Using ML Methods for SoLID Beam Test Analysis

For most experiments in particle physics, some form of particle identification is necessary. Existing industry standard methods of particle identification are currently based on performing cuts around certain particle properties, such as momentum or Cherenkov radiation. Data analysis utilizing other pieces of data gleaned from the sensors used in a particular experimental set-up is also useful, but one runs into the problem of diminishing returns quickly. In the interest of maximizing efficiency, our project investigated the usage of machine learning for particle identification for use with the upcoming SoLID detector at the Thomas Jefferson National Accelerator Facility.

The neural network ML model was first developed using data from the SoLID Electromagnetic Calorimeter beam test. Well-constrained “pencil” simulation data with little variation from the center of the experimental configuration was used to establish an absolute baseline, then the full beam test simulation with background events was considered. The bulk of the project has consisted of applying PID methods to the actual beam test data. Preliminary results indicate that the use of the ML model is better suited for particle identification than momentum cuts on the simulation data: future work on the project will include a statistical comparison to test both how closely the simulation matches the beam test data and if the increased efficiency holds true for the beam test data.

Keywords: Particle Identification, Machine Learning, Artificial Intelligence, Neural Networks