Dark Machines anomaly detection challenge published¶
The results of our community-wide anomaly detection challenge are out in SciPost Physics. We benchmarked hundreds of unsupervised ML models on over 1 billion simulated LHC events.
The Dark Machines initiative brought together physicists and machine learning researchers to tackle a fundamental question: can we find new physics at the LHC without knowing what we're looking for?
We generated a large benchmark dataset corresponding to 10 fb\(^{-1}\) of 13 TeV \(pp\) collisions and tested a wide range of algorithms — auto-encoders, normalising flows, deep sets, variational approaches, and more — in realistic analysis environments.
Key takeaways¶
- No single model dominates across all signal types.
- Ensemble approaches and normalising flows performed consistently well.
- The benchmark dataset and code are publicly available for future studies.
The benchmark data is available at phenoMLdata.org and the analysis code at GitHub.
Links: arXiv:2105.14027 · CERN Courier coverage