sPHENIX MAPS tracker discussion

[Hideki Okawa <Hideki.Okawa AT cern.ch>]

Heavy Flavor Topical Group Meeting (2021/6/15)

https://indico.bnl.gov/event/12007/

Introduction - Jin

Congratulations to Yasser, Cameron, Xuan and Sanghoon

for the RHIC/AGS Merit Awards and Zhaozhong for the

poster award!

Hideki: We'll need b/c-filtered jet samples as well, not just

   for the machine learning b-tagger, but also for the

   substructure and D-jet correlation studies.

Jin: Indeed. Will update the intro slides.

HF tools - Cameron

p.5

Jin: Are these with all 50M events?

Cameron: Yes.

Jin: Statistics seems low. Better to double check.

Zhaozhong: B+->D0(->K+ pi-) pi+ is a statistically challenging

   channel. B_s is even lower. By the way, are J/psi, psi(2s) from

   the bbbar sample?

Cameron: Yes.

p.6

Wei: The asymmetry is not seen in other peaks. Is it due to the

   eta cut?

Cameron: No eta cut applied. B+ suffers from statistics. For J/psi,

   it could be due to the photon radiation.

Wei: In principle, they should have the same asymmetry.

Jin: Can we cut triple daughter events?

Cameron: It is not straightforward.

p.7

Jin: Better to double-check by applying |eta|<0.8, not just on

   track 1, but also track 2.

(Updates from Cameron sent to the mailing list after the meeting)

Xin: How is the precision of D0 mass position in the central region?

Cameron: I was planning to have a look, but not yet.

Jin: How was the performance in the previous workfest, where we had

   no mismatch in the magnetic field?

Cameron: The simulation was simplified back then, so I am not sure

  how reliable it would be to compare.

Muon ID w/ machine learning - Ming

p.1

Jin: Better to keep in mind the contributions from noise in the EMCal

   and HCal.

p.11

Jin: Lots of hadrons from the background are not considered here yet.

Ming: I need the MB sample.

p.12

Hideki: What does "Early Stopping" mean here?

Ming: It stops the training with a small set of train samples when the

   performance begins to saturate.

Jin, Cameron, Zhaozhong: These functionalities (Random Forest, Boosted

Trees, etc.) exist in TMVA.

Cameron: I will send you some code.

Jin: Pytorch also has these functionalities implemented.

Jin: Are 50~100M MB events sufficient?

Ming: 100M should be decent.

Cameron: We should also be careful with the over-training.

Jin: For the electrons, it is better to also check the dE and dphi with the

   3x3 clusters.

Hideki: How is the cluster energy computed? Any cut on the noise?

Jin: Basic noise and shape cuts are applied.

Xin: What is the granularity of the HCal and EMCal?

Jin: 0.1x0.1 for HCal, 0.025x0.025 for EMCal. 2x2 containes the shower

in HCal and 3x3 for the EM shower.

Xin: How is the occupancy?

Jin: Ming's plots are with pp. HCal should be quite overwhelmed in AuAu.

MDC1 vertexing performance - Dan

p.3

Jin: Is this the number of reconstructed or truth tracks?

Dan: Reco matched to the truth tracks from the PV.

Xin: Truth hits or reco hits?

Dan: Reco hits.

Xin: It would be better to make the x-axis in the log scale to see

   the low pt part in detail.

Hideki: Why are the errors significantly different between the x and z

   resolution for Num. Tracks ~15-20?

Dan: I need to check. The plots are made with TProfile.

Jin: You should check the histograms in 1D. For the resolution, you can

   use TH2::FitProfileY().

Hideki: Values in TProfile can visibly shift due to tails. You should check

   whether this is happening or not.

Xin: What is the eta cut?

Dan: |eta|<1.1

Jin: The vertex resolution looks bad. It seems that the tail effect is

   sizable.

Xin: What is the beam width?

Jin: 100 microns

Xin: What about pileup?

Jin: Not considered in these samples.

Xin: We should consider the 2D decay length instead of 3D. The z

   component may not work.

Jin: Which sample are you using?

Dan: Inclusive MB. I will also have a look at the HF samples.

Jin: Do you apply a qhat cut?

Dan: I don't know actually. Will check.

Jin: Two checks to be done: (1) check the 1D histogram for # of tracks~5,

   (2) pick a few events and compare the truth and reco. There can be

   mis-identification of the truth vertices.

Jin: Where are the RAVE vertexing plots?

Dan: I don't have the ones with MDC1 yet.

Jin: What about the ones with your own Pythia samples?

Dan: I have those plots and will add them to the slides after the meeting.

   (Slides are now updated on the indico)