Average size of looper when charge overlap starts

I.e. when do we start to find looping legs as tracks and not just mangled clusters 

• Average pad-width [cm] of clusters vs. row number

• Use clusters between 60 < row < 100 as average: <sigma_pad> ~ 0.4cm
• Assume overlap starts with <3*sigma_pad separation: 2.4cm ; Cluster height between 60 < row < 100: ~1.1cm
• Need 10 clusters for tracking looper legs
  Geometric conditions:
  • Charge mangling: sin(30) = 2.4cm / R
  • Min. NCl for tracking: sin(70) = 5*1.1cm / R
    • R_{looper legs}^{min} = max(condition 1 , condition 2) = 5.9cm = 5.3 pads in height = 9.8 pads in width
• This coincides well with one of the plots from my thesis

• Corresponding pT: pT_{looper legs}^{min} = 8.9 MeV/c

Receiver-Operator-Characteristic (ROC) and Area-under-Curve (AUC)

• 3D > 2D, but no significant difference between all 3D networks
• Classifier with >90% AUC is typically categorized as excellent, >85% as very good

RTX5080 vs RTX3090

• Forget the memory usage for now...
• NN's are approx 2.1-2.4x faster on modern GPU in both FP32 and FP16