Hardware-aware anomaly detection for LHC event selection
Results · Getting started · Scientific report
RareGate studies the trade-off between anomaly detection and computational cost in event selection inspired by the Large Hadron Collider (LHC). It compares compact neural networks with classical statistical methods on simulated collisions, measuring signal acceptance and CPU time.
The aim is to test whether a smaller model still offers a useful advantage when feature construction, scoring and selection are measured together.
- Prepare the events. Verify public HDF5 data, remove exact duplicates and separate training, selection, calibration and evaluation. The final ADC2021 study uses 12 aggregates of missing transverse momentum, electrons, muons and jets.
- Compare the detectors. Train compact and reference autoencoders (271 and 1,999 parameters) on background events, alongside three-component PCA and regularized Mahalanobis distance. Fit preprocessing on training data only; reserve signals for evaluation.
- Measure quality and cost. Freeze background-calibrated cuts, compare signal efficiencies with paired uncertainty bounds, and measure CPU cost across batch sizes. Check numerical parity and selection effects.
Earlier LHC Olympics (LHCO) studies cover anomaly-detection baselines, synthetic queues and bounded streaming prototypes.
The final ADC2021 comparison uses 1,000,000 background events and 40,000 per signal, with a lepton prefilter. Below are signal efficiencies at a 1% target background acceptance, with CPU times per batch of 32 events:
| Method | A → 4ℓ | LQ → bτ | CPU (µs) |
|---|---|---|---|
| Compact AE | 37.33% | 6.90% | 284.9 |
| Reference AE | 50.53% | 10.63% | 295.4 |
| PCA3 | 22.72% | 11.91% | 266.3 |
| Mahalanobis | 50.13% | 15.84% | 269.1 |
CPU times cover in-memory objects → features/scaling → scores, masks and owned outputs. Data loading and HDF5 I/O are excluded; these are not detector latencies.
Signal efficiency versus joint CPU cost. Vertical intervals include calibration uncertainty; points use the frozen empirical cuts.
The smaller autoencoder lost sensitivity without meeting the required CPU saving. The paired bounds excluded 90% retention of reference efficiency on both signals, and the joint cost reduction fell short of 25%.
Mahalanobis is the recommended baseline for these features and this workload: it outperformed compact on both signals. Its absolute 15% LQ utility target remains unresolved, with a simultaneous bound of 14.65–16.97%. No method met all the preregistered quality and cost requirements.
Measurement details
Neural efficiencies average three separately trained seeds, without ensembling. CPU times are medians of eight block means from the recorded single-thread macOS arm64 run, with neural seeds averaged within each block. The figure's intervals bound efficiency at the population background quantile, including calibration uncertainty.
The scientific report provides the uncertainty analysis, selection profiles, timing boundaries and result tables.
Use Python 3.11 or 3.12 (3.12 recommended) and Git. A CPU is sufficient. The installation includes NumPy, scikit-learn, PyTorch, HDF5 support, Matplotlib and development tools.
git clone https://github.com/matteosgobba/RareGate.git
cd RareGate
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e ".[dev,data,analysis]"
python scripts/smoke_test.pyThe smoke test checks a small tensor computation. On Windows, activate with
.venv\Scripts\activate. For
CPU-only Linux, install PyTorch from its
CPU wheel index before the editable install.
To run the development checks:
python -m ruff check .
python -m ruff format --check .
python -m pytestCI is configured for Python 3.11 and 3.12. Tests use generated inputs.
Data, processed arrays and most models stay outside Git. For experiment setup, see the ADC protocol, data preparation and usage guide.
src/raregate/ Data, models, training, evaluation, benchmarks and streaming
configs/ Data, model and optional-runtime specifications
scripts/ Data preparation, validation and diagnostics
tests/unit/ Generated numerical, model and data-contract tests
results/ Selected scientific figures and numerical tables
artifacts/ Selected small historical model artifacts
docs/ Scientific findings, methodology, data and usage guides
data/ Local datasets and derived stores; excluded from version control
assets/ Project branding
Further reading: evaluation methods · installation and workflow.
Results depend on the simulation, object preselection, features and training budgets. They do not establish physical LHC rates or discovery significance; the software is not a deployment-ready trigger. Engineering prototypes retain the limitations documented in their reports.
RareGate is an independent project with no official affiliation with or endorsement by CERN, ATLAS, CMS or CERN openlab.
Created by Matteo Sgobba · GitHub · LinkedIn
This work uses public datasets from the ADC2021 authors (data descriptor) and the LHC Olympics 2020 creators, with NumPy, scikit-learn, PyTorch, h5py and Matplotlib. Dataset versions and provenance are recorded in the experiment protocols.
License: MIT for the code; CC BY 4.0 for the referenced datasets, with their respective attribution requirements.