LCFI Package and Flavour 3TeV Tomáš Laštovička Institute of Physics AS CR CLIC WG3 Meeting 9/6/2010
Page 2 LCFI Package Used for jet flavour tagging and secondary vertex reconstruction. Topological vertex finder ZVRES. Standard LCIO input/output –Marlin environment (used for both ILD/SiD) Flavour tagging based on Neural Nets. –Combine several variables… Probability Tubes Vertex Function
Page 3 NN Input Flavour Discriminating Variables There are 14 flavour discriminating variables R - and Rz- significance for 2 tracks with the highest impact parameter significance in R (“leading tracks”) Relative momenta of the leading tracks (relative to jet energy) Joint Probability in R and Rz Decay length and decay length significance (relative to jet energy) Pt-corrected vertex mass Secondary vertex probability Relative total momentum of non-primary vertex tracks and their number These inputs are re-normalised and transformed by tanh() - except joint and secondary vertex probabilities. Tracks/vertices have to pass some minimal selection cuts.
Page 4 NN Input Flavour Discriminating Variables Inputs are sent to 3 neural networks (8 inputs each) according to the number of secondary vertices found in a given jet –0 vertices: R -, Rz- significance and momenta for 2 leading tracks Joint Probability (R , Rz) –1 vertex and >1 vertices: Decay length, decay length significance, pt-corrected vertex mass, Total momentum of non-primary vertex tracks and their number, Joint Probability (R , Rz), Secondary vertex probability This is not a dogma, inputs can be added/removed –Requires some coding. –Studies better done outside the package (I fancy FANN package for this purpose).
Page 5 Input Variables – Additional Topics Joint Probability Calculation –Estimated using fits to impact parameter distributions. –Might depend on detector geometry and sim/rec effects. K s, and conversion tagger –Part of the package, depends on detector geometry. Cuts on tracks/vertices for NN Inputs –There is a number of parameters to tune up the package (see next slide).
Page 6 LCFI Package Optimisation Optimisation is not only a matter of Neural Net retraining. The package has plenty of parameters: –Track selection params –ZVRES params –Flavour Tag params –Vertex Charge params
Page 7 Example 1 SiD FastMC 500GeV ISR removed by M inv cut b-jets (red) c-jets (green) Light-jets (black) R 1 R 2 Z 1 JP R JP Z M 1 M 2 DL SDL Pt CMRM #t V#V SVPE Z 2
Page 8 Further Examples I compared various samples (sorry for too many plots). Let’s start with the same setup but for 3 TeV –It’s pretty much similar as far as input variables are concerned.
Page 9 SiD FastMC 3TeV ISR removed by M inv cut SiD FastMC 500GeV ISR removed by M inv cut b-jets (red) c-jets (green) Light-jets (black) R 1R 2 Z 1Z 2 JP R JP ZM 1M 2 DL SDLPt CMRM #t V#VSVPE
Page 10 Further Examples I compared various samples (sorry for too many plots). Let’s start with the same setup but for 3 TeV –It’s pretty much similar as far as input variables are concerned. ff 2-jet 3 TeV
Page 11 3TeV ISR removed by M inv cut ILD Full Sim/Rec 3TeV DST files area normalised M inv cut R 1R 2 Z 1Z 2 JP R JP ZM 1M 2 DL SDLPt MCRM #t V#VSVPE b-jets (red) c-jets (green) Light-jets (black)
Page 12 Further Examples I compared various samples (sorry for too many plots). Let’s start with the same setup but for 3 TeV –It’s pretty much similar as far as input variables are concerned. ff 2-jet 3 TeV H 0 A 0 4-jet events –First reconstructed with the SiD FastMC, –then with the full simulation and reconstruction. –Please, ignore c-jets.
Page 13 3TeV ISR removed by M inv cut SiD FastMC H 0 A 3TeV no M inv cut 4 jet events area normalised b-jets (red) c-jets (green) Light-jets (black) b-jets (red) c-jets (green) Light-jets (black) R 1R 2 Z 1Z 2 JP R JP ZM 1M 2 DL SDLPt MCRM #t V#VSVPE
Page 14 ILD Full Sim/Rec H 0 A 3TeV DST files 224 – 231, jet events area normalized R 1R 2 Z 1Z 2 JP R JP ZM 1M 2 DL SDLPt MCRM #t V#VSVPE b-jets (red) c-jets (green) Light-jets (black) SiD FastMC H 0 A 3TeV no M inv cut 4 jet events area normalised
Page 15 Discussion SiD FastMC consistent for 500GeV and 3TeV. –And consistent to full SiD 500GeV. Then things get bit more complicated to compare –Different events, detectors, reconstruction, low statistics. –ff events comparable for b- and c-tag. Light jets probably polluted (?). –H 0 A 0 events: b-events more or less OK, however: Differences between FastMC and full simulation reconstruction (e.g. P t corrected mass secondary vertex reconstruction?). Different input distribution compared to the reference one worse performance with default nets.
Summary LCFI package has a number of flavour tag sensitive variables, these can be revised/modified. We’ve looked at a couple of samples using SiD FastMC as well as DST files from Marco (full simulation and reconstruction). Future Plans: b-tag will be studied more closely. c- and uds- mistag efficiencies. Optimisation of the LCFI package.