M.Sc. Project: Identifying Hadronically Decaying Tau Leptons Using Machine Learning

This repository presents the complete workflow for classifying hadronically decaying tau jets versus light quark jets in high-energy physics using jet image-based representation and Convolutional Neural Networks (CNNs). The analysis leverages simulated $e^+e^-$ collision events and evaluates performance across multiple center-of-mass energies.

---

📘 Project Overview

Tau lepton identification is crucial for both precision Standard Model (SM) measurements and searches for physics beyond the Standard Model (BSM).

A central motivation stems from Higgs boson analyses:

• The Higgs decays to tau lepton pairs ($H \to \tau^+ \tau^-$) with a branching ratio of ~6.3%, making it a key channel for probing Yukawa couplings.
• Hadronic tau decays, constituting ~65% of all tau decays, are challenging to isolate from quark-induced jets.
• Effective tau tagging is critical for $H \to \tau\tau$ reconstruction, as well as identifying new physics signatures involving tau leptons.

This project investigates CNN-based image classification to distinguish tau jets from light quark jets, using simulated data from:

• $e^+e^- \to \tau^+ \tau^-$ (tau jets)
• $e^+e^- \to jj$ (light quark jets, $j = u,d,s,c,\bar{u},\bar{d},\bar{s},\bar{c}$)

---

🎯 Project Goals

• Construct jet images in the $(\eta, \phi)$ plane from final-state particles.
• Apply a full preprocessing pipeline: centering, rotating, and normalizing.
• Train CNN models to classify tau vs. quark jets.
• Evaluate model generalization across various center-of-mass energies.

---

Dataset Details

• Event generation:

  • MadGraph5_aMC@NLO for parton-level events
  • Pythia8 for hadronization and decay
  • FastJet with anti-$k_T$ algorithm (R = 0.4) for jet clustering

• Jet images:

  • Format: $32 \times 32$ grayscale images
  • Plane: $(\eta, \phi)$
  • Labels:
    • Tau jet → 1
    • Quark jet → 0

• Energy configurations:

  • Center-of-mass energies: 100, 200, 300 GeV
  • Each dataset contains ~220k–240k jets

• Core jet images:

  • Also generated with R = 0.15 to capture compact energy deposits

---

🧠 CNN Architecture

Developed using TensorFlow / Keras:

• Convolutional Backbone:

  • 4 blocks: Conv2D → BatchNorm → MaxPooling

• Classifier Head:

  • Flatten → Dense(128, ReLU) → Dropout(0.5) → Dense(1, Sigmoid)

• Training Config:

  • Loss: BinaryCrossentropy
  • Metrics: Accuracy, AUC, Precision, Recall
  • Optimizer: Adam with exponential learning rate decay
  • Class imbalance handled using class_weight

---

📊 Results Summary

• Model A (Trained @ 100 GeV):

  • High accuracy at 100 GeV, but limited generalization at higher energies.

• Model B (Trained @ 200 GeV):

  • Strong performance across multiple test sets with high AUC.

• Model C (Trained @ 300 GeV):

  • Best overall classifier with AUC up to 99.78% and consistent generalization.

• Evaluation Range:

  • Test energies: 100, 150, 200, 250, 300, 350, 400 GeV
  • Performance includes: score distributions, ROC curves, tagging/mistagging vs $p_T$

---

📂 Repository Structure

├── LHE_Files/           # MadGraph-generated parton-level events
├── Tau_Pipeline/        # Tau jet simulation and image generation
├── JJ_Pipeline/         # Quark jet simulation and image processing
├── NPY_Files/           # Saved jet image arrays and labels
├── CNN_Model/           # CNN architecture, training, and evaluation scripts
├── Results/             # Plots, metrics, and final performance evaluations