T5 inefficiencies for mildly displaced tracks

[slava77]

to keep a note that this needs some follow up.
https://indico.cern.ch/event/1598658/#10-contribution is a starting point

suggests that the slightly displaced T5s are inefficient mainly in the more forward direction

[GNiendorf]

Failure cases: https://gist.github.com/GNiendorf/395dff7e53a183a30a406e3a2df3090f

At first glance, it seems like the T3 DNN may be contributing to some failures at large r_vertex?

[GNiendorf]

Here are cases where a T4 candidate exists, but a TC was not created: https://gist.github.com/GNiendorf/2bdd95e6f33562a71113bfb6002f2747

[GNiendorf]

@slava77 Improvement from turning off the T3 DNN

[slava77]

@slava77 Improvement from turning off the T3 DNN

Does FR explode in this case?

[GNiendorf]

https://gist.github.com/GNiendorf/c9d6fb6f483a2759180bdb8ab684a803

Same printout if T3 DNN is turned off. Seems like a mix of them getting killed by one of the T5 cuts, some having no MD's in layers 1 or 2 (only layer 3), and some where the MD's exist but the LS's aren't formed.

[GNiendorf]

@slava77 No, the FR doesn't increase much from the T3 DNN being turned off. The T3 DNN just gave a moderate timing improvement, removing like ~60% of fake T3's so they don't need to be considered during T5 building.

[slava77]

@slava77 Improvement from turning off the T3 DNN

wouldn't this (potentially) further increase T4 efficiency?

The plot is almost unchanged for rvertex<5 cm, where T4s are apparently recovering the inefficiency for |eta|>1.5.
At least for this region it looks like the relaxed/removed T3 DNN is not really the solution.

[GNiendorf]

wouldn't this (potentially) further increase T4 efficiency?

Yeah it's possible, same for any corrections we make the module maps.

[GNiendorf]

Looking at Jade's plot today, seems like it's a T5 cut responsible for the <5 cm loss? Probably the T5 DNN?

[GNiendorf]

@slava77 Comparison with both T3 and T5 DNN turned off. I think that explains where it's coming from.

[slava77]

@slava77 Comparison with both T3 and T5 DNN turned off. I think that explains where it's coming from.

and yet, the first two bins don't move, the place where Jade's plots with |eta|>1.5 show a significant addition from T4s.

[GNiendorf]

@slava77 Comparison with both T3 and T5 DNN turned off. I think that explains where it's coming from.

and yet, the first two bins don't move, the place where Jade's plots with |eta|>1.5 show a significant addition from T4s.

I would imagine it's a different T5 cut causing that, probably the RZ-chi squared cut. Look at the green/red comparison for the first few bins in Jade's plot.

[GNiendorf]

And look at how big the error bars are on the first couple of bins for the |eta|>1.5 plot, and that's at 10k events...

[GNiendorf]

@Slava Here is with T3 DNN + T5 DNN off + tightCutFlag always set to true.

[GNiendorf]

Slides from today: Issue #209_ Inefficiencies for Displaced Tracks.pdf

Note: All plots here are from the low-pT config of the code.