Computers can analyze complex subatomic particle data using neural networks

Overview

Particle Physicists use Liquid Argon Time Projection Chambers (LArTPCs) for the study of subatomic particles, especially neutrinos. However, the current LArTPC data analysis requires manual intervention or difficult to create custom algorithms which are time consuming and inaccurate. These algorithms can have additional problems generalizing to new data and need to be made with the specific experiment in mind. The engineering goal of this project is to create a generalized, accurate solution to rapidly analyze large volumes of LArTPC data by identifying the particle type (classification) and labeling the different parts of the image data at the pixel level (segmentation). Image data analysis is one area where deep learning excels. A neural network model to classify and segment the LArTPC data was chosen from the possible solutions of decision trees and neural networks. The dataset for training and testing the model was obtained from Deep Learn Physics. Multiple models were trained but the final iteration was a sparse convolutional network for classification and a fully sparse convolutional network for segmentation. The final classification model had an accuracy of 97.55% and an F1 score of 95.60% for classifying between EM and track particles. It had an 81.53% accuracy and an 83.54% F1 score for classifying between all five particles. The segmentation model had 99.75% accuracy and 99.61% Jaccard score. This solution allows for LArTPC data to be accurately analyzed in seconds. This automated process will help physicists learn more about subatomic particles and better understand the fundamental laws of particle physics.