Overview
This work introduces novel evaluation metrics for generative models in high-energy physics, providing a comprehensive framework for assessing model performance and reliability.
Highlights
- Introduced two new metrics: Fréchet Physics Distance (FPD) and Kernel Physics Distance (KPD)
- Conducted systematic evaluation of metric sensitivity to various failure modes
- Demonstrated practical application in comparing transformer and GAN models
- Provided open-source implementation in JetNet Python library
Technical Contributions
- Metric Development: Novel physics-aware distance measures
- Sensitivity Analysis: Comprehensive testing of metric behavior
- Framework Integration: Implementation in existing HEP tools
- Practical Validation: Real-world application to model comparison