R&D of the EM Calorimeter Energy Calibration with Machine Learning based on the low-level features of the Cluster

Suzuna Morimasa, M. Iwasaki, Taikan Suehara, Junichi Tanaka, Masahiko Saito, Hajime Nagahara, Yuta Nakashima, Noriko Takemura, T. NakanoEPJ 2024 · DOI

Abstract

We have developed an energy calibration method using machine learning for the International Linear Collider (ILC) electromagnetic (EM) calorimeter (ECAL), a sampling calorimeter consisting of Silicon-Tungsten layers. In this method, we use a deep neural network (DNN) for a regression to determine the energy of incident EM particles, improving the energy calibration resolution of the ECAL. The DNN architecture takes cluster hit data as low-level features of the cluster. In this paper, we report the status of our R&D and present results on energy calibration accuracy.

Introduction: The Calibration Challenge

In high-energy physics, accurately measuring the energy of particles like electrons and photons is vital for discovery. The ILC’s Electromagnetic Calorimeter (ECAL) is a "sampling calorimeter" made of alternating layers of Tungsten (to shower the particles) and Silicon (to detect them).

Traditionally, energy is calibrated using a "simple reconstruction" method where detected hits are summed and multiplied by a single coefficient. However, this approach suffers from non-linearity at high energies due to shower leakage, particle-type dependence, and significant angular ($\phi$) dependence. The SiD detector design uses twelve trapezoidal modules that overlap for mechanical stability, creating "projective gaps" and non-uniform responses depending on the particle's entry angle.

Cross-section of the SiD ECAL showing overlapping trapezoidal modules and the resulting projective gaps.

Methodology: Leveraging Low-Level Features

To overcome these limitations, we moved away from "high-level features" (such as center-of-mass position) and utilized low-level features. We employed a Multi-Layer Perceptron (MLP) to treat energy calibration as a regression problem.

The model's input is the raw data from every "hit" in a particle cluster. Each hit is defined by five parameters: hit energy, spatial coordinates (x, y, z), and the specific detector layer number. By providing the exact spatial and layer information, the model "learns" the internal structure of the particle shower, allowing it to automatically compensate for the detector's complex geometry.

MLP architecture using low-level hit features (energy, coordinates, and layer) as input.

Because the number of hits varies per cluster, we used a fixed input size of $5 \times 1900$ with 0-padding for smaller clusters. The final architecture is a 4-layer MLP with 100 nodes in the hidden layer, implemented using the PyTorch framework.

Datasets and Simulation

The study utilized Monte Carlo (MC) simulation data for the SiD ECAL design, focusing on single electrons and photons in the energy range of 2 to 40 GeV. Training was conducted on separate electron and photon datasets, as well as a mixed dataset to test robustness.

Key Results

The DNN-based approach demonstrated significant improvements over conventional methods and previous high-level machine learning models:

  • Energy Resolution: We achieved a resolution of 16–18% for electrons and 16–22% for photons. This is remarkably close to the SiD ECAL design goal of 17% and a notable improvement over high-level feature models (which achieved 22–25%).
  • Angular Stability: The $\phi$ distribution showed no angular dependence. The model successfully "ignored" the mechanical overlaps of the detector modules, providing a uniform response regardless of entry angle.
  • Mixed Performance: Accuracy remained robust at approximately 18–21% when trained on a mix of electrons and photons.

Conclusion and Future Work

Our research demonstrates that low-level hit data is the essential component for achieving uniform energy resolution in complex detector geometries. While the MLP performed well, the fixed-size vector input is rigid for sparse detector data. Future work will explore Graph Neural Networks (GNNs), which are naturally suited for "point cloud" style data, and extend the method to the ILD (Interposed Large Detector) design and more complex multi-particle environments.

Summary updated

Abstract

We have developed an energy calibration method using machine learning for the ILC electromagnetic (EM) calorimeter (ECAL), a sampling calorimeter consisting of Silicon-Tungsten layers. In this method, we use a deep neural network (DNN) for a regression to determine the energy of incident EM particles, improving the energy calibration resolution of the ECAL. The DNN architecture takes cluster hit data as low-level features of the cluster. In this paper, we report the status of our R&D and present results on energy calibration accuracy.
BibTeX
@inproceedings{morimasa2024r,
  author = {Suzuna Morimasa and M. Iwasaki and Taikan Suehara and Junichi Tanaka and Masahiko Saito and Hajime Nagahara and Yuta Nakashima and Noriko Takemura and T. Nakano},
  title = {{R\&D of the EM Calorimeter Energy Calibration with Machine Learning based on the low-level features of the Cluster}},
  booktitle = {EPJ},
  year = {2024},
  doi = {10.1051/epjconf/202431503012},
}
Figure