Background Suppression with Trigger Neuro Team Suppression Simulation Background Algorithm NeuroTrigger Belle II the Belle II Neural Network Introduction Outline Mar 19, 2018 Max-Planck-Institut für Physik Sebastian Skambraks Trigger S. Bähr, C. Kiesling, S. Pohl, S. Skambraks
Introduction - Belle II at SuperKEKB u kikou Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) located in Tsukuba, Japan at KEK High Energy Accelerator Research Organization 2/ 14 高エネルギー加速器研究機構 K ¯ o E nerug ¯ ı K asokuki kenky ¯ e + 4 GeV • asymmetric e + e − collider • Υ (4S) resonance B 0 B 0 / B + B − e − 7 GeV � • L = 8 × 10 35 cm − 2 s − 1 (40 × KEKB) • average p T : 500 MeV • average track multiplicity: 11
Introduction - The Belle II Detector Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) 3/ 14 e − e +
Introduction - The Belle II Detector Central Drift Chamber 56 layers Input for L1 Track Trigger Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) 3/ 14 e − e +
Introduction - Belle II Background z Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) Belle Beam Background Tracks NeuroTrigger Goals - beam gas - radiative Bhabha back scatters - Touschek efgect 4/ 14 • reject tracks from z � = 0 cm background s c i • single track z -vertex resolution < 2 cm s y h p • latency < 1 µs e − e + Z distribution # of events / 5 mm 1000 • tracks generated at the beam-line & -wall with vertices z � = 0 cm 800 • increase with luminosity 600 • main processes: 400 200 0 -40 -30 -20 -10 0 10 20 30 40 z (cm) ⇒ need z vertex reconstruction at 1 st trigger level
Introduction - Belle II First Level Trigger CDC ECL KLM PID GDL 30 kHz tracking 5 µs Requirements Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) 5/ 14 • 30 kHz trigger rate • 5 µs latency ⇒ deadtime-free pipelined operation
Introduction - Belle II First Level Trigger CDC Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) Track Segment Finder (3D Tracks) Neural Network (2D Tracks) Hough Transformation GDL 3. Tracker 2. Finder 1. TSF CDC Trigger Pipeline CDC Requirements 5 µs tracking 30 kHz GDL PID KLM ECL 5/ 14 • 30 kHz trigger rate • 5 µs latency ⇒ deadtime-free pipelined operation
Introduction - CDC Trigger SL Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) Stereo SL confjguration -68.5 – -74.0 8 63.1 – 70.0 6 -55.3 – -64.3 4 45.4 – 45.8 2 angle (mrad) 6/ 14 5 axial stereo layer axial layer z layers super 4 stereo layers super ≈ 2 . 4 m ≈ 1 . 2 m ≈ 16 cm • 56 layers combined to 9 super layers (SL) • 2336 track segments (TS) in 9 SL
Introduction - CDC Trigger angle (mrad) Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) information of TS priority wires NeuroTrigger Input Track Segment Stereo SL confjguration -68.5 – -74.0 8 63.1 – 70.0 6 -55.3 – -64.3 4 45.4 – 45.8 5 axial 2 SL super super layers 4 stereo layers axial layer z stereo layer 6/ 14 ≈ 2 . 4 m ≈ 1 . 2 m ≈ 16 cm • 56 layers combined to 9 super layers (SL) • 2336 track segments (TS) in 9 SL ≈ 15 mm • position, drift time and left/right • 2D track estimates ( p T , ϕ )
Introduction - CDC Trigger x [mm] Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) (nonlinear) xt – relation t [ns] 300 150 10 0 -10 7/ 14 • axial layers • stereo layers • Υ( 4 S ) event • background noise • track segments (TS)
Introduction - CDC Trigger x [mm] Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) (nonlinear) xt – relation t [ns] 300 150 10 0 -10 7/ 14 • axial layers • stereo layers • Υ( 4 S ) Event • background noise • rack segments (TS)
Introduction - CDC Trigger x [mm] Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) (nonlinear) xt – relation t [ns] 300 150 10 0 -10 7/ 14 • axial layers • stereo layers • Υ( 4 S ) Event • background noise • track segments (TS)
Introduction - CDC Trigger x [mm] Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) (nonlinear) xt – relation t [ns] 300 150 10 0 -10 7/ 14 • axial layers • stereo layers • Υ( 4 S ) Event • background noise • track segments (TS)
NeuroTrigger - Multi Layer Perceptron sl . . Properties w ji t sl sl t sl sl t sl t sl . sl t sl sl input layer hidden layer output layer Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) . w kj . i Training i z True . Net i input one TS Hit per SL per track 8/ 14 z • robust function approximator • massively parallel processing • short deterministic runtime ϕ rel ϕ rel ϕ rel ϕ rel ϕ rel • neuron: y = tanh ( w i x i + w 0 ) • network: z k = f ( w kj f ( w ji x i )) α sl α sl α sl α sl α sl θ � � 2 • minimize � − z • RPROP (backpropagation) (position ϕ rel , α and time t ) output z , θ estimate
NeuroTrigger - Input Representation drift time t Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) r 2 D 2D arc length to TS TS position relative to 2D track arc length s axial 9/ 14 axial stereo V axial stereo U axial stereo V axial stereo U ϕ rel ϕ rel : • use track estimates provided by 2D fjnder • 3 inputs per SL, values: ( t , ϕ rel , α ) α : • dedicated networks for missing hits
Background Simulation 10 Brems 1 kHz Luminosity Machine Triggered Particles background Track Multiplicity 1 10 Coulomb BhabhaS BhabhaM BhabhaL 3 Bhabha cases (In the dominating t-channel, the Bhabha cross section strongly depends on the scattering angle) Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) 15 kHz 10/ 14 2 kHz 2D trigger Bhabha M Bhabha S 6 kHz TwoPhoton 20 kHz rate 52 kHz Bhabha L 26 kHz process Touschek intra bunch scatt. 60 e + e − e + e − e + e − → ���� 40 kHz e + e − γγ 20 e + e − → e + e − γ 0 e + + e p other particles Touschek Brems BhabhaL TwoPhoton Coulomb BhabhaM BhabhaS e ± N → e ± N e ± N → e ± N γ θ e + [ ◦ ] 120 100 80 kHz 60 40 20 0 . 5 0 1 2 3 tracks θ e − [ ◦ ] 0 . 51
Background - Material Scattering Initial Bkg Particles before Scattering Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) from the beam pipe only from the IP 2D trigger bkg particles 11/ 14 30 Touschek Brems BhabhaL TwoPhoton 25 Coulomb BhabhaM BhabhaS 20 r / cm 15 10 5 0 60 rate: 121.5 kHz 150 100 50 0 50 100 150 z / cm • primary generated 40 kHz • only events with a 20 • luminosity bkg • machine bkg 0 150 100 50 0 50 100 150 z / cm
Background - Material Scattering Tracks seen in the Trigger Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) 2D trigger tracks detector simulation 11/ 14 30 Touschek Brems BhabhaL TwoPhoton 25 Coulomb BhabhaM BhabhaS 20 r / cm 15 10 25 5 0 rate: 121.5 kHz 150 100 50 0 50 100 150 z / cm 20 15 • particles after kHz 10 • bkg particles matched to 5 ≈ 80 kHz reducible ( z � = 0) ≈ 40 kHz irreducible ( z = 0) 0 150 100 50 0 50 100 150 z / cm
Background - Material Scattering Reducible Background Tracks Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) 2D trigger tracks detector simulation 11/ 14 30 Touschek Brems BhabhaL TwoPhoton 25 Coulomb BhabhaM BhabhaS 20 r / cm 15 10 5 0 rate: 81.4 kHz 150 100 50 0 50 100 150 8 z / cm 6 • particles after kHz 4 • bkg particles matched to 2 ≈ 80 kHz reducible ( z � = 0) ≈ 40 kHz irreducible ( z = 0) 0 150 100 50 0 50 100 150 z / cm
Background - Reconstruction Neural Network Track Estimates Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) with the neural network 12/ 14 100 Touschek Brems BhabhaL TwoPhoton Coulomb BhabhaM BhabhaS 50 z NN / cm 0 50 100 12.5 rate: 76.2 kHz 100 50 0 50 100 z MC / cm 10.0 • 3D reconstructed bkg 7.5 kHz 5.0 • neuro z range limited to [ − 50 , 50 ] cm 2.5 0.0 150 100 50 0 50 100 150 z / cm
Background - Suppression Z Cut (Tracks not from IP) Background Suppression with the Belle II Neural Network Trigger (Sebastian Skambraks) neural network z after a cut on the 13/ 14 Touschek Brems BhabhaL TwoPhoton 12.5 rate: 76.2 kHz Coulomb BhabhaM BhabhaS 10.0 7.5 kHz 5.0 120 2.5 0.0 150 100 50 0 50 100 150 100 z / cm 80 • only tracks with | z MC | ≥ 1 cm kHz 60 • cumulative bkg rate 40 20 • z cut is varied in 5 cm steps 0 0 10 20 30 40 50 z cut / cm
Recommend
More recommend
