Machine Learning Projects — Particle Physics

Thomas Griffiths — Data Analysis and Machine Learning, University of Edinburgh

Two independent notebooks applying machine learning to LHC-style collider data: an unsupervised anomaly detector for finding new physics without labeled BSM examples, and a supervised binary classifier used to sharpen a real search for a heavy resonance. Both work with simulated detector-level kinematics (jets, leptons, missing energy) and both ultimately ask the same question — can ML separate "interesting" events from Standard Model background better than hand-picked cuts — but from opposite ends: one with no signal labels at all, one with full signal/background truth.

Projects

BSM_Search.ipynb — Anomaly Detection for Exotic Event Identification

Trains an autoencoder on Standard Model (SM) events only, then uses reconstruction error as an anomaly score to flag nine different Beyond-Standard-Model (BSM) signal models (SUSY chargino/neutralino production, $Z'$ decays, gluino/RPV-SUSY) — without the network ever seeing BSM data during training. The model-independent strategy mirrors how a real search for unknown new physics would have to work.

• Headline result: the Gluino and Stlp (RPV-SUSY-like) models separate clearly from SM (~4σ / ~3σ); the SUSY chargino/neutralino and $Z'$ models tested look largely SM-like to this model.
• Full details, preprocessing steps, architecture, and known limitations: see BSM_Search.ipynb — the notebook is self-contained with inline documentation throughout.

ATLAS_VZ.ipynb — Exotic Searches with ATLAS and ML Classification

A full search analysis for a hypothetical 1 TeV heavy Higgs decaying via $H \to ZV$ ($Z \to \ell^+\ell^-$, $V \to qq'$ as a boosted fat-jet), benchmarked against three SM background processes (diboson, Z+jets, top). The analysis runs the same search twice: once with traditional rectangular kinematic cuts and a binned $\chi^2$ fit, once with a supervised neural network classifier trained to separate signal from background, each followed by a Wilks' theorem significance calculation so the two approaches can be compared directly. A final section deliberately includes the search observable (reco_zv_mass) as a training feature to demonstrate the data leakage and background-sculpting this causes.

• Headline result: rectangular cuts alone give a ~6.83σ discovery significance; adding the NN classifier (trained without the mass variable) raises this to ~14.24σ. Including the mass variable as a training feature further boosts the apparent S/B but invalidates the search by sculpting the background to mimic the signal peak.
• Full details, physics model, fit procedure, and discussion: see ATLAS_VZ.ipynb.

Shared context

Both notebooks deal with the same category of LHC event data: per-event kinematics (energy, $p_T$, $\eta$, $\phi$) for reconstructed particles, plus missing transverse energy (MET) as a proxy for invisible particles like neutrinos. Both also use PyTorch for their respective neural network components and a SM-events-only training philosophy, just applied to different problems — unsupervised anomaly detection in BSM_Search.ipynb versus supervised classification in ATLAS_VZ.ipynb.

Running

Each notebook expects its own dataset locally (data/Proj2/... for BSM_Search.ipynb, data/W18_Proj4/... for ATLAS_VZ.ipynb — see each notebook's data-loading cell for exact filenames). Requirements across both: numpy, pandas, matplotlib, seaborn, torch, scikit-learn, scipy, tqdm. Run each notebook top to bottom independently; they don't share state.

Known limitations

See each notebook's own limitations discussion for project-specific issues (e.g. BSM_Search.ipynb has a particle-counting bug that zeroes out two of its input features — documented inline).

Acknowledgements

BSM_Search.ipynb: DAML Project 2 — Anomaly Detection for Exotic Event Identification at the Large Hadron Collider, Robert Currie, University of Edinburgh. Background and BSM datasets from The Dark Machines Anomaly Score Challenge (arXiv:2105.14027).

ATLAS_VZ.ipynb: DAML Project 4 — Exotic Searches with ATLAS and ML Classification, Christos Leonidopoulos, University of Edinburgh, March 2026.