Search for New Physics at the Large Hadron Collider Using Unsupervised Machine Learning for Anomaly Detection

Urgency in LHC Anomaly Hunts

The Large Hadron Collider has delivered unprecedented data in recent years, with 2025 marking a record integrated luminosity exceeding previous benchmarks during the ongoing Run 3 phase. This surge in collisions provides the raw material for increasingly sophisticated analyses, particularly those employing unsupervised machine learning to flag anomalies that deviate from Standard Model predictions without presupposing specific new particles or processes.

Argonne National Laboratory's high-energy physics group has contributed notably to this approach, developing techniques centered on autoencoders—neural networks trained to reconstruct typical collision events, thereby highlighting outliers as potential indicators of new physics. Such methods gained traction after pioneering ATLAS results using Run-2 data and have evolved with recent arXiv publications and experimental implementations, including tools like ADFilter for processing events in a model-independent way.

The timing proves critical because Run 3 concludes around mid-2026, after which Long Shutdown 3 commences, halting physics operations for roughly four years to install upgrades for the High-Luminosity LHC. This upgrade promises a tenfold increase in integrated luminosity starting in 2030, dramatically expanding discovery potential but also introducing pile-up challenges that will demand even more advanced data-handling strategies. Researchers thus face a narrow window to apply these anomaly detection innovations to the current dataset, where any signal could guide preparations for the HL-LHC era.

Stakes extend beyond academia: failure to uncover deviations now could delay understanding of fundamental questions like dark matter composition or Higgs sector extensions, while the multibillion-euro HL-LHC investment hinges on justifying continued exploration of ever-higher energies. Non-obvious tensions arise in balancing computational demands of real-time ML applications against finite resources, and in reconciling the broad sensitivity of anomaly searches with the risk of false positives that could misdirect experimental priorities.

These developments reflect a broader methodological shift in particle physics, away from hypothesis-driven searches toward data-centric discovery, prompted by the persistent absence of clear beyond-Standard-Model signals despite decades of effort.