Unsupervised Quark/Gluon Jet Tagging With Poissonian Mixture Models.

Area of Science:

• High-energy physics
• Particle physics
• Machine learning applications in physics

Background:

• Distinguishing quark and gluon jets is crucial for New Physics searches at high-energy colliders.
• Current jet classification methods often rely on Monte Carlo simulations, introducing potential theoretical and systematic uncertainties.
• There is a need for unbiased and robust jet classification techniques.

Purpose of the Study:

• To develop an unsupervised learning algorithm for classifying quark and gluon jets without relying on simulations.
• To learn the SoftDrop Poissonian rates and fractions for quark- and gluon-initiated jets directly from data.
• To construct and evaluate an interpretable quark-gluon tagger with minimal assumptions.

Main Methods:

• An unsupervised learning algorithm was developed to analyze jet samples.
• Maximum Likelihood Estimates were used to determine mixture parameters and posterior probabilities.
• A quark-gluon tagger was constructed based on the learned parameters.
• Unsupervised metrics were employed for hyperparameter selection.

Main Results:

• The algorithm successfully learned SoftDrop Poissonian rates and fractions for quark and gluon jets.
• The resulting unsupervised quark-gluon tagger achieved an estimated accuracy of 0.65-0.7 in actual data.
• The tagger's performance remained robust even with simulated detector effects (angular smearing).
• Unsupervised metrics proved effective for hyperparameter optimization.

Conclusions:

• The proposed unsupervised learning algorithm provides a viable, interpretable, and less biased alternative for quark-gluon jet classification.
• The method demonstrates competitive performance compared to supervised approaches, particularly in reducing reliance on simulations.
• This approach offers a promising direction for future New Physics searches at colliders.