CERN ATLAS / bachelor’s thesis / 2025
CERN ATLAS
Classifying rare ttH events.
I trained a self-attention classifier on simulated ATLAS events in the same-sign dilepton plus hadronic-τ channel, then tested what its output changed in an expected signal-strength fit.
Expected statistical uncertainty · simulated ATLAS events
A narrower expected range.
Event count onlySingle bin · no classifier
µ = 1 ± 0.28Transformer outputTen score bins
µ = 1 ± 0.24Final z-score model 0.804 ROC AUC 0.704 macro F1
The work
From events to a fit.
- PrepareParticle-specific features from simulated signal and background events.
- ClassifyA six-layer, multi-head self-attention model trained on z-score standardized inputs.
- EvaluateTRExFitter compared an event-count-only fit with a ten-bin fit using the model output.