Skip to content

T-tracks CatBoost filter

Selections with T-tracks that can't be gated by a rare process such as a detached J/ψ→µµ show a significant slowdown due to the fact that all (clone-killed) T-tracks are passed through Kalman filtering and RICH reconstruction.

Since a majority of T-tracks are of little interest, we propose an MVA-based selection using CatBoost to select likely signal T-tracks and only pass these through the reconstruction chain. This also reduces the combinatorics in the selections.

The MVA has as features the transverse momentum, the radial distance from the beam pipe, the z-momentum, the Y position of the first measurement and the pseudo rapidity.

Currently trained on Lb2JpsiLambda (Lambda -> proton pion T tracks) signal.

A cut of prob > 0.3 would reduce the number of tracks by around 65% and only reduce signal by around 15%.

Presented in WP2 meeting on 30/5/23 and WP2 meeting on 13/6/23

Have additionally added a track pair MVA filter -- this is trained on MC using any true track pair as signal, any non-pair as background. This is essentially to reject tracks that can't easily be paired to another track.

With a cut of single track score > 0.3 and track pair score > 0.75, about 90% of tracks are rejected (and about 95% of track combinations)

Configured in Moore!2367 (merged)

ParamFiles MR: lhcb-datapkg/ParamFiles!62 (merged)


  • Add CatBoost model binary to paramfiles
  • Use ParamFileSvc to access CatBoost model file
  • Update model location in the algorithm
  • Add qmtest in Moore
Edited by Izaac Sanderswood

Merge request reports