Mixture-of-experts graph transformers for interpretable particle collision detection

Area of Science:

• High-energy particle physics
• Machine learning applications in science

Background:

• Large Hadron Collider (LHC) generates vast, complex datasets from particle collisions.
• Graph Neural Networks (GNNs) show promise for analyzing collision data but often lack explainability.
• The
• black box
• nature of GNNs hinders trust in their predictions for scientific discovery.

Purpose of the Study:

• To develop a machine learning model that offers both high predictive accuracy and inherent explainability for high-energy physics data.
• To address the limitations of current GNNs in providing transparent decision-making processes.
• To enhance trust in AI-driven analysis for particle physics research.

Main Methods:

• Proposed a novel architecture combining a Graph Transformer model with Mixture-of-Expert layers.
• Leveraged attention maps and expert specialization for interpretable insights.
• Evaluated the model on simulated data from the ATLAS experiment, focusing on Supersymmetric signal versus Standard Model background discrimination.

Main Results:

• Achieved competitive classification accuracy in distinguishing rare signal events from background noise.
• Demonstrated that the model provides interpretable outputs linked to physics-informed features.
• The model's predictions align with established physics principles, validating its transparency.

Conclusions:

• The developed model offers a robust and transparent tool for high-energy physics data analysis.
• Embedding explainability directly into the architecture is crucial for building trust in AI for scientific discovery.
• This approach paves the way for more reliable AI-driven insights in fundamental physics research.