Framework developments

• https://github.com/AliceO2Group/AliceO2/pull/14378
  • Bug-fix for time boundary value filling
  • Added deconvolution kernel setting for NN evaluation (makes it fully compatible with current clusterizer for flag setting)

Physics

• Now working on both data types: SC distorted and non-distorted; enforcing 0-5 centrality for better occupancy coverage
  • Training data: Combination of 0-100 centrality, woSC (50%) and 0-5 centrality, SC (50%) 
• Tested two settings for MC clusterizer, accumulation window (pad, time): (2, 4) and (3, 8)
  • For current CoG: (abs(charge_pos - CoG) < window && mc_id_charge == mc_id_cog) ? accumulate : don't accumulate
• Conclusion: Large window performs slightly worse
  • Most probably due to misassignment of MC clusters to peaks
  • More MC CoGs triggers the looper tagger earlier -> Potentially some regions are not tagged with a larger window and the network training data gets some "confused" samples

(left: larger window, right: smaller window)

• Similar observation for CoG position vs. occupancy: At higher occupancies, the smalle window size works better
• Efficiency and fake rate improve for NN, however clone rate goes up for both primaries and secondaries. (Example in figure below: clone rate for secondaries)
  Total number of clone tracks stays the same (within the uncertainty) -> Higher clone rate comes from overall reduction in number of tracks

Steps ahead

• Getting the full PID calibration to work
  • Working until skim tree creation (at least it runs through)
• More thesis writing