HLT1 NN based ghost killer
To be test with: lhcb-datapkg/ParamFiles!76 (merged)
Neural-Network-based ghost killer: The model has been retrained and optimized using the MinBias MC data expected for 2024 (consisting of 100k MagUp and 100k MagDown events, split into a 50% training set and a 50% test set).
Expected performance:
| Matching | Forward |
|---|---|
 |
 |
We expect a negligible efficiency impact (signal efficiency ~ 99%) and a significant ghost rejection rate (ghost rejection >~ 40%).
Performance plots - Efficiency
| Matching | Forward | Forward then matching |
|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Performance plots - Ghost rate
| Matching | Forward | Forward then matching |
|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Note: Changes in expected efficiency and ghost rates have been observed for Matching and Forward. A slightly larger impact on efficiency is noted in Forward then Matching. It's unclear if this is a cause for concern, but we might consider applying a different ghost killer threshold for Forward then Matching case, such as a tight cut in Forward followed by a looser cut in Matching.
Updating the checks in phi
| Efficiency | Ghost rate | |
|---|---|---|
| Matching |  |
 |
| Forward |  |
 |
| Forward then matching |  |
 |
Misalignement study
Using misaligned FT conditions
Efficiency at Bs2PhiPhi (With FT Misaligned)
| Matching | Forward | Forward then matching |
|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Ghost rate at Bs2PhiPhi (With FT Misaligned)
| Matching | Forward | Forward then matching |
|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
The network design is being presented for the first time at https://indico.cern.ch/event/1249044/#5-first-studies-to-reduce-ghos