Deep Learning Approaches for BSM Physics: Evaluating DNN and GNN Performance in Particle Collision Event Classification

TL;DR

This paper compares deep learning models, DNNs and GNNs, for classifying particle collision events to detect BSM signals, showing high accuracy and highlighting GNNs' ability to model particle relationships.

Contribution

It introduces and evaluates GNN architectures for BSM event classification, demonstrating their potential alongside traditional DNNs in high-energy physics.

Findings

All models achieved AUC > 94%.

DNN slightly outperformed GNNs overall.

GNNs effectively capture inter-particle relationships.

Abstract

Detecting Beyond Standard Model (BSM) signals in high-energy particle collisions presents significant challenges due to complex data and the need to differentiate rare signal events from Standard Model (SM) backgrounds. This study investigates the efficacy of deep learning models, specifically Deep Neural Networks (DNNs) and Graph Neural Networks (GNNs), in classifying particle collision events as either BSM signal or background. The research utilized a dataset comprising 214,000 SM background and 10,755 BSM events. To address class imbalance, an undersampling method was employed, resulting in balanced classes. Three models were developed and compared: a DNN and two GNN variants with different graph construction methods. All models demonstrated high performance, achieving Area Under the Receiver Operating Characteristic curve (AUC) values exceeding $94\%$. While the DNN model slightly outperformed GNNs across various metrics, both GNN approaches showed comparable results despite different graph structures. The GNNs' ability to explicitly capture inter-particle relationships within events highlights their potential for BSM signal detection.