Hi everyone,
I am currently working with ATLAS Open Data to test a Machine Learning algorithm for background estimation. I am manually streaming/downloading ROOT files rather than using the standard atlas-open-magic package.
Due to some network interruptions during the download process, I have ended up with a partial dataset. To account for these missing files, I have implemented the following workflow:
-
Luminosity Calculation
I applied the Good Runs List (GRL) to my data. To handle the missing files, I estimated an effective integrated luminosity by scaling the target luminosity by the ratio of processed events to the total expected events from the GRL. I am processing 2016 pp collision data and got an effective luminosity of approximately 25 fb⁻¹.
-
MC Normalization
For my backgrounds (Multijet Pythia8 JZ slices, Sherpa W/Z+jets, ttbar, etc.), I am using metadata that includes the cross-section ($\sigma$), number of events ($N$), and the sum of weights ($\sum w$). My scaling factor for each MC sample is cross section*L_eff/(sum of weights)
After passing both Data and MC through identical preselection cuts to define my Control Regions, the MC yields are significantly overshooting the observed Data, often by a large factor, despite the luminosity scaling. I suspect I may be mishandling the event weights. Specifically, I am looking at the branch (EventInfoAuxDyn_mcEventWeights)
I have 2 questions:
- Are there additional internal weights (like pile-up reweighting or SFs) in the Open Data branches that are mandatory to bring the MC down to the expected Data levels?
- Am I scaling the MC samples wrong?
Thank you very much
Best,
Francis