the benefits of complex classical algorithms, such as those used by ATLAS 13 and CALICE 14, with a training latency that allows for rapid comparison of detector configurations. While most literature focuses on fixed instruments, recent studies have begun exploring the interplay between design and ML-based estimation 15,16. We focus on Graph Neural Networks (GNNs) 17 and DeepSets 18, which treat the calorimeter data as point clouds, a method proven powerful in recent experimental studies 19.

  1. A. Accardi et al., Eur. Phys. J. A 52 no. 9, (2016) 268. ↩
  2. R. Abdul Khalek et al., Nucl. Phys. A 1026 (2022) 122447. ↩
  3. L. de Oliveira et al., JHEP 01 (2019) 069. ↩
  4. CMS Collaboration, CERN-LHCC-2017-023. ↩
  5. ATLAS Collaboration, ATL-PHYS-PUB-2020-018. ↩
  6. C. Neubüser et al., Front. in Phys. 9 (2021) 715479. ↩
  7. N. Akchurin et al., JINST 16 no. 09, (2021) P09036. ↩
  8. N. Akchurin et al., Phys. Rev. D 105 no. 5, (2022) 052001. ↩
  9. ATLAS Collaboration, ATL-PHYS-PUB-2022-040. ↩
  10. S. R. Qasim et al., Eur. Phys. J. C 82 no. 8, (2022) 753. ↩
  11. J. Kieseler, Eur. Phys. J. C 82 no. 1, (2022) 79. ↩
  12. H. L. He et al., "Deep learning for calorimeter signal reconstruction," [arXiv:2103.00197]. ↩
  13. ATLAS Collaboration, Eur. Phys. J. C 77 no. 7, (2017) 490. ↩
  14. CALICE Collaboration, JINST 7 (2012) P09017. ↩
  15. N. Akchurin et al., [arXiv:2104.05608]. ↩
  16. C. Neubüser et al., [arXiv:2111.01188]. ↩
  17. J. Shlomi et al., Machine Learning: Science and Technology 2 no. 2, (2020) 021001. ↩
  18. M. Zaheer et al., "Deep Sets," NIPS'17 (2017) 3391–3401. ↩
  19. ATLAS Collaboration, ATL-PHYS-PUB-2022-040. ↩