Quark Gluon Classification using Contrastive Learning Representations
Supervisor: Dr. Tilman Plehn
Our building blocks are atomic nuclei, which in turn comprise protons and neutrons, which are made up of quarks and gluons. This project aims to distinguish the quarks (signal) from the gluons (background).
When a collision takes place in the Large Hadron Collider, constituent particles get scattered. The set of scattered particles are called jets. Snapshots of jets are called jet images and have been conventionally used to classify between quark and gluon initiated jets. In this research work, we adopt an alternative technique. We train a transformer with contrastive loss to map particle jets to a representation space. Spherical position, particle type and transverse momentum of the constituent particles are taken into consideration while obtaining the representation of jets.
These linear representations are fed into machine learning models for binary classification tasks (to distinguish between quarks and gluons.)
Technical Details
- Language: Python
- Framework: Pytorch
- Models Experimented With: Logistic Regression, Linear Discriminant Analysis, Random Forest, XGBoost
- Loss Functions: Hinge Loss, Squared Hinge Loss, Mean Squared Error
- Dataset Size:
- Train: 90000 jets
- Test: 10000 jets
Results
We’ve obtained an AUC of 0.8 so far. Research work is still ongoing.