Roundtable
Paul
CI-bridge: planned to get update in for proper performance tracking in a less hacky way. SHould be finished this week.
ODD-dataset: working on EDM4hep conversion. Local to bound correction is now in review and measurements can persist. Plan to improve IO in examples. EDM4hep read-in and then run whole reconstruction there.
Geometry: Some updates. Athena-update needed. Plan to merge this week.
Alex
Athena GX2F:
It seems, there is a prefit not considering scattering angles. Then the strongest scatterer is checked for and used as a kink. Full derivatives are repeated. This seems odd, because we would need too many iterations.
Continued to comment the GX2F and marked unclear points. It could turn out, some of them are bugs or just magic numbers.
Doga
QP: Last report was well received. Discussion on physics performance samples.
Remark was that physics group should have been contacted.
Keep suggested parameters as default.
Preparing for EP-meeting on Thursday.
Noemi
Implement additional material layers in tracking geometry. Update can be steered with a flag.
There seems to be some bugs still, need to work on that. Will have discussion afterwards.
PF
Tested new tracking geometry with sensitive material layers in between. Went through covariance updates in all steps. Material layers 10mm fro surface and in between. Material gets projected to surfaces and only fraction to intermediate layer. Strip[s: We basically get the filtering with small noise from material. It is very close to the previous case.
Noemi: Igor's trackfinding: 40% material per layer. -> Andreas Salzburger: We want a different approach! Optimise the maps and then in the trackfinding, we could do Igor's trick. Keep material model exact and allow for more chi2 deviation. After smoothing, we get a nice track.
Paul: If we can have it finetuned between layers, it might be beneficial.
Paul: Idea: Pull out material from geometry and pass it as a config.
Andreas Salzburger: Maybe a problem with different material in filtering and smoothing.
Currently, material is before sensitive but it should be right after. This represents more how the detector is built. Then we would also open up the covariance for the next suface
Approach: First map where it is in real detector and then do some tricks.
Sizes of covariance matrices. Local x and y get to micron level because set in the finding. This might be too small. In legacy, the calibration happens after the finding. Now we use calibrated measurements because: 1. errors are small and 2. charge should switch. Efficincy gets slightly better. Charge gets negative from TRT and we still need to check if valid. THis could be still fine, because all pixels have different calibration curves.
Paul: This only effects us, when we have small errors. Just for the finding, we could use the uncalibrated measurements. -> Problem is that we will have a problem after filtering.
Andreas Stefl: Legacy chi2 is meaningless/with a weird scale.
Paul: Pull track hypothesis with measurement? -> Might influence the chi2-distribution -> If we do it only in the finding, it might not contributeto the final distribution.
It should be also cheap to refit the track after the finding.
Andreas Stefl: Further discussion on the calibration is needed anyway. Current method is not correct for finding.
Paul: Maybe we can separate the 2 requirements, then we might not need to touch the calibration.
Mikkel
GNN-training started and talking to Noemi, how to continue.
Dimitra
Validation of tracking geometry. Surfaces canditating needs to be increased. Taking first simhit in muon. Were seeing outliers in HDT. Not grid-naviagtion and also with try-all it persists. ATLAS magnetic field: Direction does not change in propagation (muons and charged geantinos).
Neza
Updated to lates traccc version
Plan to do some timing on traccc to find hotspots in own code.
Stephen
traccc seeding. Export both seedings as csv and compare.
Currently, we have from N hits (N choose 3) hits. With the graph-based method, we might reduce that number. We still need to figure out, which one is most important
Andreas Salzburger: Igor classified seeds as PSP (PixelStripPixel), PPS, etc. Then for each detector region, there is an optimal type and the others can be thrown away.
Noemi: This might make problems, when the detecotr changes (e.g. parts turned off)
Noemi: seeds with same bottom and middle and different top: if there is a certain match, we might merge the particles into a single one.
Paul: Deduplication is a problem with multithreading. Maybe run in batches, to synchronise?
Currently 200k-300k are visible. -> PF: 8-10k per event and process 5k after deduplication -> Maybe one simple cut possible?
Andreas Salzburger: On GPU, we don't want to do too much book keeping.
Andreas Salzburger
Python-plugins: Working version exists, but needs clean-up.
Started migrating some of the ROOT functionality for material. Plan to have a simple shell around for material mapping with Athena. Currently not sure, if we need more rules for the mapping.
Benjamin
Prepared testbed and will present tomorrow first measurements. Preparing Athena integration, which is a big milestone for the mid-cycle.
GSF: Next week: overview on measurements with mixed values
Andreas Stefl
Followed trackfinding discussions. Carlo's calibration effects are interesting for deduplication. It seems, we can tackle all problems now.
Thesis work.
