[PPT] - Particle Flow at 40 MHz with the CMS L1 Trigger Christian Herwig, PowerPoint Presentation

SLIDE 1

Particle Flow at 40 MHz with the CMS L1 Trigger

Christian Herwig, for the CMS L1PF Team CPAD Instrumentation Frontier Workshop December 8-10, 2019

SLIDE 2

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Outline

2

Motivation and the High-luminosity LHC
Particle Flow reconstruction
PUPPI Pileup subtraction
The Phase-II Upgrade to the L1 CMS Trigger
Progress of PF+PUPPI implementation

SLIDE 3

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

3

We are here

SLIDE 4

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

4

We are here Phase-II upgrades 10x dataset increase

SLIDE 5

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Discover Higgs!

5

SLIDE 6

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Constraints on BSM Physics (especially strongly produced)

6

Observed limits Expected limits

1

139.0 fb 1 χ ∼ Wb → 1 t ~ 1L, [ATLAS-CONF-2019-17]

1

36.1 fb 1 χ ∼ Wb → 1 t ~ / 1 χ ∼ t → 1 t ~ 0L, [1709.04183] 1 χ ∼ bff' → 1 t ~ / 1 χ ∼ Wb → 1 t ~ / 1 χ ∼ t → 1 t ~ 1L, [1711.11520] 1 χ ∼ bff' → 1 t ~ / 1 χ ∼ Wb → 1 t ~ / 1 χ ∼ t → 1 t ~ 2L, [1708.03247] 1 χ ∼ bff' → 1 t ~ monojet, [1711.03301] 1 χ ∼ t → 1 t ~ , t t [1903.07570] 1 χ ∼ c → 1 t ~ c0L, [1805.01649] 1 χ ∼ c → 1 t ~ monojet, [1711.03301]

1

= 8 TeV, 20 fb s Run 1, [1506.08616]

200 300 400 500 600 700 800 9001000 ) [GeV]

1

t ~ m( 100 200 300 400 500 600 700 ) [GeV]

1

χ ∼ m(

) = 0 1 χ ∼ , 1 t ~ m( Δ W + m b ) = m 1 χ ∼ , 1 t ~ m( Δ t ) = m 1 χ ∼ , 1 t ~ m( Δ

1

= 13 TeV, 36.1-139 fb s July 2019 ATLAS Preliminary

production

1

t ~

1

t ~ Limits at 95% CL

200 210 220 230 30 40 50 60 70

SLIDE 7

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

7

SM hh Rare+Exotic Higgs EWK BSM

m(NLSP)

200 400 600 800 1000 1200 1400

m(NLSP, LSP) [GeV] Δ

1 10

2

10

Higgsino-like EWK processes

HL-LHC 3/ab, 14 TeV (soft-lepton A) HL-LHC 3/ab, 14 TeV (soft-lepton B) HE-LHC 15/ab, 27 TeV (soft-lepton B) FCC-hh (HE-LHC approx. rescaling) , 0.5/ab 500 ILC , 1/ab 1000 ILC 380 / FCC-ee 380 CLIC , 2.5/ab 1500 CLIC , 5/ab 3000 CLIC HL-LHC monojet LHeC monojet-like (proj) HE-LHC monojet FCC-eh monojet-like FCC-hh monojet m(NLSP,LSP) not displayed Δ Monojet reach in CLIC: extrapolated below 5 GeV

threshold [GeV]

T

Minimum jet p

45 50 55 60 65 70 75 80

Loss in signal significance [%]

5 10 15 20 25 30 (14 TeV)

1

3000 fb

CMS Phase-2

Simulation Preliminary

b b b b → HH

(GeV)

miss T

Minimum threshold on E

150 200 250 300 350 400 (%)

SM

σ inv)/ → B(H × σ 95% CL upper limit on 5 10 15 20 25 30

1

= 300 fb

data

L

1

= 1000 fb

data

L

1

= 3000 fb

data

L

HL-LHC 14 TeV

CMS Phase-2

Simulation Preliminary

SLIDE 8

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

8

40 mhz 35 pp/event 100 khz (400x rej) 1 khz (100x rej) Typically limited to information from a single sub-detector (calorimeter, muons) L1 HLT

SLIDE 9

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

9

40 mhz 200 pp/event 750 khz (50x rej) 7.5 khz (100x rej) L1 HLT Naively scales with luminosity

SLIDE 10

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Challenges to Phase-II L1 Trigger

10

L1 Accept rate scales ~ linearly with luminosity increase
Must maintain performance in hostile environment!

SLIDE 11

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Challenges to Phase-II L1 Trigger

11

L1 Accept rate scales ~ linearly with luminosity increase
Must maintain performance in hostile environment!

[GeV] 5 10 15 20 25 30 Normalized entries 0.02 0.04 0.06 0.08 0.1 0.12 0.14

= 6

PV

N = 10

PV

N = 14

PV

N = 18

PV

N

ATLAS Simulation 20 ≤ < 21

Pythia8 dijets, √s = 8 TeV from LCW topo-clusters

Take hh production in 4b (or bbττ) decay mode

Higher pileup → Extra stochastic energy enters into the jet cone More low-pT jets to "measure high" than vice versa → Higher trigger rate

SLIDE 12

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Challenges to Phase-II L1 Trigger

12

L1 Accept rate scales ~ linearly with luminosity increase
Must maintain performance in hostile environment!

It gets worse !! Background (uncorrelated coincidences) ~ (lumi)2 beamspot "cigar"~30cm Not new problems, solved offline with Particle Flow Reco+

SLIDE 13

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

13

Tracks Muon segments HCal

Muons Electrons

(Isolated) photons

Charged hadrons Neutral hadrons

ECal

Particle Flow Reconstruction

Idea: combine measurements across all sub-detectors

to achieve best possible resolution per object

Algorithm returns a list of single-particle candidates

SLIDE 14

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

14

Tracks Muon segments HCal

Muons Electrons

(Isolated) photons

Charged hadrons Neutral hadrons

ECal

Particle Flow Reconstruction

Idea: combine measurements across all sub-detectors

to achieve best possible resolution per object

Algorithm returns a list of single-particle candidates

(GeV)

Ref T

p

20 100 200 1000

Energy resolution

0.2 0.4 0.6

CMS

Simulation

Calo PF , R = 0.4

T

Anti-k | < 1.3

Ref

η |

(GeV)

miss T,Ref

p

50 100 150 200 250

resolution

miss T

Relative p

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Calo PF

CMS

Simulation

improved jet pT resolution improved missing pT resolution

Improved Jet pT resolution Improved pT-miss resolution

SLIDE 15

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Pileup Per Particle Identification

15

Weight

weight>0.01

N/ N Δ

5 −

10

4 −

10

3 −

10

2 −

10

1 −

10 1 10

2

10

3

10

Neutral Particles Data MC

Weight

0.2 0.4 0.6 0.8 1 Data/MC 1 2

(13 TeV)

1

0.36 fb

CMS

(13 TeV)

1

0.36 fb

CMS

Preliminary

C i

α

5

5 10 15

fraction of particles

0.02 0.04 0.06

charged LV charged PU neutrals LV neutrals PU

Pileup Leading Vertex

Idea: get probability that a neutral PF candidate is pileup

based on local activity from the leading vertex

α ∼ X

i∈cone

pT,i ∆Ri

1407.6013

SLIDE 16

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Pileup Per Particle Identification

16

Idea: get probability that a neutral PF candidate is pileup

based on local activity from the leading vertex

5 10 15 20 25 30 35 40 45 50

5 10 15 20 25 30 35 40

ee → Z

miss T

PF p ee → Z

miss T

PUPPI p Uncertainty

CMS

Preliminary

Response-corrected

(13 TeV)

1

35.9 fb ) [GeV] ( u σ

5 10 15 20 25 30 35 40 45 50

Number of vertices

0.6 0.7

JME-18-001

Improved pT-miss resolution

SLIDE 17

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Architecture of the Phase-II L1 Trigger

17

SLIDE 18

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Architecture of the Phase-II L1 Trigger

18

vertices 2-3 GeV tracks |η|<2.5 9 ɸ sectors

SLIDE 19

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Architecture of the Phase-II L1 Trigger

19

Layer 2: Algorithms using PF+PUPPI inputs Layer 1: Run the PF+PUPPI algorithm itself

SLIDE 20

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

20

Take advantage of the inherent locality of PF+PUPPI
Distribute computation across many processing units
Processing is divided into three main steps:
Regionalization (VHDL)
PF+PUPPI calculation (High Level Synthesis C++)
Algorithms using PF+PUPPI inputs (HLS C++)
HLS: no expertise required!
Fast prototyping, debugging, comparison of alg variants

Strategy for L1 Implementation

Layer 1 Layer 2

SLIDE 21

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Inputs versus η, PF+PUPPI regions

21

TMUX 18→6

SLIDE 22

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Regionizer validation

22

VHDL algorithm validated with simulated data inputs

Input link index 10 20 30 40 50 60 70 80 90

# objects 10 20 30 40 50 60 70 80

CMS Internal

Tracks EM calo Calo Muons

Region index 2 4 6 8 10 12 14 16

# objects 20 40 60 80 100

CMS Internal

Simulation Tracks EM calo Emulation Calo Muons

ttbar events μ~200

100% match!

SLIDE 23

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

23

Regionalization → small # of objects to link (truncation)
Cluster input pre-processing: exploit shapes
PUPPI 'linearized'; smaller cone size

Work in Progress

Classify cluster:
Hadronic or EM-like?
Remove pileup deposits
Less work for PUPPI!

HW Particle Flow + PUPPI

SLIDE 24

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

24

Many ΔR calculations for track-calo linking drives DSP
Scales as (#tracks)*(#calo clusters)
PUPPI weights drive BRAM usage
To compute pT/ΔR quickly requires division tables
DSPs also used to map (pT, ΔR) → PUPPI weights

Resource drivers

Resource LUT FF BRAM DSP Usage 528k 785k 871 1020 % VU9P 45% 33% 40% 15%

PF+PUPPI resources for 22 tracks, 15+13 calo clusters

SLIDE 25

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Regionalization schemes

25

Resources vs. various initiation intervals and region sizes

SLIDE 26

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Hardware Prototype

26

PF+ PUPPI

Regional sorting Link infra ATCA carrier card development lead by APx consortium Placed preliminary algorithm on VU9P

SLIDE 27

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Layer 2 algorithms — Jets and MET

27

100 200 300 400 500

(GeV)

miss T

Generator E

0.2 0.4 0.6 0.8 1

Eff (L1 rate 50 kHz)

PU 200 (14 TeV)

CMS Phase-2 Simulation

t Signal: t

miss T

E

miss T

E Calo > 430 GeV tk Δz > 58 GeV Puppi > 90 GeV

miss T

E

500 1000 1500 2000

Generator m(jj)

0.2 0.4 0.6 0.8 1

Eff (L1 rate 20 kHz)

PU 200 (14 TeV)

CMS Phase-2 Simulation

VBF selection invisible → Signal: VBF H HF threshold = 0 GeV HF threshold = 15 GeV HF threshold = 50 GeV

Work in Progress Work in Progress

Use PF+PUPPI candidates to build jets, energy sums

SLIDE 28

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Layer 2 algorithms — Tau ID NN

28

0.1 0.2 0.3 0.4 0.5 0.6 0.7 > 20 GeV)

Gen T

min p

h

τ (

s

∈ 10

2

10 Trigger Rate(kHz)

h

τ Di- NN NN+Puppi CMS Phase-2 Simulation PU 200 (14 TeV)

Identify hadronic tau decays using PF+PUPPI candidates

LUT FF DSP Latency 90k 150k 1400 210ns 7% 6% 20%

Work in Progress

Inputs: 10 nearby PF

candidates (pT,η,ɸ,id)

Dense w/ 3 hidden layers

(25,25,10) → 1 MVA ID

This implementation:
Up to 18 PF+PUPPI

candidates / event A proof-of-principle prototype Developed using hls4ml

SLIDE 29

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Layer 2 algorithms — Tau ID NN

0.1 0.2 0.3 0.4 0.5 0.6 0.7 > 20 GeV)

Gen T

min p

h

τ (

s

∈ 10

2

10 Trigger Rate(kHz)

h

τ Di- NN NN+Puppi CMS Phase-2 Simulation PU 200 (14 TeV)

Identify hadronic tau decays using PF+PUPPI candidates

Work in Progress

A proof-of-principle prototype Developed using hls4ml

LUT FF DSP Latency 90k 150k 1400 210ns 7% 6% 20%

Inputs: 10 nearby PF

candidates (pT,η,ɸ,id)

Dense w/ 3 hidden layers

(25,25,10) → 1 MVA ID

This implementation:
Up to 18 PF+PUPPI

candidates / event See hls4ml talk / Sergo + L1 Muon / Jia Fu + ML trigger / Zhenbin

SLIDE 30

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Conclusion

30

The Level-1 Particle Flow Trigger is an ambitious addition

to the Phase-II upgrade

Correlation of all major sub-detectors allows

unprecedented event reconstruction at 40mhz

Capability promises to significantly enhance CMS

sensitivity to interesting weak-scale physics

SLIDE 31

Backup

SLIDE 32

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

32

Idea: combine measurements across all sub-detectors to

achieve best possible resolution per object

Algorithm returns a list of single-particle candidates

Tracks Muon segments HCal

Muons Electrons

(Isolated) photons

Charged hadrons Neutral hadrons

ECal

The full story is a bit more complicated…

Kinked tracks, brem recovery Cluster splitting Secondary vertices

+

Fake tracks "Indirect linking"

+

Particle Flow Reconstruction

SLIDE 33

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Pileup Per Particle Identification

33

Idea: assign a probability that a neutral PF candidate is

pileup based on local activity from the leading vertex

Discriminant favor nearby, high-pT particles (in cone)
QCD is collinear, while pileup is diffuse

αi = log

∑

j6=i,∆Rij<R0

pTj ∆Rij !2

∆R2

ij = ∆η2 ij + ∆φ2 ij

Weight

weight>0.01

N/ N Δ

5 −

10

4 −

10

3 −

10

2 −

10

1 −

10 1 10

2

10

3

10

Neutral Particles Data MC

Weight

0.2 0.4 0.6 0.8 1 Data/MC 1 2

(13 TeV)

1

0.36 fb

CMS

(13 TeV)

1

0.36 fb

CMS

Preliminary

C i

α

5

5 10 15

fraction of particles

0.02 0.04 0.06

charged LV charged PU neutrals LV neutrals PU

sum over nearby charged particles

Pileup Leading Vertex

SLIDE 34

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Pileup Per Particle Identification

34

Idea: assign a probability that a neutral PF candidate is

pileup based on local activity from the leading vertex

Discriminant favor nearby, high-pT particles (in cone)
QCD is collinear, while pileup is diffuse

Weight

weight>0.01

N/ N Δ

5 −

10

4 −

10

3 −

10

2 −

10

1 −

10 1 10

2

10

3

10

Neutral Particles Data MC

Weight

0.2 0.4 0.6 0.8 1 Data/MC 1 2

(13 TeV)

1

0.36 fb

CMS

(13 TeV)

1

0.36 fb

CMS

Preliminary

Compare α w/ expected distribution, given the level of pileup (chi2 test)

btain weights!

Re-scale 4-vectors: 50 GeV particle w/ 0.4 PUPPI weight considered as a 20 GeV particle

SLIDE 35

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Latency budget

35

Upstream (+4 ns)

Layer 2

SLIDE 36

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Architecture of the Phase-II L1 Trigger

36

Calorimeter clusters in 3 regions: High-granularity endcap calorimeter (see Z. Gecse's talk)

SLIDE 37

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Architecture of the Phase-II L1 Trigger

37

"Standalone" muons "Global" muons

SLIDE 38

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

38

TM18 TM18 TM1 TM18 L1: outputs @ TM6, 6 eta regions L2: outputs @ TM6, 1 eta region

Architecture of the Phase-II L1 Trigger

SLIDE 39

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

FPGA / TMUX / Region view

39

SLIDE 40

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Firmware - Regionalization

40

Regionizer

Nlinks in

(different upstream detector)

PF Block

Tracker Calo Muon

M(=5) buffers hold N(=25) tracks for each small region Each set of M holds B(=2) small regions

}

PF region 1

}

PF region N

}

PF region 1

}

PF region 1

PF region 1 PF region N

PF regions streamed in

“multi-tap shift registers array”

Nlinks BRAMs ~15-20 clocks

(needed to avoid clashes into same region and deal with overlaps)

405 URAMs 5+3 clocks From links ~3 clocks Fabric (LUT) resources here ~15% of VU9P

SLIDE 41

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

PF+PUPPI algo

41

SLIDE 42

Dec. 9, 2019
C. Herwig — CPAD Instrumentation Frontier Workshop

Regionalization — 'board regions’

42