The LHC experiments use unsupervised neural networks (autoencoder) for anomaly detection to increase sensitivity to BSM models and remove trivial Standard Model backgrounds. Such neural networks are trained using a small fraction of data. This web service is designed to calculate acceptance corrections for any BSM or SM event records.

Compressed LHE, ProMC and ROOT files and slimmed Delphes ROOT files with the size less than 150MB are supported. These files can be downloaded from HepSim repository. Before upload, transform Delphes ROOT files as explained here.

To start processing, upload input file (*.root, *.promc, *.lhe.gz)

Choose autoencoder:
Single leptons pT>60 GeV CM=13 TeV (ATLAS, Phys. Rev. Lett. (2024) 132, 081801) Jets SM MC 2→2 1 TeV CM=14 TeV (Eur. Phys. J. Plus (2024) 139:237) ATLAS multitrigger (not for public) CMS ? (not available)

Output Results:

• ROOT file with reconstructed objects (and RMM) for TensorFlow input
• ROOT file with events, cutflow and invariant masses in the anomaly region.
• Log file with all steps

---

One can also process Monte Carlo events and data using DAOD_PHYS and DAOD_PHYSLIGHT, Run2+Run3 autoencoder and multiple trigger streams. Contact the authors of this tool for the instruction. You need to be an ATLAS member. Here is the list of triggers:

1. MET above 200 GeV
2. 2 leptons with pT.g.30 GeV
3. 1 photon (pT.gt.140 GeV)
4. 2 photons (pT.gt.30 GeV)
5. 1 jet (pT.gt.500 GeV)
6. 4 jets (pT.gt.100 GeV)

---

The ATLAS / JDM anomaly search team: S.Chekanov, W.Islam, R.Zhang, S.Mohiuddin, N.Luongo
For ATLAS-specific autoencoders, ask somebody from this team.