CCDB fetching

• PR pending: https://github.com/AliceO2Group/AliceO2/pull/14841 (MacOS CI problem unrelated)
• Works in full system test
  • What to do about: *mConfParam = mConfig->ReadConfigurableParam()
  • What to do about metadata -> Would be good to have in order to separate networks (e.g. with different inputs)
• Uploaded right now: Networks from the commissioning runs. Do we need more?
• Upload sheet: https://docs.google.com/spreadsheets/d/1BGgDFqKnvYLlCK05hn5paQsDaiiE5HrwCErekvDqTv4/edit?usp=sharing

pp simulation

• 3000 min bias events, LHC24af, 1 MHz
• Evaluation with PbPb networks

Almost no occupancy coverage with this data. No advantage to be gained on the cluster properties (regression).

Some improvement for the qTot estimation for wide clusters (qTot / qMax large):

Still, clusters are being rejected with higher thresholds

Efficiency increases at region where highest cluster rejection occurs. No fake-rate improvement: Fake rate is already extremely low!

Tracks are so well separated that there is no real improvement to be found

Cluster error / split clusters

10 EV, 38 kHz PbPb, 0-5% centrality enforced

Investigation of split clusters with the NN

• Option 1: Use all MC charges, search for maxima that have no assigned ideal label
  • Problem: This can find maxima per MC label which might not correspond to location of maxima in digits
• Option 2: Check training data: Find all training data inputs that have class label 0 (no attached ideal cluster) but exactly one peak in the 5x5 neighbourhood with assignment (red), or multiple peaks each with assignment (blue). If the network rejects such maxima, this will correlate to a reduction in split clusters.

Examples:

• For all clusters with class label 0 (regardless of neighbouring peaks) at 0.1 threshold: 44.7% rejection. This includes looper clusters and noise peaks -> Only small drop in efficiency for split clusters: They are similarly well identifiable.