USING ARTIFICIAL NEURAL NETWORK TO PREDICT PARTICLE TYPE IN HIGH-ENERGY PHYSICS
No Thumbnail Available
Date
2024-06-13
Authors
Othman, Iman
Journal Title
Journal ISSN
Volume Title
Publisher
An-Najah National University
Abstract
In the realm of high-energy physics, such as particle collision experiments in particle accelerators like the Large Hadron Collider (LHC), complex collision events occur, leading to the formation of particle jets. Precisely identifying and describing these jets is crucial for understanding fundamental particles and their interactions. However, traditional jet identification algorithms face challenges in capturing subtle features and interactions within jets, especially in dense and complex environments. Thus, predicting the type of particle in high-energy physics is essential to advancing our scientific understanding of fundamental particles and their interactions.
Artificial intelligence is a prominent research field that offers optimal solutions across various domains, especially in the context of deep learning. Many models have been studied, such as deep neural networks, jet classification, the use of neural networks, and recurrent neural networks. This study addresses the challenge of using neural networks to classify jets into five distinct categories (light quarks (q), gluons (g), W and Z bosons, and top quarks) with the highest possible accuracy. Using a model within the TensorFlow/Keras framework, we leveraged data from the Zenodo platform consisting of 150 particles with 16 attributes used for jet classification. The methods included building various neural network architectures in depth, including single-layer networks, two-layer networks, and three-layer networks. We explored different activation functions, the number of training epochs, and optimizers. Additionally, we adopted a strategy to control for overfitting and identify prominent features to improve classification performance.
The best results were achieved by building a three-layer neural network using Softmax, Sigmoid, and Selu activation functions, with the Adamax optimizer. These results were obtained after training the model for approximately 200 epochs, achieving an accuracy of 0.7400. This research highlights the potential of neural networks to achieve high levels of jet classification accuracy and provides insights into improving neural network architectures for similar tasks in particle physics research.