FCC (Future Circular Collider) ROC Plot Systems For Jet Flavor Tagging Machine Learning Models

The Future Circular Collider (FCC) is the next-generation supercollider at CERN. This project focused on using machine learning techniques to optimize the detector geometry and budget for the FCC. Two proposed detectors, IDEA and CLD, were studied, both based on general designs of existing detectors at the LHC, such as CMS. The models were trained on simulated data to classify particle jets—a process known as jet tagging. Jet tagging involves reconstructing the original particle from the particles observed in the detectors. The models were then evaluated and compared using ROC curves to assess their misidentification rates on the simulated jets. This misidentification rate allows us to access which models are most accurate to be used for detector optimization.

This project focused on two pieces of software. The first was a program that allowed inference to be run on dozens of models simultaneously. The main challenge here is creating generalized code that could edit ROOT leaves with different naming conventions and then compare them against each other. The next piece was the actual ROC plotting software. We modified the current ROC plotting software to allow us to directly plot a variety of different model systems against each other. For example, the new system allows plotting of many variants of a single model for one jet flavor, or vice versa. This system allows the user to decide exactly which models and jet flavors they wish to compare without modifying the code.