Where do we use Machine learning and where do want to improve?

Size: px

Start display at page:

Download "Where do we use Machine learning and where do want to improve?"

Martina Owen
6 years ago
Views:

1 Where do we use Machine learning and where do want to improve? Sascha Stahl, CERN Paul Seyfert, INFN On behalf of LHCb

2 The LHCb detector Vertex and track finding Particle identification and first event trigger LHC Run2: ~45 khz bb pairs and ~1 MHz cc pairs (Signal!) New York, Slide 2

3 Tracking challenges x z Relatively high fake track rate Secondary particles from material interactions No tracking stations inside B-field Dipole magnet: (T) B y VELO TT T1 T2 T3 LHCb-DP Tracking needs to be fast and efficient z (m) New York, DS@HEP, Tracking@LHCb Slide 3

Trigger New for Run2: Same track reconstruction in trigger and offline Allows offline quality analyses in trigger Track reconstruction separated in two stages

4 Trigger New for Run2: Same track reconstruction in trigger and offline Allows offline quality analyses in trigger Track reconstruction separated in two stages Fast stage for long tracks with high p T (p T > 500 MeV) Full stage achieves offline efficiency and precision New York, DS@HEP, Tracking@LHCb Slide 4

Trigger (Upgrade 2020) Mandatory for Upgrade: Same track reconstruction in trigger and offline Allows offline quality analyses in trigger Track reconstruction

5 Trigger (Upgrade 2020) Mandatory for Upgrade: Same track reconstruction in trigger and offline Allows offline quality analyses in trigger Track reconstruction separated in two stages Fast stage for long tracks with high p T Full stage achieves offline efficiency and precision New York, DS@HEP, Tracking@LHCb Slide 5

6 Fast stage (1) LHCb-PUB LHCb-PUB LHCb Velo tracking Velo TT : Initial momentum estimate TT T-stations: Full long track (forward tracking) Full Kalman filter: Optimal track parameters Partical event reconstruction, runs at 1 MHz input rate. Upgrade 30 Mhz. Velo tracking: Search for straight tracks in Velo (no B-field) Velo TT tracking: Velo TT Extend straight Velo track to TT Search for nearby clusters (Low B-field ) TT T-Stations T-Stations Match VeloTT track with hits in T-Stations New York, DS@HEP, Tracking@LHCb Slide 6

7 Fast stage (2) LHCb-PUB LHCb-PUB LHCb Velo tracking Velo TT : Initial momentum Primary estimate vertices TT T-stations: Full long track (forward tracking) Full Kalman filter: Optimal track parameters Time: ~18 % ~3 % ~22 % ~50 % Track fit needed to get optimal track parameters Quite slow: Impact parameter resolution Propagation through many different materials (mitigated by using a simplified geometry) Propagation through B-Field (Runge-Kutta method used) Can we use ML to get good parameter estimates and make it faster? New York, DS@HEP, Tracking@LHCb Slide 7

8 Full stage Full event reconstruction, tracking and particle identification, runs at khz input rate. Upgrade O(Mhz)? 1) Start with tracks from fast stage 2) Forward recovery loop for low p T tracks 3) Standalone T-track reconstruction (on all T-station hits, very time consuming) 1) Velo track + T-track matching Complementary long track algorithm 2) TT clusters + T-tracks Downstream tracks 4) Kalman Filter + clone and fake track rejection Uses Neural Net classifier New York, DS@HEP, Tracking@LHCb Slide 8

9 Ghost probability Fast neural net based fake track rejections Inputs (21): Track fit quality, number of hits, momentum, detector occupancies. First (2012), heavily used in centralised event selection Reduced output to manageable levels Second (2015), implemented in final track selection Rejects fake long tracks and increases Downstream track efficiency Third (2016), implemented in first software trigger stage Reduces significantly rate to second level with small efficiency loss ) 2 candidates / (4 MeV/c fake track rejection LHCb preliminary ghost probability track fit 2 /dof LHCb preliminary 0 all D fakes m K K 2 [MeV/c ] effi ciency New York, DS@HEP, Tracking@LHCb Slide 9

10 Make it faster Ghost probability must be computed fast (numbers for TMVA) Neural network faster than BDT (40x) Compile network instead of loading at runtime (4x) Tune auto-generated network code by hand (2x) Faster network activation function (uncharted, 4x) Drop support for >5yr old CPUs (10x) Make auto-generated code better? Lower is better BDT with.class.c MLP with.class.c hand tuned MLP BDT with.xml MLP with.xml MLP with improved.class.c AVX New York, Slide 10

Treat Simulation data differences ML trained with simulated data Degraded performance on real data? Physics analyses put a lot of effort into correcting simulation Can we do better? (i.e. let the computer deal with the problem) Started to explore Domain adaption (arxiv:1409.

11 Treat Simulation data differences ML trained with simulated data Degraded performance on real data? Physics analyses put a lot of effort into correcting simulation Can we do better? (i.e. let the computer deal with the problem) Started to explore Domain adaption (arxiv: ) network oblivious of the domain. No success so far New York, DS@HEP, Tracking@LHCb Slide 11

12 Back to fast stage LHCb-PUB LHCb-PUB LHCb Velo tracking Velo TT : Initial momentum estimate TT T-stations: Full long track (forward tracking) Full Kalman filter: Optimal track parameters x u v x x u v x x u v x Search window in T-Stations defined by track estimate and minimal p T Project x-hits into a reference plane Clusters Fit clusters and remove outliers Add stereo hits Track candidate handed to Kalman filter Recovery loop for track candidates with hits in only 4 x-layers New York, DS@HEP, Tracking@LHCb Slide 12

Improvement in Forward tracking Trained two neural networks: For rejection of bad clusters with only 4 out of 6 x-layers For final track selection Type:

13 Improvement in Forward tracking Trained two neural networks: For rejection of bad clusters with only 4 out of 6 x-layers For final track selection Type: MLP with 3 hidden layers Trained for background rejection at given efficiency (97% or 99 %) LHCb unoffical New York, DS@HEP, Tracking@LHCb Slide 13

14 Improvement in Forward tracking Results: Increased efficiency! Significantly reduced fake rate! 5% less time in first trigger stage Time in second trigger stage increases: LHCb unoffical MVA rejects more candidates 4 layer recovery loop called more often Compensated by fewer tracks to fit and rejection of 90% of 4 x-layer clusters Numbers relative to not using NN Scenario A: No recovery of 4 x- layer candidates Scenario B: Recovery of 4 x-layer candidates New York, DS@HEP, Tracking@LHCb Slide 14

15 Improvement in Forward tracking Results: Increased efficiency! Significantly reduced fake rate! 5% less time in first trigger stage Time in second trigger stage increases: MVA rejects more candidates 4 layer recovery loop called more often Compensated by fewer tracks to fit and rejection of 90% of 4 layer clusters LHCb unoffical Can we do better? Final track selection works well! Early cluster rejection promising but room for improvement! Numbers relative to not using NN Scenario A: No recovery of 4 x- layer candidates Scenario B: Recovery of 4 x-layer candidates New York, DS@HEP, Tracking@LHCb Slide 15

16 Going further (T-Seeding) LHCb Standalone T-track reconstruction Needed to reconstruct very displaced particles Increases Long track efficiency Algorithm quite efficient with relatively low fake rate Takes 25% of the reconstruction time 50 % spend in fit of track candidates Can we identify and select good clusters with ML faster? New York, DS@HEP, Tracking@LHCb Slide 16

17 Conclusion LHCb uses successfully neural networks in pattern recognition and track selection Good to reject fake tracks Can be used to be more efficient Can we make MVA more robust against Data-Simulation differences? Speed is always a question! A big thanks to the DS@HEP organizers for this event and the Simons Foundation for hosting us. New York, DS@HEP, Tracking@LHCb Slide 17

18 New York, Slide 18

19 The tracking system x z Velo: Silicon strips (r,φ sensors) TT: Silicon strips (x, ±5 sensors) T-Stations Inner region: Silicon strips Outer region: Drift tubes (x, ±5 sensors) Dipole magnet: (T) B y VELO TT T1 T2 T3 Velo tracks: LHCb-DP New York, DS@HEP, Tracking@LHCb Slide 19 z (m) Vertex reconstruction Long tracks: Main input to analyses Downstream tracks: e.g. for Ks and Lambda decay products

20 Fast stage (1) LHCb-PUB LHCb-PUB LHCb Velo tracking Velo TT : Initial momentum Primary estimate vertices TT T-stations: Full long track (forward tracking) Full Kalman filter: Optimal track parameters Velo tracking: Search for straight tracks in Velo Start from end with r-sensors First candidates with four hits, then three hit candidates Add hits from φ sensors New York, DS@HEP, Tracking@LHCb Slide 20

21 Going further (Kalman filter) Momentum resolution Impact parameter resolution Int.J.Mod.Phys. A30 (2015) no.07, Track fit needed for best track parameter estimates Kalman filter very time consuming in current reconstruction and in reconstruction for the upgrade Propagation through many different materials (mitigated by using a simplified geometry) Propagation through B-Field (Runge-Kutta method used, maybe a fast parametrization possible?) Can we use ML to get good parameter estimates and make it faster? New York, DS@HEP, Tracking@LHCb Slide 21

22 Usage of Machine learning Relatively high fake track rate Secondary particles from material interactions No tracking stations inside B-field Neural net to reject fake tracks Used in selection of best tracks Usually tighter cut used in analyses Neural nets implemented in Forward tracking (New in 2016) In clustering fewer clusters to consider Final track selection Fewer tracks Kalman filtered. Also improves efficiency New York, Slide 22

23 Where do we want to improve? Improve fake track rejection More efficient and faster More robust against Data-MC differences Use more machine learning in pattern recognition Identification of clusters in tracking detectors Selection of final track candidates Improve and speed up track fit Needed for best track parameter estimate Kalman filter is the main time consumer in our reconstruction New York, Slide 23

Machine learning and parallelism in the reconstruction of LHCb and its upgrade

Machine learning and parallelism in the reconstruction of LHCb and its upgrade Marian Stahl on behalf of the LHCb collaboration Physikalisches Institut der Universität Heidelberg, Germany E-mail: marian.stahl@cern.ch