Higgs Diphoton Analysis with Graph Attention Models

ISEF Category: Physics and Astronomy

Subcategory: Nuclear and Particle Physics  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A tiny bump in a huge background is how particle physicists found the Higgs boson. In the diphoton channel, that bump can hide inside a flood of ordinary photon pairs. Your job is to make the signal stand out more clearly. This project turns real collider data into a test of whether machine learning beats hand-tuned cuts.

What Is It?

This project studies how well you can recover the Higgs boson in the H→γγ channel, which means a Higgs decaying into two photons. Think of it like trying to hear one flute in a marching band. The signal is rare, and most events look like background, meaning ordinary processes that mimic the same final state.

ATLAS Open Data gives you recorded collision events from the Large Hadron Collider. You can sort those events by photon quality, build a diphoton mass distribution, and look for the Higgs peak near 125 GeV. A graph attention network is a machine learning model that learns which pieces of input matter most, so it can rank events better than simple hand-picked cuts.

Why This Is a Good Topic

This is a strong science fair topic because you can measure a real physics result, compare two analysis methods, and quantify the change with clear statistics. It connects to particle discovery, detector performance, and machine learning in physics. You can learn event selection, classification, uncertainty estimation, and signal-to-background thinking, all from public data and a well-defined final target.

Research Questions

• How does a graph attention network change the signal significance of the H→γγ selection compared with cut-based selection?
• What is the effect of adding photon-shape variables on the classifier's ability to separate Higgs-like events from background?
• Does event-level graph structure improve diphoton mass resolution after selection?
• To what extent does the classifier preserve the Higgs peak region while reducing continuum background?
• Which input features contribute most to classification performance when you rank them with attention weights?
• How does the chosen decision threshold change the balance between efficiency and background rejection?

Basic Materials

• A laptop or desktop computer with at least 16 GB RAM.
• ATLAS Open Data 13 TeV event files.
• Python installed with Jupyter Notebook.
• Pandas for table handling.
• NumPy for array work.
• Matplotlib for plotting mass spectra and performance curves.
• Scikit-learn for baseline classifiers and metrics.
• A text editor or code notebook for documenting analysis choices.

Advanced Materials

• Access to a GPU workstation or university computing cluster.
• ATLAS Open Data 13 TeV event files.
• Python with PyTorch or TensorFlow.
• PyTorch Geometric or a similar graph learning library.
• ROOT for particle physics data handling.
• XGBoost for strong baseline comparisons.
• SciPy for statistical tests and uncertainty estimates.
• ImageJ only if you also test calorimeter image style inputs from derived datasets.

Software & Tools

• Python: Processes event tables, trains classifiers, and makes plots for the analysis.
• Jupyter Notebook: Lets you document code, notes, and figures in one place.
• Pandas: Organizes event features and selection tables.
• Scikit-learn: Builds baseline classifiers and computes ROC and significance metrics.
• PyTorch Geometric: Trains graph attention models on event-level particle features.
• ROOT: Reads particle physics data formats and helps with standard histogram analysis.

Experiment Steps

1. Define the physics question, the exact event features you will compare, and the metric that will decide whether the graph model helps.
2. Choose a baseline cut-based selection, then write down the same signal and background samples it will face.
3. Design the graph representation for each event, including which particles or detector objects become nodes and which relationships become edges.
4. Build a fair comparison plan so the machine learning model and the cut-based method use the same training, validation, and test split.
5. Plan how you will turn classifier output into physics results, including a diphoton mass fit and a signal-strength estimate.
6. Decide which uncertainty checks you will run, such as threshold scans, feature ablations, and performance stability across subsamples.

Common Pitfalls

• Training and testing on overlapping events, which makes the classifier look better than it really is.
• Optimizing only ROC AUC and ignoring whether the Higgs mass peak survives the final selection.
• Using photon-quality inputs that leak direct signal labels, which turns the comparison unfair.
• Comparing the graph model to weak cut choices, which makes the improvement hard to trust.
• Skipping uncertainty checks on the signal-strength estimate, which leaves the final result too fragile.

What Makes This Competitive

A competitive project would not just train a model. It would compare methods under the same event split, quantify uncertainty on the signal strength, and test whether the gain survives different thresholds and feature sets. Strong work would also connect classifier performance to the actual Higgs mass peak, not just a machine learning score. If you can show a clear physics benefit with careful controls, your project looks much more serious.

Project Variations

• Use jet-related event features instead of photon-quality features and test whether a graph model still improves background rejection.
• Compare a graph attention network with XGBoost, a multilayer perceptron, and a cut-based baseline on the same ATLAS sample.
• Repeat the analysis on a different diphoton selection region, then check whether the model still preserves the Higgs peak.

Learn More

• ATLAS Open Data: Search the ATLAS Open Data portal for 13 TeV collision samples, analysis notes, and example code.
• CERN Open Data Portal: Look for public particle physics datasets and documentation on event formats.
• MIT OpenCourseWare, Introduction to High Energy Physics: Use lecture notes and problem sets to review collider basics and particle detection.
• Particle Data Group: Search the Review of Particle Physics for Higgs boson properties and standard notation.
• PubMed: Search for review articles on graph neural networks in particle physics to see how researchers frame similar analyses.