This project explores how machine learning can identify subatomic particles using simulated data from CERN’s LHCb experiment. The goal is to classify electrons, muons, protons, kaons, and pions from their detector signatures.
Each particle interacts with the detector in a unique way. By studying signals such as momentum, energy deposits, and Cherenkov radiation, the model learns to recognize patterns that distinguish one particle type from another.
The dataset includes several detector readings for each recorded particle track.
TrackP Total momentum of the particle EcalE, HcalE Energy deposited in electromagnetic and hadronic calorimeters RICH_DLL* Log-likelihood difference scores from the RICH detector BremDLLbeElectron Score related to electron braking radiation ParticleType True particle label (electron, muon, pion, kaon, proton) Model
Objective: Multi-class classification Algorithm: Gradient Boosted Decision Trees (XGBoost) Inputs: Detector signal features Output: Predicted particle type
The model uses these physical properties to learn how to separate different particle types. It captures real physical patterns rather than arbitrary correlations.
- Momentum Distribution
Shows how each particle type occupies a different range of total momentum.
- Calorimeter Energy Map
Illustrates how electrons deposit most of their energy in the electromagnetic calorimeter, while heavier hadrons deposit energy deeper into the hadronic layer.
- RICH Detector Comparison
Displays how kaons and protons overlap in RICH detector outputs, explaining why they are difficult to distinguish.
- Anomaly Detection
Shows particles that the model found most confusing. These may represent rare or noisy detector events.
- Metric Value Accuracy: 77.41%
- Key Features: Momentum, EcalE, HcalE, RICH_DLLbeKaon, RICH_DLLbeProton
- Hardest Classes: Kaon vs Proton
The model easily learned to separate electromagnetic particles (electrons and muons) from hadronic particles (protons, kaons, pions).
The main difficulty came from distinguishing kaons and protons, which are physically similar in both mass and behavior.
The anomalies revealed genuine detector ambiguities rather than model errors.