Particle Flow and PUPPI in the Level-1 Trigger at CMS for the HL-LHC

Benjamin Kreis1,∗ for the CMS Collaboration 1Fermi National Accelerator Laboratory, Batavia, IL 60510, USA

Abstract. With the planned addition of tracking information to the Compact Muon Solenoid (CMS) Level-1 trigger for the High-Luminosity Large Hadron Collider (HL-LHC), the trigger algorithms can be completely reconceptualized. We explore the feasibility of using particle flow-like reconstruction and pileup per particle identification (PUPPI) pileup mitigation at the hardware trigger level. This represents a new type of multi-subdetector pattern recognition chal- lenge for the HL-LHC. We present proof-of-principle studies on both physics and hardware-resource performance of a prototype algorithm for use by CMS in the HL-LHC era.

1 Introduction

In the mid-2020s, the High-Luminosity Large Hadron Collider (HL-LHC) at CERN will begin colliding protons with an instantaneous luminosity leveled to 5 × 1034 cm−2s−1 with the goal of delivering approximately 300 fb−1 per year for ten years [1]. This increase in luminosity will enable the general-purpose experiments at the HL-LHC, CMS and ATLAS, to maintain progress in their rich physics programs of precision measurements of the standard model and searches for physics beyond the standard model, such as supersymmetry and dark matter. The challenge for the HL-LHC experiments will be to maintain detection efficiency for interesting physics processes at the electroweak energy scale occurring in the same bunch crossings as additional inelastic interactions, known as pileup. Pileup produces extra hits in the tracking subdetectors and extra energy depositions in the calorimeter subdetectors that must be separated from the process of interest. In addition, the subdetectors must be able to withstand the radiation damage caused by the particles produced by these many proton interactions. At the HL-LHC, the average number of pileup interactions is expected to reach 200. The CMS experiment [2] will undergo a comprehensive upgrade to maintain performance in the high-luminosity conditions of the HL-LHC. In this paper, we focus on upgrades to the Level-1 trigger, the first stage of data processing that has a coarse, global view of the detector and determines which collision events to fully read out for further analysis. We explore the feasibility of using particle-level reconstruction and pileup mitigation, the best performing algorithms in current offline processing, in the upgraded trigger. In Section 2, we describe the inputs to the upgraded Level-1 trigger and its architecture. The inputs will

∗e-mail: [email protected] Particle Flow and PUPPI in the Level-1 Trigger at CMS for the HL-LHC Benjamin Kreis, Fermi National Accelerator Laboratory, Batavia, IL for the CMS Collaboration [email protected] Abstract include tracks forWith the the first planned time, and addition the trigger of the will tracking have a correlatorinformationthat in formsthe Level the global1 trigger in CMS for the HL-LHC, the algorithms for Level 1 trigger can view of each collisionbe completely event. Particle-level reconceptualized. reconstruction Following and the pileup example mitigation for offline algorithms reconstruction in CMS to use complementary subsystem information and for the correlator are described in Section 3. In Section 4, results on physics performance and mitigate pileup, we explore the feasibility of using Particle Flow-like and pileup per particle identification (PUPPI) techniques at the uses lead, copper, and copper-tungsten alloy as absorbers while copper and steel are used for the hardware resource requirementshadronic part. Siliconbased sensors are used in on the region a with proof-of-principle the highest level of radiation, whilst study are presented. hardwareplastic trigger scintillator with on-tile level. silicon photomultipliers This are usedrepresents for the part of the CE-H where thea new type of multi-subdetector pattern recognition challenge for the HL-LHC. We present proof-of- radiation level is the lowest. The key parameters of the HGCAL are summarised in figure 1. Surfaces of about 600 m2 of silicon sensors and 500 m2 of scintillator tiles are used. The silicon principle sensorsstudies are finely segmented on into 0.both5 and 1 cm2 cells physics with active thicknesses and of 300 µm, 200 resourceµm usage performance of a prototype algorithm for use by CMS in the HL-LHC era. or 120 µm depending on the region of the detector. These cell parameters are chosen such that 2 Level-1 Trigger forthe signal-over-noise HL-LHC ratio remains high enough for the measurement of MIP signals over the full lifetime of the HGCAL. Scintillator cells of 1° and 1.25° are used, arranged in a ⌘- grid. 2018 JINST 13 C02043

CMS Detector Upgrades for HL-LHC Chapter 1. The CMS Phase-2 muon detector Particle Flow Algorithm for Level-1 Trigger 18 2.3. Overview of the upgraded tracker concept 21 Upgrades to the CMS detector for HL-LHC, described in detailη elsewhere0.1 0.2 0.3 0.4 0.5 0.6 0.7 [3,0.8 0.9 4],1.0 1.1 will bring θ° 84.3° 78.6° 73.1° 67.7° 62.5° 57.5° 52.8° 48.4° 44.3° 40.4° 36.8° η θ° 8 ) 1.2 33.5° DTs Hexagonal modules based on Si sensors m CSCs standalone muons

R ( MB4 RPCs charged hadrons radiation regions of CE-H RB4 1.3 30.5° 7 trigger primitivesGEMs linked new and improved inputs to the Level-1 trigger. The inputs, referredWheelto 1 as , “Cassettes”: multiple modules mounted on Wheel 0 Wheel 2 22 Chapter 2. Overview of the Phase-2 Tracker Upgrade iRPCs ME0 muons cooling plates with electronics and absorbers MB3 1.4 27.7° CMS-TDR-17-001 6 RB3 T(EM)

readout in ME1/3 muon/track RE1/3 RE2/3 RE3/3 RE4/3 1.5 25.2° (track) 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 T electrons neutral hadrons 1.6 MB2 track are formed by trigger primitive generators (TPGs). For the first time5 at CMS,RB2 theME2/2 Level-1ME3/2 ME4/2 1.6 22.8° trig- T 1200 linking 1.7 20.7° MB1 (closest in pT) (track), or p 1.8 RB1 T

r [mm] 1000 ~2.3m 4 1.8 18.8° RE1/2 RE2/2 RE2/2 RE3/2 RE4/2 unlinked tracks 2.0 1.9 17.0° 800 ME1/2 ger will receive trigger primitives representing tracks with transverseSolenoid magnet momentum pT2.0 15.4°> 2-3 tracks photons 2.2 3 2.1 14.0° 2.2 12.6° (EM) ~ p 600 T (track) 2.4 2.3 11.5° p ME2/1 ME3/1 ME4/1 T 2.6 HCAL RE3/1 RE4/1 2.4 10.4° T(EM) Figure 2.4: Illustration of the module concept. (a) Correlation of signals in closely-spaced GE2/1 significantly larger than p 400 pT 2.8 2 2.5 9.4° 3.0 GE1/1 due to electron+photon

sensors enables rejection of low-p particles; the channels shown in green represent the selec- ME1/1 GeV from a new outer trackingT subdetector. The TPGs of a new endcapECAL calorimeter will pro- 2.8 7.0° 200 tion window to define an accepted stub. (b) The same transverse momentum corresponds to a Steel 3.0 5.7° (calo) significantly 4.0 1 HGCAL T larger distance between the two signals at large radii
for a given sensor spacing. (c) For the end- Silicon ME0 tracker 0 4.0 2.1° cap discs, a larger spacing between the sensors is needed to achieve the same discriminating CMS-TDR-016 (track) due to linked, or unlinked low p unlinked calorimeter, 0 500 1000 1500 2000 2500 z [mm] 5.0 0.77° electromagnetic/ no track, T or p or p 13 C02043 0 larger than p vide energy depositionpower as in the barrel atinformation the same radius. at improved~2m granularity. New0 1 TPGs2 3 4 in5 6 the7 8 barrel9 10 11 12 electro- Figure 2.3: Sketch of one quarter of the tracker layout in r-z view. In the Inner Tracker the z (m) significantly larger J.-B. Sauvan 2018 JINST green lines correspond to pixel modules made of two readout chips and the yellow lines to track linking than p 473 old depends on the acceptance window, which can be tuned to a certain level by programming Figure 1.4: An R-z cross section of a quadrant of the CMS detector, including the Phase-2 up- electromagnetic electron+photon pixel modules with four readout chips.474 Inthe the respective Outer Tracker setting in the the blue readout and chip. red Forlines the representpT modules a few different values of sensor grades (RE3/1, RE4/1, GE1/1, GE2/1, ME0). The acronym iRPCs in the legend refers to the (closest in ΔR) new improved RPC chambers RE3/1 and RE4/1. The interaction point is at the lower left cor- calorimeter/track mageticthe two types ofcalorimeter modules described in the475 text.spacing are used, will optimized to also achieve the desired improvepT filtering in different regions granularity, of the de- while the TPGs of the hadron calorimeter unlinked tracks ner. The locations of the various muon stations are shown in color (MB = DT = Drift Tubes, ME calorimeter All new silicon476 tector (Fig. 2.4 (b) and (c)). tracker For a pitch of about 100 µ m betweenH silicon i gstrips (or h macro-pixels, g r a n u l a r i t y New Cathode Strip 477 as detailed below) in the transverse plane, sufficient pT resolution can be achieved down to a = CSC = Cathode Strip Chambers, RB and RE = RPC = Resistive Plate Chambers, GE and ME0 linking 478 radius of about 200 mm in a barrelFigure geometry, 1. Schematic thanks to longitudinal the 3.8 T magnetic cross field section of CMS. of the The HGCAL calorimeter and its key parameters. The = GEM = Gas Electron Multiplier). M denotes Muon, B stands for Barrel and E for Endcap. (based on pT and ΔR) Labelling details are given in Section 1.2.2. The magnet yoke is represented by the dark gray will be upgraded479 concept without is therefore applicablesilicon in the Outerchanging andscintillator Tracker, and parts limited of in the angular detector acceptancegranularity. are highlighted to about in green and dark blue respectively. Finally, the TPGs of the muon subdetec- with pixel-strip and calorimeter in endcap areas.Chambers, G a s clustering, electromagnetic+hadronic 480 h < 2.4. | | calibration 481 2.3.2 The Outer Tracker For the CE-E part, silicon sensors are sandwiched between copper-tungsten baseplates and Near the interaction region a silicon tracker, composed of an inner pixel detector surrounded subtract two-strip modules in region. CE-E measures byElectron a silicon strip detector, measures vertices andMultipliers momenta of charged particles. The elec- , tors will upgraded482 The Outer and Tracker is populated will with includemodules, implementing the L1trigger trigger functionality. primitives from new muon detectors added to the PCBs assembledpT into hexagonal modules, and mounted on both sides of a copper cooling plate. electrons/photons 483 The pT module concept relies on the fact that the strips of the top and bottom sensors of a tromagnetic calorimeter (ECAL) and the hadronic calorimeter (HCAL) are located inside the Stainless steel tubes inside the cooling plate carry two-phase CO2 to maintain the silicon sensors at a 484 module are parallel to each other. With the strip direction being parallel to the z axis in the solenoid, measuring electromagnetic and hadronic showers with lead tungstate crystals and a hadronic calorimeter scintillator-brass sampling detector, respectively. o u t e r r485 ebarrel g and i nearly o radial n in thetemperature endcaps, t h this a prevents of t30 ° theC.electromagnetic concept Each cooling of stereo plate strips is to covering be used to an azimuthal angle of 60° over theshowers, full radius of and upgraded Resistive endcap region. 486 measure the coordinate ( coordinate) in the barrel (endcaps). For this reason two versions z r each layer. Plates of lead and stainless steel of the same dimensions cover the sensors and electronics. The current silicon tracker must be replaced before the start of Phase-2, since it will suffer 487 of pT modules have been realized: modules with two strip sensors (2-strip or 2S modules) significant radiation damage by the end of Run 3. To maintain excellent track reconstruction provide 488 pairsand modules with a stripof and This a macro-pixel assemblyhits sensor is called (pixel-stripa a “cassette”, n or PS dmodules). and 14 CDetailscassettes areE are - assembled H in a m stack to e form a the CE-E.s u r e s atPlate high pileup, the granularity Chambers of both the inner pixel tracker and thefor outer tracker the will be 489 provided in Chapter 3. The strips inFor the the 2S modules CE-H part, have the a length absorber of about is made 5 cm, while of full those disks of stainless steel forming a rigid mechanical increased, by decreasing the pixel size and by shortening the strip lengths. For the first time The data flow490 in the from PS modules are about the 2.4 cm long. CMS In PS modules one subdetectors of the two sensors is segmented to the TPGs and through the Level-1 trigger is at CMS a momentum measurement will become possible within a few microseconds, and this 491 into macro-pixels of about 1.5structure mm length, in providing which the the activez coordinate elements measurement. are inserted The PS as cassettes of 60°. The 60° cassettes are split information can be used in the Level-1 (L1) trigger. The track trigger will greatly sharpen the compatible492 modules with are deployed in thetracks firstin 30 three° units layers toof the keep Outer theirhadronic Tracker, size in manageable. the radial region The of 200– first 8 layers showers. are fully instrumented with silicon The L1muonpT resolution, which willsystem. reduce the trigger rate at a given transverse momentum. Thus 493 600 mm, i.e. down to radii at which the stub pT resolution remains acceptable and the data by combining input from the tracker and muon systems the pT threshold for the single muon PUPPI Algorithm for Level-1 Trigger shown in Figure494 reduction 1. effective. The The 2S modules barrelsensors, arewhile deployed the in calorimeterrest the outermost of the layers three are layers, made in the of radial scintillator trigger tiles in the low pseudorapidity processes region triggertrigger can be kept low primitives despite the high rate at HL-LHC. to form high- 495 region above 600 mm. In the endcaps the modules are arranged in rings on disc-like structures, with pT>2 GeV atand silicon the sensors inactive the high pseudorapidity elements region. As opposed to the CE-E are part, the detectorsilicon The endcap calorimeters will also suffer significant radiation damage. The replacement planned 496 with the rings at low radii, up to about 700 mm, equipped with PS modules, while 2S modules Figure 2.4: Average number of module layers traversed by particles, including both the Inner for Phase-2, the High Granularity Calorimeter (HGCAL), will have an electromagnetic and a 497 are used at larger radii. The modulesz coordinates are provided assembled by the only three on PS one barrel side layers of the constrain copper cooling plate. tracks vertexing resolutionTracker (red) and the Outer clusters. Tracker (blue)498 modules,the origin as ofThe well the trigger as the tracks complete barrel to a portiontracker of (black). the luminousand Par- region endcap of about 1 mm, which is muon track finders identify muons. The output of ticle trajectories are approximated by straight499 sufficiently lines, using precise a flat to partially distribution discriminate of primary particles vertices coming from different vertices. full LHC collision rate. and scintillator based. primary within z0 < 70 mm, and multiple scattering is not included. the barrel| | calorimeter trigger, muon track–2– finders, and the trigger primitives of the tracking vertex The following section summarizes the main concepts and features of the upgraded tracking system. One quarter of the Phase-2 tracker layout can be seen in Fig. 2.3. Figure 2.4 shows charged pileup charged particles charged particles subdetector,the average number of active layersendcap that are traversed bycalorimeter, particles originating from the lumi- and forward hadronic calorimeter reach the correlator at a subtraction from primary vertex nousCMS region, for the complete Level-1 tracker as well as for the InnerTrigger Tracker and the Outer Tracker for HL-LHC separately. particle bandwidthThe number of layers has of been optimisedO(10) to ensure robust Tbps, tracking, i.e. basically where unaffected the information is correlated to form a global view of the performance when one detecting layer is lost in some parts of the rapidity acceptance. The six flow layersAn of the Outer all Tracker are thenew minimum required FPGA-based to ensure robust track finding at the L1 Level 1 Trigger will process data from CMS PUPPI reweighted trigger in the rapidity acceptance of h < 2.4, as discussed in more details in Section 3.1. neutral particles event. This is passed| | on to the global trigger, where the trigger decision is made. weight neutral particles subdetector trigger primitive generators (TPGs). 3.6. New Trigger Objects based3.6. onNew Particle Trigger Flow Objects Reconstruction based on Particle Flow Reconstruction 27 27 tracker calorimeters muon detectors The PUPPI weight for neutral particle i is computed using a lookup table trigger-primitive clusters ortrigger-primitive trigger-primitive clusters muons. or We trigger-primitive analyse2 the linked muons. trigger-primitive We analyse the linked trigger-primitive information to create a globalpT,j list of L1 particle candidates: muons, charged hadrons, electrons, barrel barrel forward resistive cathode gas information to create a global↵i list= of L1 particle candidates: muons,⇥(R0 chargedRij) hadrons, electrons, endcap drift addressed by the metricphotons, and neutral hadrons.Rij We then run the L1 PUPPI algorithm on the list of global particle tracker electromagnetic hadronic hadronic plate strip electron photons, and neutral hadrons. Wej thencharged run from theL1 PV PUPPI✓ algorithm◆ on the list of global particle calorimeter tubes 2 X calorimeter calorimeter calorimeter chambers chambers multipliers candidates to filter the eventcandidates into the most to filter leading, the event vertex into compatible the most particle leading, candidates. vertex compatible As particle candidates. As Algorithmwith the PF Performance algorithm, thewith PUPPI the algorithmic PF algorithm, complexity the PUPPI is reduced algorithmic and only complexity uses simplified is reduced and only uses simplified integer operations. Since PUPPIinteger requires operations. input Since information PUPPI requires from track input vertexing information but the from PF track vertexing but the PF TPG TPG TPG TPG TPG TPG TPG TPG TPG algorithm does not, such aalgorithm vertexing stepdoes is not, naturally such a performed vertexing step in parallel is naturally with the performed PF oper- in parallel with the PF oper- Performanceations. The is output studied of the combinedations. using The PF+PUPPI output simulated of trigger the combined algorithm t-tbar PF+PUPPI is a set ofevents triggervertex filtered algorithm decaying can- is a set of vertex semileptonically filtered can- didates that can then be useddidates to reconstruct that can then and beidentify used totriggerable, reconstruct prompt and identify physics triggerable, objects. prompt physics objects. as a signal with large This missing approach to transversepileup mitigation needs energy the primary (MET) vertex to be and properly scalar reconstructed, sum as of jet barrel muon endcap muon This approach to pileup mitigation needs the primary vertex to be properly reconstructed, as barrel calorimeter trigger can be done easily in events with a large multiplicity of high tracks (see Section 3.2). track finder track finder energy (HcanT be) and done easily minimum-bias in events with a large multiplicity events of highas pbackgroundT tracks (see Section 3.2).. pT

CMS Phase-2 SimulationCMS,=140Phase-2 Simulation,=140

O(10) Tbps ~ 2.5 μs Semileptonic tt Semileptonic tt correlator 10 10 Rate [MHz] Rate [MHz] global trigger 1 1

The Correlator combines all of this information into trigger objects, which are MET (Calo) MET (Calo) Figuresent 1. Schematicto the Global of the Trigger CMS Level-1 to make Trigger the forLevel-1 HL-LHC. trigger A number decision. of possible, additional direct MET (TK Δz) MET (TK Δz) −1 −1 links are not pictured (see [4]). 10 10 MET (PF+PUPPI) MET (PF+PUPPI) Particle-Level Reconstruction 0.1 0.2 0.3 0.4 0.5 0.6 0.70.1 0.80.2 0.9 0.31 0.4 0.5 0.6 0.7 0.8 0.9 1 Signal effciency Signal effciency Figure 3.8: (Left) efficiency for selecting signal and background for three different Emiss trigger miss Particle Flow reconstruction correlates detector signals to identify final-state Figure 3.8: (Left) efficiency for selecting signal and backgroundT for three different ET trigger 34 5 Performance in simulation algorithms and (Right) efficiencyalgorithms turn-on and curves (Right) for efficiency three different turn-onHT curvestrigger for working three different points: trigger working points: Secondly, in low pileup environments, there is less information available locally just due to HT Theparticles Level-1 and trigger measure systems their will use properties. Field Programmable This approach Gate Arrays provides (FPGAs) better for allFPGA aImplementation calorimeter-only algorithma calorimeter-only (purple); a track-only algorithm trigger (purple); algorithm a track-only using tracks trigger consistent algorithm using tracks consistent the lack of pileup. This means the ↵ distribution is not as well populated and the uncertainty with the primary vertex (red); and a PF+PUPPI trigger algorithm (blue). The H thresholds performance than single-detector reconstruction and is already used in the with the primary vertex (red); and a PF+PUPPI triggerT algorithm (blue). The HT thresholds data processing. Modern high performance FPGAs have high speed input/output able to were chosen so that each H trigger path corresponds to a rate of 20 kHz. The study was on PU is larger. Particle flow and PUPPIwereT chosen are so inherentlythat each HT trigger regional path corresponds algorithms. to a rate of 20 kHz. AsThe study a was proof-of- handleHigh the Level large Trigger data bandwidth and offline and reconstruction. hundreds of thousands of logic cells to do pipelined and conducted using a backgroundconducted sample using of minimum-bias a background collisions sample ofand minimum-bias a signal sample collisions from t¯t and a signal sample from t¯t Anti-kT (R=0.7), n = 80 Anti-kT (R=0.7) JHEP 10 (2014) 59 principle,simulated we divide events decaying the semileptonically, detector corresponding into η tox an ɸ average ~ 0.55 PU of 140. x 0.55 regions accepting up PU CMS-PRF-14-001 simulated events decaying semileptonically, corresponding to an average PU of 140. 0.3 0.2 LV LV CMS Anti-kT, R = 0.4 Calo CMS Anti-kT, R = 0.4 Calo parallelizedµ data processing. The correlator|η0.6| < 2.5 has a latency budget0.6 ofpT = approximately[100-200] GeV, |η| < 2.5 2.5 µs and While the L1 PF+PUPPI triggerWhile algorithm the L1 PF+PUPPI remains an trigger early algorithm prototype remainsat the time an ofearly this prototype writ- at the time of this writ- µ Simulation |ηRef| < 1.3 PF Simulation 1.6 < |ηRef| < 2.5 PF to 25 tracks and 20 calorimeter clusters each. )/pT )/pT ing, the initial performance has been studied and compared with more traditional, standalone PFlow PFlow ing, the initial performance has been studied and compared with more traditional, standalone LV LV must beneutral pipelined to the LHC collision rate of 40 MHz. 0.15 trigger algorithms based on a single detector (e.g. either calorimeter-only or tracker-only). As HCAL PFlowCHS PFlowCHS Implementedtrigger in algorithms Xilinx based Vivado on a single HLS detector v2016.4 (e.g. either calorimeter-only for an Ulstrascale+ or tracker-only). As VU9P hadron 0.2 clusters PUPPI PUPPI an early proxy of the potentialan early gains proxy expected of the by potential this approach, gains expected we study by the this missing approach, trans- we study the missing trans- 0.4 0.4 miss verse momentumFPGA, (E T four) and the regions scalar summed miss canpT over beall jets inprocessed an event, typically referred with ~40% of the FPGA verse momentum (ET ) and the scalar summed pT over all jets in an event, typically referred Detector , (pT - pT , (pT - pT 0.1 to as H . Gains are also expected for jet substructure studies for heavy-particle tagging, and photon Energy resolution Energy resolution T σ σ resources.to as TheHT. Gains latencies are also expected for for jet substructure Particle studies forFlow heavy-particle and tagging, PUPPI and are lepton isolation, though welepton leave thoseisolation, topics though for future we leave studies. those An topics example for future of the early studies. per- An example of the early per- 3 Particle-LevelParticle Flow Algorithms0.1 0.2 for Level-1 Trigger0.2 formance of the L1 PF+PUPPI algorithm, using a Phase-2 detector simulation but using tracks

fitted fitted approximatelyformance 500 of the L1ns PF+PUPPI and 100 algorithm, ns, using respectively. a Phase-2 detector simulation but using tracks ECAL 0.05 reconstructed with an offline algorithm and mocked-up to have similar performance (resolu- clusters reconstructed with an offline algorithm and mocked-up to have similar performance (resolu- tion, pT threshold of 2 GeV, etc) as that expected from the L1 track finder, is provided in Fig. 3.8. The availability of tracking information in the Level-1 trigger correlator opens up the pos- tion, pT threshold of 2 GeV, etc) as that expected frommiss the L1 track finder, is provided in Fig. 3.8. The figures show the signal vs background selection efficiency for different ET triggers, as miss 0 0 0 The figures show the signal vs background selection efficiency for different ET triggers, as 0 10020 200100 200 3001000 400 500 20 10050200 1000 100 well as the efficiency turn-on curves for different triggers, using a semi-leptonic t¯t signal Outlook well as the efficiency turn-onHT curves for different H triggers, using a semi-leptonic t¯t signal sibility of performing reconstruction and pileup mitigationRefpT (GeV) at the particle level,Ref the bestn per- miss T p (GeV) p (GeV) PU having true E , and they compare threemiss different algorithms based on standalone calori- T T T having true E , and they compare three different algorithms based on standalone calori- meter, standalone track, as well as PF+PUPPI.T The H efficiency turn-on curves show three Figure 13: Jet energy resolution as a function of pRef in the barrel (left) and in the endcap These proof-of-principlemeter, studies standalone track, indicate asT well as PF+PUPPI. the feasibility The H efficiency of turn-on performing curves show three Particle Figure 7:Jetp resolution vs. p (left) for n T= 80 for ⌘ < 2.5 and jet p resolution vs. different working-point thresholds, one for each algorithm, that are constrained to haveT a fixed Particle flow enables the use(right) of regions. particle-levelT The lines, added toT guidepileup the eye,PU correspondmitigation, to| fitted| functions including withT ad hoc the Flow reconstruction anddifferent PUPPI working-point pileup thresholds, mitigation one for each algorithm, in the that are CMS constrained HL-LHC to have a fixed Level-1 numberparametrizations. of pileup interactions (right) for jets with pT between 100 and 200 GeV for ⌘ < 2.5. PUPPI algorithm, which attempts to remove the effects of both charged | and| Trigger. Significant performance improvements over traditional trigger

neutral particles originating4.2 from Jet Shapes pileup (JHEP 10 (2014) 59). algorithms have been shown and may be extended with further R&D.

Similar to our study of pT distributions, we can study resolution and its pileup dependence for jet shapes. Here we show results for jet mass which is considered a reasonable proxy for generic0.2 jet shapes and is used in many applications0.2 such as boosted object tagging (see CMS |ηRef| < 1.3 CMS |ηRef| < 1.3 [34–36] and referencesSimulation therein).Parton flavor Simulation Parton flavor Anti-k R = 0.4, Calo Anti-k R = 0.4, PF 0.15 T 0.15 T Anti-kT (R=0.7), n = 80 ud Anti-kT (R=0.7), udn = 80 PU PU s s 0.3 0.3 pT = [100-200] GeV, |η| < 2.5c LV pT = [100-200] GeV,c |η| < 2.5 LV PFlow PFlow Connecting the Dots 2018, Seattle, WA 0.1 b PFlowCHS 0.1 b PFlowCHS PUPPI PUPPI

0.2 0.2 fraction of jets 0.05 0.05fraction of jets Difference to gluon response Difference to gluon response 0.1 0 0 0.1 20 100 200 1000 20 100 200 1000 pRef (GeV) pRef (GeV) T T 0 0 Figure0 14: Absolute difference 50 in jet energy 100 response between0 quark and gluon 50 jets as a function 100 Ref of pT for Calo jets (left) and PF jets mass (right). (GeV) trimmed mass (GeV)

Figure 8: The single jet mass resolution for nPU = 80 for jets with 100 GeV

First we look at jet mass for central jets with 100 GeV

– 15 – Particle Flow and PUPPI in the Level-1 Trigger at CMS for the HL-LHC Benjamin Kreis, Fermi National Accelerator Laboratory, Batavia, IL forming approach in CMS’s software-level trigger and offline reconstruction [5, 6]. To per- form these algorithms in the correlator, they must be implemented in FPGAs and meet the for the CMS Collaborationpipelining and latency constraints. [email protected] Abstract 3.1 Particle-Flow Reconstruction in Level-1 Trigger With the planned addition of the tracking information in the Level 1 trigger in CMS for the HL-LHC, the algorithms for Level 1 trigger can be completely reconceptualized. Following the example for offline reconstruction in CMS to use complementary subsystem information and mitigate pileup, we explore the feasibility of using Particle Flow-like andParticle-flow pileup per reconstruction particle identification correlates tracks (PUPPI) from the tracking techniques and muon at the subdetectors and uses lead, copper, and copper-tungsten alloy as absorbers while copper and steel are used for the clusters from the calorimeter subdetectors to identify each final-state particle and combines hadronic part. Silicon sensors are used in the region with the highest level of radiation, whilst hardwareplastic trigger scintillator with on-tile level. silicon photomultipliers This are usedrepresents for the part of the CE-H where thea new type of multi-subdetector patternthe recognition subdetector measurements challenge tofor reconstruct the HL-LHC. the identified We present particles’ proof-of- properties [5]. Here radiation level is the lowest. The key parameters of the HGCAL are summarised in figure 1. Surfaces of about 600 m2 of silicon sensors and 500 m2 of scintillator tiles are used. The silicon principle sensorsstudies are finely segmented on into 0.both5 and 1 cm2 cells physics with active thicknesses and of 300 µm, 200 resourceµm usage performance of a prototypewe describe algorithm a proof-of-principle for use by CMS algorithm in the for performingHL-LHC era. particle-flow reconstruction in the or 120 µm depending on the region of the detector. These cell parameters are chosen such that the signal-over-noise ratio remains high enough for the measurement of MIP signals over the full FPGAs of the Level-1 trigger correlator. A schematic of the algorithm is shown in Figure 2. lifetime of the HGCAL. Scintillator cells of 1° and 1.25° are used, arranged in a ⌘- grid. 2018 JINST 13 C02043

CMS Detector Upgrades for HL-LHC Chapter 1. The CMS Phase-2 muon detector Particle Flow Algorithm for Level-1 Trigger 18 2.3. Overview of the upgraded tracker concept 21 η 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 θ° 84.3° 78.6° 73.1° 67.7° 62.5° 57.5° 52.8° 48.4° 44.3° 40.4° 36.8° η θ° 8 ) 1.2 33.5° DTs Hexagonal modules based on Si sensors m CSCs standalone muons

R ( MB4 RPCs charged hadrons radiation regions of CE-H RB4 1.3 30.5° 7 GEMs linked “Cassettes”: multiple modules mounted on Wheel 0 Wheel 1 Wheel 2 22 Chapter 2. Overview of the Phase-2 Tracker Upgrade iRPCs ME0 muons cooling plates with electronics and absorbers MB3 1.4 27.7° CMS-TDR-17-001 6 RB3 T(EM)

readout in ME1/3 muon/track RE1/3 RE2/3 RE3/3 RE4/3 1.5 25.2° (track) 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 T electrons neutral hadrons 1.6 MB2 track 1200 5 RB2 ME2/2 ME3/2 ME4/2 1.6 22.8° linking T 1.7 20.7° MB1 (closest in pT) (track), or p 1.8 RB1 T

r [mm] 1000 ~2.3m 4 1.8 18.8° RE1/2 RE2/2 RE2/2 RE3/2 RE4/2 unlinked tracks 2.0 1.9 17.0° 800 ME1/2 Solenoid magnet 2.0 15.4° tracks photons 2.2 3 2.1 14.0° 2.2 12.6° (EM) ~ p 600 T (track) 2.4 2.3 11.5° p ME2/1 ME3/1 ME4/1 T 2.6 HCAL RE3/1 RE4/1 2.4 10.4° T(EM) Figure 2.4: Illustration of the module concept. (a) Correlation of signals in closely-spaced GE2/1 significantly larger than p 400 pT 2.8 2 2.5 9.4° 3.0 GE1/1 due to electron+photon sensors enables rejection of low-pT particles; the channels shown in green represent the selec- ME1/1 ECAL 2.8 7.0° 200 tion window to define an accepted stub. (b) The same transverse momentum corresponds to a Steel 3.0 5.7° (calo) significantly 4.0 1 HGCAL T larger distance between the two signals at large radii
for a given sensor spacing. (c) For the end- Silicon ME0 tracker 0 4.0 2.1° cap discs, a larger spacing between the sensors is needed to achieve the same discriminating CMS-TDR-016 (track) due to linked, or unlinked low p unlinked calorimeter, 0 500 1000 1500 2000 2500 z [mm] 5.0 0.77° electromagnetic/ no track, T or p or p power as in the barrel at the same radius. 13 C02043 0 larger than p Figure 2.3: Sketch of one quarter of the tracker layout in r-z view. In the Inner Tracker the ~2m 0 1 2 3 4 5 6 7 8 9 10 11 12 z (m) significantly larger J.-B. Sauvan 2018 JINST green lines correspond to pixel modules made of two readout chips and the yellow lines to track linking than p 473 old depends on the acceptance window, which can be tuned to a certain level by programming Figure 1.4: An R-z cross section of a quadrant of the CMS detector, including the Phase-2 up- electromagnetic electron+photon pixel modules with four readout chips.474 Inthe the respective Outer Tracker setting in the the blue readout and chip. red Forlines the representpT modules a few different values of sensor grades (RE3/1, RE4/1, GE1/1, GE2/1, ME0). The acronym iRPCs in the legend refers to the (closest in ΔR) the two types of modules described in the475 text.spacing are used, optimized to achieve the desired pT filtering in different regions of the de- new improved RPC chambers RE3/1 and RE4/1. The interaction point is at the lower left cor- unlinked tracks calorimeter/track ner. The locations of the various muon stations are shown in color (MB = DT = Drift Tubes, ME calorimeter All new silicon476 tector (Fig. 2.4 (b) and (c)). tracker For a pitch of about 100 µ m betweenH silicon i gstrips (or h macro-pixels, g r a n u l a r i t y New Cathode Strip 477 as detailed below) in the transverse plane, sufficient pT resolution can be achieved down to a = CSC = Cathode Strip Chambers, RB and RE = RPC = Resistive Plate Chambers, GE and ME0 linking 478 radius of about 200 mm in a barrelFigure geometry, 1. Schematic thanks to longitudinal the 3.8 T magnetic cross field section of CMS. of the The HGCAL calorimeter and its key parameters. The = GEM = Gas Electron Multiplier). M denotes Muon, B stands for Barrel and E for Endcap. (based on pT and ΔR) Labelling details are given in Section 1.2.2. The magnet yoke is represented by the dark gray 479 concept is therefore applicablesilicon in the Outer andscintillator Tracker, and parts limited of in the angular detector acceptance are highlighted to about in green and dark blue respectively. with pixel-strip and calorimeter in endcap areas.Chambers, G a s clustering, electromagnetic+hadronic 480 h < 2.4. | | calibration 481 2.3.2 The Outer Tracker For the CE-E part, silicon sensors are sandwiched between copper-tungsten baseplates and Near the interaction region a silicon tracker, composed of an inner pixel detector surrounded subtract two-strip modules in region. CE-E measures byElectron a silicon strip detector, measures vertices andMultipliers momenta of charged particles. The elec- , 482 The Outer Tracker is populatedPCBs with assembledpT modules,implementing into hexagonal the L1 modules, trigger functionality. and mounted on both sides of a copper cooling plate. electrons/photons 483 The pT module concept relies on the fact that the strips of the top and bottom sensors of a tromagnetic calorimeter (ECAL) and the hadronic calorimeter (HCAL) are located inside the Stainless steel tubes inside the cooling plate carry two-phase CO2 to maintain the silicon sensors at a 484 module are parallel to each other. With the strip direction being parallel to the z axis in the solenoid, measuring electromagnetic and hadronic showers with lead tungstate crystals and a hadronic calorimeter 485 barrel and nearly radial in thetemperature endcaps, this prevents of 30° theC. concept Each cooling of stereo plate strips is to covering be used to an azimuthal angle of 60° over the full radius of scintillator-brass sampling detector, respectively. o u t e r r e g i o n t h a t electromagnetic showers, and upgraded Resistive 486 measure the z coordinate (r coordinate)each layer. in the Plates barrel of (endcaps). lead and stainless For this reason steel two of the versions same dimensions cover the sensors and electronics. The current silicon tracker must be replaced before the start of Phase-2, since it will suffer 487 of pT modules have been realized: modules with two strip sensors (2-strip or 2S modules) significant radiation damage by the end of Run 3. To maintain excellent track reconstruction provide 488 pairsand modules with a strip of and This a macro-pixel assemblyhits sensor is called (pixel-stripa a “cassette”, n or PS dmodules). and 14 CDetailscassettes areE are - assembled H in a m stack to e form a the CE-E.s u r e s atPlate high pileup, the granularity Chambers of both the inner pixel tracker and thefor outer tracker the will be 489 provided in Chapter 3. The strips inFor the the 2S modules CE-H part, have the a length absorber of about is made 5 cm, while of full those disks of stainless steel forming a rigid mechanical increased, by decreasing the pixel size and by shortening the strip lengths. For the first time 490 in the PS modules are about 2.4 cm long. In PS modules one of the two sensors is segmented at CMS a momentum measurement will become possible within a few microseconds, and this 491 into macro-pixels of about 1.5structure mm length, in providing which the the activez coordinate elements measurement. are inserted The PS as cassettes of 60°. The 60° cassettes are split information can be used in the Level-1 (L1) trigger. The track trigger will greatly sharpen the 492 modules are deployed in the first three layers of the Outer Tracker, in the radial region of 200– L1 resolution, which will reduce the trigger rate at a given transverse momentum. Thus Figure 2. Schematic of particle flow algorithm for CMS Level-1 trigger correlator. compatible with tracksin 30° units to keep theirhadronic size manageable. The first 8 layers showers. are fully instrumented with silicon The muonpT system. 493 600 mm, i.e. down to radii at which the stub pT resolution remains acceptable and the data by combining input from the tracker and muon systems the pT threshold for the single muon PUPPI Algorithm for Level-1 Trigger 494 reduction effective. The 2S modulessensors, arewhile deployed the in rest the outermost of the layers three are layers, made in the of radial scintillator tiles in the low pseudorapidity region trigger can be kept low despite the high rate at HL-LHC. 495 region above 600 mm. In the endcaps the modules are arranged in rings on disc-like structures, with pT>2 GeV atand silicon the sensors inactive the high pseudorapidity elements region. As opposed to the CE-E are part, the detectorsilicon The endcap calorimeters will also suffer significant radiation damage. The replacement planned 496 with the rings at low radii, up to about 700 mm, equipped with PS modules, while 2S modules Figure 2.4: Average number of module layers traversed by particles, including both the Inner for Phase-2, the High Granularity Calorimeter (HGCAL), will have an electromagnetic and a 497 are used at larger radii. The modulesz coordinates are provided assembled by the only three on PS one barrel side layers of the constrain copper cooling plate. tracks vertexing Tracker (red) and the Outer Tracker (blue)498 modules,the origin as of well the trigger as the tracks complete to a portiontracker of (black). the luminous Par- region of about 1 mm, which is ticle trajectories are approximated by straight499 sufficiently lines, using precise a flat to partially distribution discriminate of primary particles vertices coming from different vertices. full LHC collision rate. and scintillator based. primary within z0 < 70 mm, and multiple scattering is not included. | | –2– The algorithm inputs are muonvertex candidates from the muon track finders, tracks from the The following section summarizes the main concepts and features of the upgraded tracking system. One quarter of the Phase-2 tracker layout can be seen in Fig. 2.3. Figure 2.4 shows charged pileup charged particles charged particles the average number of active layers that are traversed by particles originating from the lumi- tracking subdetector TPGs, and electromagneticsubtraction from and primary hadronic vertex calorimeter energy deposits nousCMS region, for the complete Level-1 tracker as well as for the Inner Trigger Tracker and the Outer Tracker for HL-LHC separately. particle The number of layers has been optimised to ensure robust tracking, i.e. basically unaffected from the barrel calorimeter trigger and endcap calorimeter TPGs. The algorithm reconstructs performance when one detecting layer is lost in some parts of the rapidity acceptance. The six flow layersAn of the Outer all Tracker are the new minimum required FPGA-based to ensure robust track finding at the L1 Level 1 Trigger will process data from CMS particles centered within a restricted region in pseudorapidityPUPPI (η) andreweighted azimuthal angle (φ) trigger in the rapidity acceptance of h < 2.4, as discussed in more details in Section 3.1. neutral particles | | weight neutral particles subdetector trigger primitive generators (TPGs). based3.6. New on Trigger inputs Objects from based3.6. an onNew Particle enlarged Trigger Flow Objects Reconstruction region based on to Particle correlate Flow Reconstruction across the27 η-φ boundaries.27 Multiple tracker calorimeters muon detectors copies of the algorithm will run in parallel to cover the whole detector. The PUPPI weight fortrigger-primitive neutral clusters particle or trigger-primitive i is computed muons. We analyse using the linked trigger-primitive a lookup table trigger-primitive clusters or trigger-primitive muons. We analyse2 the linked trigger-primitivep 2 2 Muon candidatesinformation and tracks to create closest a globalpT,j list in ofp L1 particlewithin candidates: a ∆R muons,≡ charged∆η + hadrons, ∆φ requirement electrons, are barrel barrel forward resistive cathode gas information to create a global↵i list= of L1 particle candidates: muons,⇥(RT0 chargedRij) hadrons, electrons, endcap drift addressed by the metricphotons, and neutral hadrons.Rij We then run the L1 PUPPI algorithm on the list of global particle tracker electromagnetic hadronic hadronic plate strip electron photons, and neutral hadrons. Wej thencharged run from theL1 PV PUPPI✓ algorithm◆ on the list of global particle calorimeter tubes linked to form particle-flow2 X muon candidates. The linked tracks are removed from further calorimeter calorimeter calorimeter chambers chambers multipliers candidates to filter the eventcandidates into the most to filter leading, the event vertex into compatible the most particle leading, candidates. vertex compatible As particle candidates. As Algorithmconsideration.with the PF Performance algorithm, Electromagnetic thewith PUPPI the algorithmic PF algorithm, calorimeter complexity the PUPPI is reduced algorithmic clusters and only complexity and uses tracks simplified is reduced closest and only in ∆ usesR simplifiedwithin a ∆R re- integer operations. Since PUPPIinteger requires operations. input Since information PUPPI requires from track input vertexing information but the from PF track vertexing but the PF TPG TPG TPG TPG TPG TPG TPG TPG TPG quirementalgorithm does are not, linked. such aalgorithm vertexing Linked stepdoes iselectromagnetic not, naturally such a performed vertexing step in energy parallel is naturally with clusters the performed PF oper- and in parallel tracks with form the PFparticle-flow oper- Performanceelectrons,ations. The is output and studied the of the linked combinedations. using The tracks PF+PUPPI output simulated are of trigger the removed combined algorithm t-tbar PF+PUPPI isfrom a set ofevents further triggervertex filtered algorithm consideration.decaying can- is a set of vertex semileptonically If filtered the electromag-can- didates that can then be useddidates to reconstruct that can then and beidentify used totriggerable, reconstruct prompt and identify physics triggerable, objects. prompt physics objects. as a signalnetic energy with cluster’s large This missingp approachis significantly to transversepileup mitigation larger needs energy than the primary that (MET) ofvertex the to betrack, and properly the scalar reconstructed, excess sump as forms of jet a barrel muon endcap muon This approach to pileup mitigationT needs the primary vertex to be properly reconstructed, as T barrel calorimeter trigger can be done easily in events with a large multiplicity of high tracks (see Section 3.2). track finder track finder energyparticle-flow (HcanT be) and done easily minimum-bias photon. in events with Particle-flow a large multiplicity events photons of highas pbackgroundT tracks are also (see Section formed 3.2).. p fromT unlinked electromagnetic energyCMS clusters.Phase-2 The Simulation electromagneticCMS,=140Phase-2 Simulation and hadronic,=140 calorimeter energy is clustered and cali- O(10) Tbps brated, and the energy of the particle-flow electrons and photons is removed. The resulting ~ 2.5 μs Semileptonic tt Semileptonic tt correlator 10 10 calorimeter clusters are linked with tracks based on ∆pT and ∆R. Linked calorimeter clusters Rate [MHz] Rate [MHz] and tracks and unlinked low-pT tracks form particle-flow charged hadrons. If the calorimeter global trigger cluster’s1 pT is significantly1 larger than that of the track, the excess pT forms a particle-flow

The Correlator combines all of this information into trigger objects, which are neutral hadron. Particle-flowMET (Calo) neutral hadronsMET (Calo) are also formed from unlinked calorimeter clus- sent to the Global Trigger to make the Level-1 trigger decision. ters. MET (TK Δz) MET (TK Δz) −1 −1 10 10 MET (PF+PUPPI) MET (PF+PUPPI)

Particle-Level Reconstruction 0.1 0.2 0.3 0.4 0.5 0.6 0.70.1 0.80.2 0.9 0.31 0.4 0.5 0.6 0.7 0.8 0.9 1 Signal effciency Signal effciency Figure 3.8: (Left) efficiency for selecting signal and background for three different Emiss trigger miss Particle Flow reconstruction correlates detector signals to identify final-state Figure 3.8: (Left) efficiency for selecting signal and backgroundT for three different ET trigger 34 5 Performance in simulation algorithms and (Right) efficiencyalgorithms turn-on and curves (Right) for efficiency three different turn-onHT curvestrigger for working three different points: trigger working points: Secondly, in low pileup environments, there is less information available locally just due to HT particles and measure their properties. This approach provides better FPGA aImplementation calorimeter-only algorithma calorimeter-only (purple); a track-only algorithm trigger (purple); algorithm a track-only using tracks trigger consistent algorithm using tracks consistent the lack of pileup. This means the ↵ distribution is not as well populated and the uncertainty with the primary vertex (red); and a PF+PUPPI trigger algorithm (blue). The H thresholds performance than single-detector reconstruction and is already used in the with the primary vertex (red); and a PF+PUPPI triggerT algorithm (blue). The HT thresholds were chosen so that each H trigger path corresponds to a rate of 20 kHz. The study was on PU is larger. Particle flow and PUPPIwereT chosen are so inherentlythat each HT trigger regional path corresponds algorithms. to a rate of 20 kHz. AsThe study a was proof-of- High Level Trigger and offline reconstruction. conducted using a backgroundconducted sample using of minimum-bias a background collisions sample ofand minimum-bias a signal sample collisions from t¯t and a signal sample from t¯t Anti-kT (R=0.7), n = 80 Anti-kT (R=0.7) JHEP 10 (2014) 59 principle,simulated we divide events decaying the semileptonically, detector corresponding into η tox an ɸ average ~ 0.55 PU of 140. x 0.55 regions accepting up PU CMS-PRF-14-001 simulated events decaying semileptonically, corresponding to an average PU of 140. 0.3 0.2 LV LV CMS Anti-kT, R = 0.4 Calo CMS Anti-kT, R = 0.4 Calo µ |η0.6| < 2.5 0.6 pT = [100-200] GeV, |η| < 2.5 While the L1 PF+PUPPI triggerWhile algorithm the L1 PF+PUPPI remains an trigger early algorithm prototype remainsat the time an ofearly this prototype writ- at the time of this writ- µ Simulation |ηRef| < 1.3 PF Simulation 1.6 < |ηRef| < 2.5 PF to 25 tracks and 20 calorimeter clusters each. )/pT )/pT ing, the initial performance has been studied and compared with more traditional, standalone PFlow PFlow ing, the initial performance has been studied and compared with more traditional, standalone LV LV neutral 0.15 trigger algorithms based on a single detector (e.g. either calorimeter-only or tracker-only). As HCAL PFlowCHS PFlowCHS Implementedtrigger in algorithms Xilinx based Vivado on a single HLS detector v2016.4 (e.g. either calorimeter-only for an Ulstrascale+ or tracker-only). As VU9P hadron 0.2 clusters PUPPI PUPPI an early proxy of the potentialan early gains proxy expected of the by potential this approach, gains expected we study by the this missing approach, trans- we study the missing trans- 0.4 0.4 miss verse momentumFPGA, (E T four) and the regions scalar summed miss canpT over beall jets inprocessed an event, typically referred with ~40% of the FPGA verse momentum (ET ) and the scalar summed pT over all jets in an event, typically referred Detector , (pT - pT , (pT - pT 0.1 to as H . Gains are also expected for jet substructure studies for heavy-particle tagging, and photon Energy resolution Energy resolution T σ σ resources.to as TheHT. Gains latencies are also expected for for jet substructure Particle studies forFlow heavy-particle and tagging, PUPPI and are lepton isolation, though welepton leave thoseisolation, topics though for future we leave studies. those An topics example for future of the early studies. per- An example of the early per- Particle Flow 0.1 0.2 0.2 formance of the L1 PF+PUPPI algorithm, using a Phase-2 detector simulation but using tracks

fitted fitted approximatelyformance 500 of the L1ns PF+PUPPI and 100 algorithm, ns, using respectively. a Phase-2 detector simulation but using tracks ECAL 0.05 reconstructed with an offline algorithm and mocked-up to have similar performance (resolu- clusters reconstructed with an offline algorithm and mocked-up to have similar performance (resolu- tion, pT threshold of 2 GeV, etc) as that expected from the L1 track finder, is provided in Fig. 3.8. tion, pT threshold of 2 GeV, etc) as that expected frommiss the L1 track finder, is provided in Fig. 3.8. The figures show the signal vs background selection efficiency for different ET triggers, as miss 0 0 0 The figures show the signal vs background selection efficiency for different ET triggers, as 0 10020 200100 200 3001000 400 500 20 10050200 1000 100 well as the efficiency turn-on curves for different triggers, using a semi-leptonic t¯t signal Outlook well as the efficiency turn-onHT curves for different H triggers, using a semi-leptonic t¯t signal RefpT (GeV) Ref n miss T p (GeV) p (GeV) PU having true E , and they compare threemiss different algorithms based on standalone calori- T T T having true E , and they compare three different algorithms based on standalone calori- meter, standalone track, as well as PF+PUPPI.T The H efficiency turn-on curves show three Figure 13: Jet energy resolution as a function of pRef in the barrel (left) and in the endcap These proof-of-principlemeter, studies standalone track, indicate asT well as PF+PUPPI. the feasibility The H efficiency of turn-on performing curves show three Particle Figure 7:Jetp resolution vs. p (left) for n T= 80 for ⌘ < 2.5 and jet p resolution vs. different working-point thresholds, one for each algorithm, that are constrained to haveT a fixed Particle flow enables the use(right) of regions. particle-levelT The lines, added toT guidepileup the eye,PU correspondmitigation, to| fitted| functions including withT ad hoc the Flow reconstruction anddifferent PUPPI working-point pileup thresholds, mitigation one for each algorithm, in the that are CMS constrained HL-LHC to have a fixed Level-1 numberparametrizations. of pileup interactions (right) for jets with pT between 100 and 200 GeV for ⌘ < 2.5. PUPPI algorithm, which attempts to remove the effects of both charged | and| Trigger. Significant performance improvements over traditional trigger neutral particles originating4.2 from Jet Shapes pileup (JHEP 10 (2014) 59). algorithms have been shown and may be extended with further R&D.

Similar to our study of pT distributions, we can study resolution and its pileup dependence for jet shapes. Here we show results for jet mass which is considered a reasonable proxy for generic0.2 jet shapes and is used in many applications0.2 such as boosted object tagging (see CMS |ηRef| < 1.3 CMS |ηRef| < 1.3 [34–36] and referencesSimulation therein).Parton flavor Simulation Parton flavor Anti-k R = 0.4, Calo Anti-k R = 0.4, PF 0.15 T 0.15 T Anti-kT (R=0.7), n = 80 ud Anti-kT (R=0.7), udn = 80 PU PU s s 0.3 0.3 pT = [100-200] GeV, |η| < 2.5c LV pT = [100-200] GeV,c |η| < 2.5 LV PFlow PFlow Connecting the Dots 2018, Seattle, WA 0.1 b PFlowCHS 0.1 b PFlowCHS PUPPI PUPPI

0.2 0.2 fraction of jets 0.05 0.05fraction of jets Difference to gluon response Difference to gluon response 0.1 0 0 0.1 20 100 200 1000 20 100 200 1000 pRef (GeV) pRef (GeV) T T 0 0 Figure0 14: Absolute difference 50 in jet energy 100 response between0 quark and gluon 50 jets as a function 100 Ref of pT for Calo jets (left) and PF jets mass (right). (GeV) trimmed mass (GeV)

Figure 8: The single jet mass resolution for nPU = 80 for jets with 100 GeV

First we look at jet mass for central jets with 100 GeV

– 15 – Particle Flow and PUPPI in the Level-1 Trigger at CMS for the HL-LHC Benjamin Kreis, Fermi National Accelerator Laboratory, Batavia, IL for the CMS Collaboration [email protected] Abstract With the planned addition of the tracking information in the Level 1 trigger in CMS for the HL-LHC, the algorithms for Level 1 trigger can be completely reconceptualized. Following the example for offline reconstruction in CMS to use complementary subsystem information and mitigate pileup, we explore the feasibility of using Particle Flow-like and pileup per particle identification (PUPPI) techniques at the uses lead, copper, and copper-tungsten alloy as absorbers while copper and steel are used for the hadronic part. Silicon sensors are used in the region with the highest level of radiation, whilst hardwareplastic trigger scintillator with on-tile level. silicon photomultipliers This are usedrepresents for the part of the CE-H where thea new type of multi-subdetector pattern recognition challenge for the HL-LHC. We present proof-of- radiation level is the lowest. The key parameters of the HGCAL are summarised in figure 1. Surfaces of about 600 m2 of silicon sensors and 500 m2 of scintillator tiles are used. The silicon principle sensorsstudies are finely segmented on into 0.both5 and 1 cm2 cells physics with active thicknesses and of 300 µm, 200 resourceµm usage performance of a prototype algorithm for use by CMS in the HL-LHC era. or 120 µm depending on the region of the detector. These cell parameters are chosen such that the signal-over-noise ratio remains high enough for the measurement of MIP signals over the full lifetime of the HGCAL. Scintillator cells of 1° and 1.25° are used, arranged in a ⌘- grid. 2018 JINST 13 C02043

CMS Detector Upgrades for HL-LHC Chapter 1. The CMS Phase-2 muon detector Particle Flow Algorithm for Level-1 Trigger 18 2.3. Overview of the upgraded tracker concept 21 η 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.1 θ° 84.3° 78.6° 73.1° 67.7° 62.5° 57.5° 52.8° 48.4° 44.3° 40.4° 36.8° η θ° 8 ) 1.2 33.5° DTs Hexagonal modules based on Si sensors m CSCs standalone muons

R ( MB4 RPCs charged hadrons radiation regions of CE-H RB4 1.3 30.5° 7 GEMs linked “Cassettes”: multiple modules mounted on Wheel 0 Wheel 1 Wheel 2 22 Chapter 2. Overview of the Phase-2 Tracker Upgrade iRPCs ME0 muons cooling plates with electronics and absorbers MB3 1.4 27.7° CMS-TDR-17-001 6 RB3 T(EM)

readout in ME1/3 muon/track RE1/3 RE2/3 RE3/3 RE4/3 1.5 25.2° (track) 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 T electrons neutral hadrons 1.6 MB2 track 1200 5 RB2 ME2/2 ME3/2 ME4/2 1.6 22.8° linking T 1.7 20.7° MB1 (closest in pT) (track), or p 1.8 RB1 T

r [mm] 1000 ~2.3m 4 1.8 18.8° RE1/2 RE2/2 RE2/2 RE3/2 RE4/2 unlinked tracks 2.0 1.9 17.0° 800 ME1/2 Solenoid magnet 2.0 15.4° tracks photons 2.2 3 2.1 14.0° 2.2 12.6° (EM) ~ p 600 T (track) 2.4 2.3 11.5° p ME2/1 ME3/1 ME4/1 3.2 PUPPI Pileup Mitigation in Level-1 Trigger T 2.6 HCAL RE3/1 RE4/1 2.4 10.4° T(EM) Figure 2.4: Illustration of the module concept. (a) Correlation of signals in closely-spaced GE2/1 significantly larger than p 400 pT 2.8 2 2.5 9.4° 3.0 GE1/1 due to electron+photon sensors enables rejection of low-pT particles; the channels shown in green represent the selec- ME1/1 ECAL 2.8 7.0° 200 tion window to define an accepted stub. (b) The same transverse momentum corresponds to a Steel 3.0 5.7° (calo) significantly 4.0 1 HGCAL T larger distance between the two signals at large radii
for a given sensor spacing. (c) For the end- Silicon ME0 tracker 0 4.0 2.1° cap discs, a larger spacing between the sensors is needed to achieve the same discriminating CMS-TDR-016 (track) due to linked, or unlinked low p unlinked calorimeter, 0 500 1000 1500 2000 2500 z [mm] 5.0 0.77° The best performing pileup mitigationelectromagnetic/ technique usedno track, T orin p CMS offline reconstructionor p is PileUp power as in the barrel at the same radius. 13 C02043 0 larger than p Figure 2.3: Sketch of one quarter of the tracker layout in r-z view. In the Inner Tracker the ~2m 0 1 2 3 4 5 6 7 8 9 10 11 12 z (m) significantly larger J.-B. Sauvan 2018 JINST green lines correspond to pixel modules made of two readout chips and the yellow lines to track linking than p 473 old depends on the acceptance window, which can be tuned to a certain level by programming Figure 1.4: An R-z cross section of a quadrant of the CMS detector, including the Phase-2 up- electromagnetic electron+photon pixel modules with four readout chips.474 Inthe the respective Outer Tracker setting in the the blue readout and chip. red Forlines the representpT modules a few different values of sensor grades (RE3/1, RE4/1, GE1/1, GE2/1, ME0). The acronym iRPCs in the legend refers to the Per Particle Identification (PUPPI), an(closest algorithm in ΔR) that removes charged particles with tracks the two types of modules described in the475 text.spacing are used, optimized to achieve the desired pT filtering in different regions of the de- new improved RPC chambers RE3/1 and RE4/1. The interaction point is at the lower left cor- unlinked tracks calorimeter/track ner. The locations of the various muon stations are shown in color (MB = DT = Drift Tubes, ME calorimeter All new silicon476 tector (Fig. 2.4 (b) and (c)). tracker For a pitch of about 100 µ m betweenH silicon i gstrips (or h macro-pixels, g r a n u l a r i t y New Cathode Strip 477 as detailed below) in the transverse plane, sufficient pT resolution can be achieved down to a = CSC = Cathode Strip Chambers, RB and RE = RPC = Resistive Plate Chambers, GE and ME0 linking = GEM = Gas Electron Multiplier). M denotes Muon, B stands for Barrel and E for Endcap. not originating at the primary vertex and downweights neutral particles based on the proba- 478 radius of about 200 mm in a barrelFigure geometry, 1. Schematic thanks to longitudinal the 3.8 T magnetic cross field section of CMS. of the The HGCAL calorimeter and its key parameters. The (based on pT and ΔR) Labelling details are given in Section 1.2.2. The magnet yoke is represented by the dark gray 479 concept is therefore applicablesilicon in the Outer andscintillator Tracker, and parts limited of in the angular detector acceptance are highlighted to about in green and dark blue respectively. with pixel-strip and calorimeter in endcap areas.Chambers, G a s clustering, electromagnetic+hadronic 480 h < 2.4. | | bility that they originate fromcalibration pileup [6, 7]. The weight is calculated from the neighboring 481 2.3.2 The Outer Tracker For the CE-E part, silicon sensors are sandwiched between copper-tungsten baseplates and Near the interaction region a silicon tracker, composed of an inner pixel detector surrounded subtract two-strip modules in region. CE-E measures byElectron a silicon strip detector, measures vertices andMultipliers momenta of charged particles. The elec- , 482 The Outer Tracker is populatedPCBs with assembledpT modules,implementing into hexagonal the L1 modules, trigger functionality. and mounted on both sides of a copper cooling plate. tromagnetic calorimeter (ECAL) and the hadronic calorimeter (HCAL) are located inside the electrons/photons 483 The pT module concept relies on the fact that the strips of the top and bottom sensors of a charged particles not originating from pileup. Here we describe a proof-of-principle imple- Stainless steel tubes inside the cooling plate carry two-phase CO2 to maintain the silicon sensors at a 484 module are parallel to each other. With the strip direction being parallel to the z axis in the solenoid, measuring electromagnetic and hadronic showers with lead tungstate crystals and a hadronic calorimeter 485 barrel and nearly radial in thetemperature endcaps, this prevents of 30° theC. concept Each cooling of stereo plate strips is to covering be used to an azimuthal angle of 60° over the full radius of scintillator-brass sampling detector, respectively. o u t e r r e g i o n t h a t electromagnetic showers, and upgraded Resistive 486 measure the z coordinate (r coordinate)each layer. in the Plates barrel of (endcaps). lead and stainless For this reason steel two of the versions same dimensions cover the sensors and electronics. The current silicon tracker must be replaced before the start of Phase-2, since it will suffer 487 of modules have been realized: modules with two strip sensors (2-strip or modules) mentation of PUPPI for the FPGAs of the Level-1 trigger correlator. A schematic is shown pT 2S significant radiation damage by the end of Run 3. To maintain excellent track reconstruction provide 488 pairsand modules with a stripof and This a macro-pixel assemblyhits sensor is called (pixel-stripa a “cassette”, n or PS dmodules). and 14 CDetailscassettes areE are - assembled H in a m stack to e form a the CE-E.s u r e s atPlate high pileup, the granularity Chambers of both the inner pixel tracker and thefor outer tracker the will be 489 provided in Chapter 3. The strips inFor the the 2S modules CE-H part, have the a length absorber of about is made 5 cm, while of full those disks of stainless steel forming a rigid mechanical increased, by decreasing the pixel size and by shortening the strip lengths. For the first time 490 in the PS modules are about 2.4 cm long. In PS modules one of the two sensors is segmented at CMS a momentum measurement will become possible within a few microseconds, and this in Figure 3. 491 into macro-pixels of about 1.5structure mm length, in providing which the the activez coordinate elements measurement. are inserted The PS as cassettes of 60°. The 60° cassettes are split information can be used in the Level-1 (L1) trigger. The track trigger will greatly sharpen the compatible492 modules with are deployed in thetracks firstin 30 three° units layers toof the keep Outer theirhadronic Tracker, size in manageable. the radial region The of 200– first 8 layers showers. are fully instrumented with silicon The L1muonpT resolution, which willsystem. reduce the trigger rate at a given transverse momentum. Thus 493 600 mm, i.e. down to radii at which the stub pT resolution remains acceptable and the data by combining input from the tracker and muon systems the pT threshold for the single muon PUPPI Algorithm for Level-1 Trigger 494 reduction effective. The 2S modulessensors, arewhile deployed the in rest the outermost of the layers three are layers, made in the of radial scintillator tiles in the low pseudorapidity region trigger can be kept low despite the high rate at HL-LHC. 495 region above 600 mm. In the endcaps the modules are arranged in rings on disc-like structures, with pT>2 GeV atand silicon the sensors inactive the high pseudorapidity elements region. As opposed to the CE-E are part, the detectorsilicon The endcap calorimeters will also suffer significant radiation damage. The replacement planned 496 with the rings at low radii, up to about 700 mm, equipped with PS modules, while 2S modules Figure 2.4: Average number of module layers traversed by particles, including both the Inner for Phase-2, the High Granularity Calorimeter (HGCAL), will have an electromagnetic and a 497 are used at larger radii. The modulesz coordinates are provided assembled by the only three on PS one barrel side layers of the constrain copper cooling plate. tracks vertexing Tracker (red) and the Outer Tracker (blue)498 modules,the origin as of well the trigger as the tracks complete to a portiontracker of (black). the luminous Par- region of about 1 mm, which is ticle trajectories are approximated by straight499 sufficiently lines, using precise a flat to partially distribution discriminate of primary particles vertices coming from different vertices. full LHC collision rate. and scintillator based. primary within z0 < 70 mm, and multiple scattering is not included. | | –2– vertex

The following section summarizes the main concepts and features of the upgraded tracking system. One quarter of the Phase-2 tracker layout can be seen in Fig. 2.3. Figure 2.4 shows charged pileup charged particles charged particles the average number of active layers that are traversed by particles originating from the lumi- subtraction from primary vertex nousCMS region, for the complete Level-1 tracker as well as for the InnerTrigger Tracker and the Outer Tracker for HL-LHC separately. particle The number of layers has been optimised to ensure robust tracking, i.e. basically unaffected performance when one detecting layer is lost in some parts of the rapidity acceptance. The six flow layersAn of the Outer all Tracker are thenew minimum required FPGA-based to ensure robust track finding at the L1 Level 1 Trigger will process data from CMS PUPPI reweighted trigger in the rapidity acceptance of h < 2.4, as discussed in more details in Section 3.1. neutral particles | | weight neutral particles subdetector trigger primitive generators (TPGs). 3.6. New Trigger Objects based3.6. onNew Particle Trigger Flow Objects Reconstruction based on Particle Flow Reconstruction 27 27 tracker calorimeters muon detectors The PUPPI weight for neutral particle i is computed using a lookup table trigger-primitive clusters ortrigger-primitive trigger-primitive clusters muons. or We trigger-primitive analyse2 the linked muons. trigger-primitive We analyse the linked trigger-primitive Figure 3. Schematic of PUPPIinformation pileup to create mitigation a globalpT,j list algorithm of L1 particle candidates: for CMS muons, Level-1 charged trigger hadrons, correlator. electrons, barrel barrel forward resistive cathode gas information to create a global↵i list= of L1 particle candidates: muons,⇥(R0 chargedRij) hadrons, electrons, endcap drift addressed by the metricphotons, and neutral hadrons.Rij We then run the L1 PUPPI algorithm on the list of global particle tracker electromagnetic hadronic hadronic plate strip electron photons, and neutral hadrons. Wej thencharged run from theL1 PV PUPPI✓ algorithm◆ on the list of global particle calorimeter tubes 2 X calorimeter calorimeter calorimeter chambers chambers multipliers candidates to filter the eventcandidates into the most to filter leading, the event vertex into compatible the most particle leading, candidates. vertex compatible As particle candidates. As Algorithmwith the PF Performance algorithm, thewith PUPPI the algorithmic PF algorithm, complexity the PUPPI is reduced algorithmic and only complexity uses simplified is reduced and only uses simplified Theinteger position operations. along Sincethe PUPPIinteger beam requires operations. axis input Since information(z) PUPPI of the requires from primary track input vertexing information vertex but the is from PF computed track vertexing from but the all PF tracks in TPG TPG TPG TPG TPG TPG TPG TPG TPG algorithm does not, such aalgorithm vertexing stepdoes is not, naturally such a performed vertexing step in parallel is naturally with the performed PF oper- in parallel with the PF oper- Performancethe event.ations. One The is output studied algorithm of the combinedations. using for The PF+PUPPI finding output simulated of trigger the combined algorithm primary t-tbar PF+PUPPI is a setvertex ofevents triggervertex is filtered algorithm to decaying find can- is a the set ofz vertex positionsemileptonically filtered can- of the peak didates that can then be useddidates to reconstruct that can then and beidentify used totriggerable, reconstruct prompt and identify physics triggerable, objects. prompt physics objects. as in a a signal histogram with of largetrack z This missingweighted approach to bytransverse pileup track mitigationpT . needs Charged energy the primary particles (MET) vertex to be and with properly tracksscalar reconstructed, not sum originating as of jet barrel muon endcap muon This approach to pileup mitigation needs the primary vertex to be properly reconstructed, as barrel calorimeter trigger can be done easily in events with a large multiplicity of high tracks (see Section 3.2). track finder track finder energynear the(HcanTprimary be) and done easily minimum-bias vertex in events with are a largeremoved. multiplicity eventsFor of highas each pbackgroundT tracks neutral (see Section particle 3.2).. pT i, we compute the sum of CMS Phase-2 Simulation,=140 pT /∆Ri j ofCMS the remainingPhase-2 Simulation charged,=140 particles j within a∆Rrequirement. The sum addresses a O(10) Tbps precomputed lookup table of weights. ~ 2.5 μs Semileptonic tt Semileptonic tt correlator 10 10 Rate [MHz] Rate [MHz] global trigger 4 Results 1 1 The Correlator combines all of this information into trigger objects, which are 4.1 Physics PerformanceMET (Calo) MET (Calo) sent to the Global Trigger to make the Level-1 trigger decision. MET (TK Δz) MET (TK Δz) −1 −1 We test the10 physics performance10 MET (PF+PUPPI) of particle-flowMET (PF+PUPPI) reconstruction and PUPPI pileup mitigation Particle-Level Reconstruction in the Level-10.1 trigger0.2 0.3 0.4 using 0.5 0.6 simulated 0.70.1 0.80.2 0.9 0.31 0.4 samples 0.5 0.6 0.7 of 0.8tt¯ 0.9events1 decaying semileptonically as signal Signal effciency Signal effciency and minimum-biasFigure 3.8: (Left) efficiency events for selecting as background. signal and background The for average three different numberEmiss trigger of pileup interactionsmiss in the Particle Flow reconstruction correlates detector signals to identify final-state Figure 3.8: (Left) efficiency for selecting signal and backgroundT for three different ET trigger 34 5 Performance in simulation algorithms and (Right) efficiencyalgorithms turn-on and curves (Right) for efficiency three different turn-onHT curvestrigger for working three different points: trigger working points: Secondly, in low pileup environments, there is less information available locally just due to simulation is 140, corresponding to a baseline HL-LHC scenario. HT particles and measure their properties. This approach provides better FPGA aImplementation calorimeter-only algorithma calorimeter-only (purple); a track-only algorithm trigger (purple); algorithm a track-only using tracks trigger consistent algorithm using tracks consistent the lack of pileup. This means the ↵ distribution is not as well populated and the uncertainty with the primary vertex (red); and a PF+PUPPI trigger algorithm (blue). The H thresholds performance than single-detector reconstruction and is already used in the Figure 4 (left) showswith the the rate primary of vertex a trigger (red); and on a PF+PUPPI missing trigger transverseT algorithm (blue). energy The H (MET)T thresholds computed were chosen so that each H trigger path corresponds to a rate of 20 kHz. The study was on PU is larger. Particle flow and PUPPIwereT chosen are so inherentlythat each HT trigger regional path corresponds algorithms. to a rate of 20 kHz. AsThe study a was proof-of- High Level Trigger and offline reconstruction. with particle-flowconducted using a backgroundcandidatesconducted sample versus using of minimum-bias a background signal collisions sampleefficiency ofand minimum-bias a signal for sample events collisions from twith¯t and a signalMET sample>100 from GeV t¯t at the Anti-kT (R=0.7), n = 80 Anti-kT (R=0.7) JHEP 10 (2014) 59 principle,simulated we divide events decaying the semileptonically, detector corresponding into η tox an ɸ average ~ 0.55 PU of 140. x 0.55 regions accepting up PU CMS-PRF-14-001 simulated events decaying semileptonically, corresponding to an average PU of 140. 0.3 0.2 generator level. We compare with two alternative MET triggers. The first computes MET LV LV CMS Anti-kT, R = 0.4 Calo CMS Anti-kT, R = 0.4 Calo µ |η0.6| < 2.5 0.6 pT = [100-200] GeV, |η| < 2.5 While the L1 PF+PUPPI triggerWhile algorithm the L1 PF+PUPPI remains an trigger early algorithm prototype remainsat the time an ofearly this prototype writ- at the time of this writ- µ Simulation |ηRef| < 1.3 PF Simulation 1.6 < |ηRef| < 2.5 PF to 25 tracks and 20 calorimeter clusters each. )/pT )/pT from calorimetering, the initial performance energy has deposits. been studied and The compared second withcomputes more traditional, MET standalone from tracks with a small ∆z PFlow PFlow ing, the initial performance has been studied and compared with more traditional, standalone LV LV neutral 0.15 trigger algorithms based on a single detector (e.g. either calorimeter-only or tracker-only). As hadron HCAL PFlowCHS PFlowCHS with respectImplemented to the primarytrigger vertex.in algorithms Xilinx Figure based Vivado on a 4 single (right) HLS detector compares v2016.4 (e.g. either calorimeter-only trigger for an e ffiUlstrascale+ orciency tracker-only). at a As fixed VU9P rate 0.2 clusters PUPPI PUPPI an early proxy of the potentialan early gains proxy expected of the by potential this approach, gains expected we study by the this missing approach, trans- we study the missing trans- 0.4 0.4 miss verse momentumFPGA, (E T four) and the regions scalar summed miss canpT pover beallH jets inprocessed an event, typically referred with ~40% of the FPGA of 20 kHz versus the scalarverse momentum sum of (E jetT ) andT ( the scalarT ) at summed the generatorpT over all jets level.in an event, The typically jets referred are clustered Detector , (pT - pT , (pT - pT 0.1 to as H . Gains are also expected for jet substructure studies for heavy-particle tagging, and photon Energy resolution Energy resolution T σ σ resources.k to as TheHT. Gains latencies are also expected for for jet substructure Particle studies forFlow heavy-particle and tagging, p PUPPI and are using thelepton anti- isolation,T thoughalgorithm welepton leave thoseisolation,with topics a though radius for future we leave ofstudies.0.4 those An [8]topics example and for future ofare the early studies. required per- An example to have of the earlyT per-> 30 GeV Particle Flow 0.1 0.2 0.2 formance of the L1 PF+PUPPI algorithm, using a Phase-2 detector simulation but using tracks

fitted fitted approximatelyformance 500 of the L1ns PF+PUPPI and 100 algorithm, ns, using respectively. a Phase-2 detector simulation but using tracks ECAL 0.05 and η

0.2 0.2 fraction of jets 0.05 0.05fraction of jets Difference to gluon response Difference to gluon response 0.1 0 0 0.1 20 100 200 1000 20 100 200 1000 pRef (GeV) pRef (GeV) T T 0 0 Figure0 14: Absolute difference 50 in jet energy 100 response between0 quark and gluon 50 jets as a function 100 Ref of pT for Calo jets (left) and PF jets mass (right). (GeV) trimmed mass (GeV)

Figure 8: The single jet mass resolution for nPU = 80 for jets with 100 GeV

First we look at jet mass for central jets with 100 GeV

– 15 – ParticleParticle Flow Flow and andPUPPI PUPPI in the in Level-1the Level-1 Trigger Trigger at CMS at CMS for thefor HL-LHCthe HL-LHC BenjaminBenjamin Kreis, Kreis, Fermi Fermi National National Accelerator Accelerator Laboratory, Laboratory, Batavia, Batavia, IL IL for thefor CMS the CMSCollaboration Collaboration [email protected]@fnal.gov AbstractAbstract With the Withplanned the additionplanned ofaddition the tracking of the informationtracking information in the Level in the 1 trigger Level 1in trigger CMS for in theCMS HL-LHC, for the theHL-LHC, algorithms the algorithms for Level for1 trigger Level 1can trigger can be completelybe completely reconceptualized. reconceptualized. Following Following the example the forexample offline for reconstruction offline reconstruction in CMS to in use CMS complementary to use complementary subsystem subsystem information information and and mitigate mitigate pileup, we pileup, explore we the explore feasibility the feasibility of using of Particle using Flow-like Particle Flow-like and pileup and per pileup particle per identification particle identification (PUPPI) techniques (PUPPI) techniques at the at the uses lead, copper, and copper-tungstenuses lead, alloy copper, as absorbersand copper-tungsten while copper alloy and as steel absorbers are used while for copper the and steel are used for the hadronic part. Silicon sensorshadronic are used part. in the Silicon region sensors with the are highest used in level the region of radiation, with the whilst highest level of radiation, whilst hardwareplastichardware trigger scintillator with on-tile level.plastic silicontrigger scintillator photomultipliers This with on-tile arelevel. used siliconrepresents for the photomultipliers part ofThis the CE-H are whereused represents for the thea partnew of the CE-H wheretype thea new of multi-subdetector type of multi-subdetector pattern recognition pattern recognition challenge challenge for the HL-LHC. for the WeHL-LHC. present We proof-of- present proof-of- radiation level is the lowest. Theradiation key parameters level is the of lowest. the HGCAL The key are parameters summarised of in the figure HGCAL1. are summarised in figure 1. Surfaces of about 600 m2 of siliconSurfaces sensors of about and 500600mm22ofof scintillator silicon sensors tiles and are used.500 m The2 of scintillatorsilicon tiles are used. The silicon principle sensorsprinciplestudies are finely segmented on sensors intostudies 0.both5 areand finely1 cm2 segmentedcells physics withon into active 0.both5 thicknessesand 1 cm 2 and ofcells 300physics withµm, active200 resourceµm thicknesses and of 300 µm, 200 usageresourceµm performance usage performance of a prototype of a prototype algorithm algorithm for use by for CMS use in by the CMS HL-LHC in the era. HL-LHC era. or 120 µm depending on theor region120 µ ofm the depending detector. on These the region cell parameters of the detector. are chosen These such cell that parameters are chosen such that the signal-over-noise ratio remainsthe signal-over-noise high enough for ratio the measurement remains high of enough MIP signals for the overmeasurement the full of MIP signals over the full lifetime of the HGCAL. Scintillatorlifetime cells of the of 1 HGCAL.° and 1.25 Scintillator° are used, cells arranged of 1° inand a ⌘1-.25grid.° are used, arranged in a ⌘- grid. 2018 JINST 13 C02043 2018 JINST 13 C02043

CMS Detector Upgrades for HL-LHC Chapter 1. The CMS Phase-2 muon detector Particle Flow Algorithm for Level-1 Trigger CMS Detector Upgrades for HL-LHC18 Chapter 1. The CMS Phase-2 muon detector Particle Flow Algorithm for Level-1 Trigger 18 2.3. Overview of the upgraded tracker2.3. concept Overview of the upgraded tracker concept 21 21 η 0.1 0.2 0.3 0.4 0.5 0.6 η 0.70.1 0.80.2 0.30.9 0.4 1.0 0.5 1.10.6 0.7 0.8 0.9 1.0 1.1 θ° 84.3° 78.6° 73.1° 67.7° 62.5° 57.5° θ° 52.8°84.3° 48.4°78.6° 73.1°44.3° 67.7°40.4°62.5° 36.8°57.5° 52.8° η 48.4°θ° 44.3° 40.4° 36.8° η θ° 8 ) 8 ) 1.2 33.5° 1.2 33.5° DTs DTs m Hexagonal modules based on Si sensorsHexagonal modules based on Si sensors m CSCs CSCs standalone muonsstandalone muons R ( MB4 R ( MB4 RPCs charged hadrons RPCs charged hadrons radiation regions of CE-H radiation regions of CE-H 1.3 30.5° 7 RB4 RB4 1.3 30.5° 7 GEMs GEMs linked linked “Cassettes”: multiple modules mounted“Cassettes”: on multiple modules mounted on Wheel 0 Wheel 1 Wheel 2 Wheel 0 Wheel 1 Wheel 2iRPCs 22 22 Chapter 2. Overview of the Phase-2Chapter Tracker 2. Overview Upgrade of the Phase-2 Tracker Upgrade iRPCs ME0 muons 1.4 27.7° ME0 muons cooling plates with electronics and coolingabsorbers plates with electronics and absorbers MB3 MB3 1.4 27.7° (EM) CMS-TDR-17-001 CMS-TDR-17-001 6 RB3 6 RB3 T T(EM)

ME1/3 muon/track readout in ME1/3 readout in RE1/3 RE2/3 RE3/3 RE4/3 muon/track 1.5 25.2°RE1/3 RE2/3 RE3/3 RE4/3 1.5 25.2° (track) 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 T electronsT(track) 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 electrons track neutral hadronsneutral hadrons 1.6 MB2 track 1.6 MB2ME2/2 ME3/2 ME4/2 1200 1200 5 RB2 5 RB2 1.6 22.8° ME2/2 ME3/2 ME4/2 1.6 22.8° linking linking T T 1.7 20.7° 1.7 20.7° (closest in pT) 1.8 MB1 MB1 (closest in pT) (track), or p (track), or p 1.8 RB1 RB1 T T r [mm] 1000

~2.3m 1.8 18.8° r [mm] 1000 4 ~2.3m 1.8 18.8°

4 RE1/2 RE2/2 RE2/2 RE3/2 RE4/2 RE1/2 RE2/2 RE2/2 RE3/2 RE4/2 unlinked tracks unlinked tracks 2.0 1.9 17.0° 1.9 17.0°

2.0 ME1/2 800 800 Solenoid magnet Solenoid magnet 2.0 15.4°ME1/2 tracks 2.0 15.4° tracks photons photons 2.2 2.2 3 3 2.1 14.0° 2.1 14.0° 2.2 12.6° (EM) ~ p (EM) ~ p 600 600 2.2 12.6° T T (track) 2.4 2.4 2.3 11.5° p p (track) ME2/1 ME3/1 ME4/1 2.3 11.5° T ME2/1 ME3/1 ME4/1 T

HCAL RE3/1 RE4/1 RE3/1 RE4/1 (EM) 2.6 2.6 HCAL 2.4 10.4° 2.4 10.4° T (EM) GE2/1 significantly larger than p T Figure 2.4: Illustration of the pT moduleFigure concept. 2.4: Illustration (a) Correlation of the ofmodule signals in concept. closely-spaced (a) Correlation of signals in closely-spaced GE2/1 significantly larger than p 400 400 2.8 pT 2.8 2 2 2.5 9.4° 2.5 9.4° GE1/1 3.0 GE1/1 due to electron+photon due to electron+photon sensors enables rejection of low-pT particles; the channels shown in green represent the3.0 selec- ME1/1 sensors enables rejection of low-pT particles; the channels shown in green represent the selec- ME1/1 ECAL ECAL 2.8 7.0° 2.8 7.0° 200 200 tion window to define an accepted stub. (b) The same transverse momentum corresponds to a Steel 3.0Steel5.7° (calo) significantly tion window to define an accepted stub. (b) The same transverse momentum corresponds to a HGCAL 3.0 5.7° (calo) significantly 4.0 1 1 HGCAL T T 4.0 ME0 larger distance between the two signalslarger at large distance radii between
for a given the sensor two signals spacing. at large (c) For radii the
for end- a given sensor spacing. (c) For the end- Silicon Silicon ME0 0 tracker tracker 0 4.0 2.1° 4.0 2.1° unlinked calorimeter, cap discs, a larger spacing betweencap the discs,sensors a largeris needed spacing to achieve between the the same sensors discriminating is needed to achieve the same discriminating CMS-TDR-016 (track) due to linked, or unlinked low p linked, or unlinked low p unlinked calorimeter, 0 500 1000 1500 2000 2500 z [mm] 5.0 0.77°CMS-TDR-016 5.0 0.77° T (track) due to or p or p 0 500 1000 1500 2000 2500 z [mm] C02043 0 electromagnetic/electromagnetic/no track, or p no track, T or p power as in the barrel at the same radius.power as in the barrel at the same radius. 13 13 C02043 0 larger than p larger than p Figure 2.3: Sketch of one quarter of the tracker layout in r-z view. In the Inner Tracker the ~2m ~2m 0 1 2 3 4 5 0 6 1 7 2 8 3 9 4 10 5 11 6 12 z (m)7 8 9 10 11 12 z (m) significantly larger Figure 2.3: Sketch of one quarter of the tracker layout in r-z view. In the Inner TrackerJ.-B. the Sauvan 2018 JINST significantly larger green lines correspond to pixel modules made of two readout chips and the yellow lines to J.-B. Sauvan 2018 JINST track linking than p than p green lines473 correspond to pixel modules made of two readout chips and the yellow lines to Figure 1.4: An R-z cross section of a quadrant of the CMS detector, including the Phase-2 up- track linking old depends on the acceptance window,473 old depends which can on be the tuned acceptance to a certain window, level which by programming can be tuned to a certain level by programming Figure 1.4: An R-z cross section of a quadrant of the CMS detector, including the Phase-2 up- electromagneticelectromagnetic grades (RE3/1, RE4/1, GE1/1, GE2/1, ME0). The acronym iRPCs in the legend refers to the electron+photon electron+photon pixel modules with four readoutpixel chips. modules474 Inthe the withrespective Outer four Tracker setting readout in the the chips. blue readout474 and Inthe chip.the red respective Outer Forlines the representTrackerp settingT modules in the the ablue readoutfew anddifferent chip. red values Forlines the representofpT sensormodules a few different values of sensor grades (RE3/1, RE4/1, GE1/1, GE2/1, ME0). The acronym iRPCs in the legend refers to the (closest in ΔR) (closest in ΔR) new improved RPC chambers RE3/1 and RE4/1. The interaction point is at the lower left cor- calorimeter/track 475 spacing are used, optimized to475 achieve the desired pT filtering in different regions of the de- new improved RPC chambers RE3/1 and RE4/1. The interaction point is at the lower left cor- unlinked tracks unlinked tracks calorimeter/track the two types of modules describedthe two in the types text. of modules described in the text.spacing are used, optimized to achieve the desired pT filtering in different regions of the de- calorimeter calorimeter 476 ner. The locations of the various muonner. stations The locations are shown of thein color various (MB muon = DT stations = Drift Tubes, are shown ME in color (MB = DT = Drift Tubes, ME All new All silicontector (Fig. new 2.4 (b) and (c)). tracker For asilicon476 pitchtector of about (Fig. 1002.4 (b)µ m and betweenH (c)). tracker For silicon i a pitch gstrips of about(or h macro-pixels, 100 µ m betweenH g silicon i gstripsr a (or h macro-pixels, n u l g a r r a i n t y u lNew a r i t yCathode New Cathode Strip Strip = CSC = Cathode Strip Chambers, RB and RE = RPC = Resistive Plate Chambers, GE and ME0 477 as detailed below) in the transverse477 as plane, detailed sufficient below)pT inresolution the transverse can be plane, achieved sufficient downp toT resolution a can be achieved down to a = CSC = Cathode Strip Chambers, RB and RE = RPC = Resistive Plate Chambers, GE and ME0 linking linking . Schematic longitudinal cross section of the HGCAL calorimeter and its key parameters. The = GEM = Gas Electron Multiplier). M denotes Muon, B stands for Barrel and E for Endcap. 478 radius of about 200 mm in a barrelFigure478 radius geometry, 1 of about thanks 200 to mm the 3.8in a T barrel magneticFigure geometry, 1 field. Schematic of thanks CMS. toThe longitudinal the 3.8 T magnetic cross field section of CMS. of the The HGCAL calorimeter and its key parameters. The = GEM = Gas Electron Multiplier). M denotes Muon, B stands for Barrel and E for Endcap. (based on pT and ΔR) (based on pT and ΔR) Labelling details are given in SectionLabelling 1.2.2. The details magnet are yoke given is inrepresented Section 1.2.2. by the The dark magnet gray yoke is represented by the dark gray 479 concept is therefore applicablesilicon in479 theconcept Outer andscintillator Tracker, is therefore and parts applicable limited of in thesilicon angular in detector the Outer and acceptance arescintillator Tracker, highlighted to and about parts limited in of green in the angular detector and dark acceptance are blue highlighted respectively. to about in green and dark blue respectively. with pixel-strip and calorimeter in endcap areas.Chambersareas. , G a s clustering, clustering,electromagnetic+hadronic electromagnetic+hadronic with480 pixel-strip and calorimeter in endcap Chambers, G a s h < 2.4. 480 h < 2.4. | | | | calibration calibration 481 2.3.2 The Outer Tracker 481 For2.3.2 the The CE-E Outer part, Tracker silicon sensorsFor the are CE-E sandwiched part, silicon between sensors copper-tungsten are sandwiched baseplates between and copper-tungsten baseplatesNear the interaction and region a silicon tracker,Near the composed interaction of region an inner a silicon pixel detector tracker, surroundedcomposed of an inner pixel detector surrounded subtract subtract two-strip modules in region. CE-E measures byElectron a silicon strip detector, measures vertices andMultipliers momenta of charged particles. The elec- , two-strip482 modules in region. CE-E measures byElectron a silicon strip detector, measures vertices andMultipliers momenta of charged particles. The elec- , The Outer Tracker is populatedPCBs482 withThe assembledpT Outermodules, Trackerimplementing into is hexagonalpopulatedPCBs the with L1 modules, trigger assembledpT modules, functionality. andimplementing into mounted hexagonal on the both L1 modules, trigger sides functionality.of and a copper mounted cooling on both plate. sides of a copper cooling plate. tromagnetic calorimeter (ECAL) and the hadronic calorimeter (HCAL) are located inside the electrons/photonselectrons/photons 483 The pT module concept relies on483 theThe factpT thatmodule the strips concept of reliesthe top on and the bottom fact that sensors the strips of a of the top and bottom sensors of a tromagnetic calorimeter (ECAL) and the hadronic calorimeter (HCAL) are located inside the Stainless steel tubes inside theStainless cooling plate steel carry tubes two-phase inside the COcooling2 to maintain plate carry the two-phase silicon sensors CO2 to at maintain a the siliconsolenoid, sensors measuring at a electromagnetic and hadronic showers with lead tungstate crystals and a 484 module are parallel to each other.484 Withmodule the are strip parallel direction to each being other. parallel With to the the stripz axis direction in the being parallel to the z axis in the solenoid, measuring electromagnetic and hadronic showers with lead tungstate crystals and a hadronic calorimeterhadronic calorimeter scintillator-brass sampling detector, respectively.scintillator-brass sampling detector, respectively. o u t e r ro485 ebarrel u g and t i nearly e o rradial n in the rtemperature endcaps, t485ebarrel h g this aand prevents ofi nearly to30 ° theradialC. nelectromagnetic concept Each in the cooling oftemperature endcaps, t stereo h plate strips this a preventsis ofto covering bet30 used° theC. toelectromagnetic conceptan Each azimuthal cooling of stereo angle plate strips isof to covering60 be° used over to anthe showers, azimuthal full radius angle of of 60° over theshowers, fulland radius of upgraded and upgraded Resistive Resistive 486 measure the z coordinate (r coordinate)each486 measure layer. in the Plates the barrelz coordinate of (endcaps). lead and (r coordinate) stainless Foreach this layer. reason steel in the Plates two of barrel the versions of same (endcaps). lead dimensions and stainless For this coverreason steel the two of the sensors versions same and dimensions electronics. cover the sensors andThe electronics. current silicon tracker must beThe replaced current before silicon the tracker start of must Phase-2, be replaced since it will before suffer the start of Phase-2, since it will suffer 487 of pT modules have been realized:487 of modulespT modules with havetwo strip been sensors realized: (2-strip modules or 2S withmodules) two strip sensors (2-strip or 2S modules) significant radiation damage by the endsignificant of Run radiation 3. To maintain damage excellent by the endtrack of reconstruction Run 3. To maintain excellent track reconstruction provide 488 pairsand modules with a strip of and This a488 macro-pixeland assembly hits modules sensor is with called (pixel-strip aa strip a “cassette”, and n orThis aPS macro-pixel dmodules). assembly and 14 sensor CDetailscassettes is called (pixel-strip areE area “cassette”, - assembled or HPS modules). and in 14a m stack Detailscassettes to e are form are a the assembled CE-E.s u in r a stack e to s form at thePlate high CE-E. pileup, the granularity Chambers of bothat high the inner pileup, pixel the tracker granularity and the offor both outer the tracker innerthe will pixel be tracker and the outer tracker will be provide489 pairs of hits a n d C E - H m e a s u r e s Plate Chambers for the provided in Chapter 3. The strips in the 2S modules have a length of about 5 cm, while those 489 Forprovided the CE-H in Chapter part, 3. the The absorber strips inFor theis made the 2S modules CE-H of full part, have disks the a length of absorber stainless of about is steel made 5 cm, forming while of full those disks a rigid of mechanical stainless steel forming a rigidincreased, mechanical by decreasing the pixel sizeincreased, and by shortening by decreasing the stripthe pixel lengths. size and For the by shortening first time the strip lengths. For the first time 490 in the PS modules are about 2.4490 cmin long. the In PS PS modules modules are one about of the 2.4 twocm long. sensors In PSis segmented modules one of the two sensors is segmented at CMS a momentum measurement willat CMS become a momentum possible within measurement a few microseconds, will become and possible this within a few microseconds, and this 491 into macro-pixels of about 1.5structure mm491 length,into macro-pixels in providing which the theof about activez coordinate 1.5 elementsstructure mmmeasurement. length, are in providing which inserted The the the PS as activez cassettescoordinate elements ofmeasurement.60 are°. The inserted60 The° PScassettes as cassettes are of split60°. The 60° cassettesinformation are split can be used in the Level-1information (L1) trigger. can The be usedtrack in trigger the Level-1 will greatly (L1) trigger. sharpen The the track trigger will greatly sharpen the 492 modules are deployed in the first492 three layers of the Outer Tracker, in the radial region of 200– L1 resolution, which will reduce the trigger rate at a given transverse momentum. Thus compatiblecompatible with tracksin 30modules° units with are to deployedkeep theirhadronic in the size tracks firstin manageable.30 three° units layers toof theThe keep Outer first theirhadronic Tracker, 8 size layers inshowers. manageable. the are radial fully region instrumented The of 200– first 8 layerswith showers. silicon are fully The instrumented withmuonpT silicon The system. L1muonpT resolution, which will system. reduce the trigger rate at a given transverse momentum. Thus 493 600 mm, i.e. down to radii at which493 600 the mm, stub i.e.pT downresolution to radii remains at which acceptable the stub andpT theresolution data remains acceptable and the data by combining input from the trackerby and combining muon systems input the frompT thethreshold tracker for and the muon single systems muon the pT thresholdPUPPI for the single muon AlgorithmPUPPI Algorithm for Level-1 for Level-1 Trigger Trigger 494 reduction effective. The 2S modulessensors,494 reduction arewhile deployed effective. the in rest the The outermost of 2S the modules layerssensors, three are are layers,while deployed made in the of in radial rest scintillator the outermost of the layers tiles three in are layers, the made low in the of pseudorapidity radial scintillator tiles region in the low pseudorapiditytrigger can region be kept low despite the hightrigger rate canat HL-LHC. be kept low despite the high rate at HL-LHC. 495 region above 600 mm. In the endcaps495 region the modules above 600 are mm. arranged In the in endcaps rings on the disc-like modules structures, are arranged in rings on disc-like structures, with pT>2with GeV pT >2 atand silicon the GeV sensors inactive the athighand pseudorapidity silicon the sensors elements region. inactive the As high opposed pseudorapidity to the CE-Eelements region. are part, As the opposed detectorsilicon to the CE-E are part, The the endcap detectorsilicon calorimeters will also sufferThe significant endcap calorimeters radiation damage. will also The suffer replacement significant planned radiation damage. The replacement planned 496 with the rings at low radii, up to496 aboutwith 700 the mm, rings equipped at low radii, with upPS tomodules, about 700 while mm, 2S equipped modules with PS modules, while 2S modules Figure 2.4: Average number ofFigure module 2.4: layers Average traversed number by particles,of module including layers traversed both the by Inner particles, including both the Inner for Phase-2, the High Granularity Calorimeterfor Phase-2, (HGCAL), the High will Granularity have an electromagneticCalorimeter (HGCAL), and a will have an electromagnetic and a 497 are used at larger radii. The modulesz 497coordinatesare used are provided atassembled larger byradii. the only The three onmodulesz PScoordinates one barrel side are layers of provided assembled the constrain copper by the only cooling three on PS one plate. barrel side layers of the constrain copper cooling plate. tracks tracksvertexing vertexing Tracker (red) and the Outer TrackerTracker (blue) (red)498 modules,the and origin the as ofOuter well the trigger Tracker as the tracks complete (blue)498 to modules,athe portiontracker origin of (black). as of the well the luminous trigger as Par- the tracks complete region to of a about portiontracker 1 mm, of (black). the which luminous Par- is region of about 1 mm, which is ticle trajectoriesfull are LHC approximatedticle by trajectoriescollision straight499 sufficiently lines, are approximated using precise a flat to partially distributionrate. by straight discriminate499 sufficiently of lines, primary particles using precise vertices acoming flat toand partially distribution from different discriminate ofscintillator vertices. primary particles vertices coming from different vertices. based. full LHC collision rate. and scintillator based. primary primary within z0 < 70 mm, and multiplewithin scatteringz < 70 is mm,not included. and multiple scattering is not included. | | 0 vertex | | –2– –2– vertex

The following section summarizesThe followingthe main concepts section summarizes and features the of main the upgraded concepts trackingand features of the upgraded tracking system. One quarter of the Phase-2system. tracker One quarterlayout can of the be seenPhase-2 in Fig. tracker 2.3. layout Figure can 2.4 be shows seen in Fig. 2.3. Figure 2.4 shows charged pileup chargedcharged pileup particlescharged particles the average number of active layers that are traversed by particles originating from the lumi- charged particlescharged particles the average number of active layers that are traversed by particles originating from the lumi- subtraction subtractionfrom primary vertexfrom primary vertex nousCMS region, for the complete Level-1 trackernousCMS region, as well for as the for complete the Level-1 Inner Trigger tracker Tracker as and well the as Outer for the Tracker Inner Trigger Trackerfor and the HL-LHC Outer Tracker for HL-LHC separately. separately. particle particle The number of layers has beenThe optimised number to of ensure layers robusthas been tracking, optimised i.e. basically to ensure unaffected robust tracking, i.e. basically unaffected performance when one detectingperformance layer is lost when in some one parts detecting of the layer rapidity is lost acceptance. in some parts The of six the rapidity acceptance. The six flow flow layersAn of the Outer all Tracker arelayers the newAn minimum of the Outer required all Tracker FPGA-based to ensure are the new robust minimum track required finding FPGA-based at to the ensure L1 robust Level track finding at the L1 1 Level Trigger 1 will Trigger process will process data from data CMS from CMS PUPPI PUPPIreweighted reweighted trigger in the rapidity acceptancetrigger of h in< the2.4, rapidity as discussed acceptance in more of detailsh < 2.4, in Sectionas discussed 3.1. in more details in Section 3.1. | | neutral particlesneutral particles | | weight weightneutral particlesneutral particles subdetectorsubdetector trigger primitive trigger primitivegenerators generators (TPGs). (TPGs). 3.6. New Trigger Objects3.6. New based Trigger3.6. onNew Objects Particle Trigger basedFlow Objects3.6. Reconstruction onNew Particle based Trigger on Flow Objects Particle Reconstruction based Flow Reconstruction on Particle Flow27 Reconstruction 27 27 27 tracker tracker calorimeters calorimeters muon detectorsmuon detectors The PUPPIThe weight PUPPI for weight neutral for particle neutral i particle is computed i is computed using a lookupusing a table lookup table trigger-primitivetrigger-primitive clusters ortrigger-primitive trigger-primitive clusters ortrigger-primitive clusters trigger-primitivemuons. or We trigger-primitive analyse2 clusters muons. the linked or We trigger-primitive muons. analyse trigger-primitive2 We the analyse linked muons. the trigger-primitive linked We analyse trigger-primitive the linked trigger-primitive information to createinformation a globalpT,j tolist create of L1 a particle globalpT,j candidates:list of L1 particle muons, candidates: charged hadrons, muons, charged electrons, hadrons, electrons, barrel barrelbarrel barrelforward forward resistive cathoderesistive gascathode gas information to createinformation a global↵i list to= create of L1 particlea global↵i candidates: list= of L1 particle muons,⇥(R candidates:0 chargedRij) hadrons, muons,⇥(R0 charged electrons,Rij) hadrons, electrons, endcap endcap drift drift addressedaddressed by the metric byphotons, the andmetric neutralphotons, hadrons. andRij Weneutral then hadrons. run theRij L1 We PUPPI then run algorithm the L1 PUPPIon the listalgorithm of global on particle the list of global particle tracker tracker electromagnetic electromagnetichadronic hadronichadronic hadronic plate stripplate electronstrip electron photons, and neutralphotons, hadrons. and We neutralj thencharged hadrons. run from theL1 PV We PUPPI✓j thencharged run algorithm◆ from theL1 PV PUPPIon✓ the list algorithm◆ of global on particle the list of global particle calorimeter calorimeter tubes tubes 2 X 2 X calorimeter calorimetercalorimeter calorimetercalorimeter calorimeter chambers chamberschambersmultiplierschambers multipliers candidates to filtercandidates the eventcandidates to into filter the the most to event filter leading,candidates theinto event the vertex most to into filter compatible leading, the the most event vertex particle leading, into compatible thecandidates. vertex most compatible particle leading, As candidates. vertex particle compatible candidates. As particle As candidates. As AlgorithmwithAlgorithm the PF Performance algorithm,with the the PFwith PUPPI Performance algorithm, the algorithmic PF algorithm, thewith PUPPI complexity the the algorithmic PF PUPPI algorithm, is reduced algorithmic complexity the and PUPPI only complexity is reduced uses algorithmic simplified is and reduced only complexity uses and simplified only is reduced uses simplified and only uses simplified integer operations.integer Since operations. PUPPIinteger requires operations. Since input PUPPIinteger Since information requires operations. PUPPI input requires from Since information track input PUPPI vertexing information requires from but track the input from vertexing PF information track but vertexing the from PF but track the vertexing PF but the PF TPG TPGTPG TPGTPG TPGTPG TPGTPG TPGTPG TPGTPG TPGTPG TPGTPG TPG algorithm does not,algorithm such aalgorithm vertexing does not, stepdoes such is anot,algorithm naturally vertexing such a performed vertexingstep does is not, naturally stepsuch in parallel is a performed naturallyvertexing with the performed step in PFparallel is oper- naturally in with parallel the performed PF with oper- the in parallelPF oper- with the PF oper- Performanceations.Performance The is output studiedations. of the The combined ations.is outputusing studied The PF+PUPPI of output the simulated combinedations. ofusing trigger the The combined PF+PUPPI algorithm output simulated t-tbar PF+PUPPI of trigger is the a set combined algorithm ofevents triggervertex t-tbar PF+PUPPI filtered algorithmis a setdecaying can-ofevents triggervertex is a set filtered algorithm of vertex decayingsemileptonically can- filtered is a set ofcan- vertex semileptonically filtered can- didates that can thendidates be used thatdidates can to reconstruct then that be can used thendidates and to beidentify reconstruct usedthat can totriggerable, reconstruct then and beidentify used prompt and totriggerable, identify reconstruct physics triggerable, objects. prompt and identify physics prompt triggerable, objects. physics objects. prompt physics objects. as a signalThisas approach with a signal to large pileup mitigationwith This missing approach large needs tothe This transversepileup missingprimary approach mitigation vertex to to transversepileup needs be energy properly mitigation the primary reconstructed, needs(MET) vertex energy the to primary as be and properly (MET) vertex scalar reconstructed, to be and properly sum asscalar reconstructed, of jet sum as of jet barrel muon barrel muonendcap muon endcap muon This approach to pileup mitigation needs the primary vertex to be properly reconstructed, as barrel calorimeterbarrel trigger calorimeter trigger can be done easilycan in beevents done with easily a large in events multiplicity with a large of high multiplicityp tracks of(see high Sectiontracks 3.2). (see Section 3.2). track finder track finder track finder trackenergy finder (HcanenergyT be) and done easily minimum-bias(Hcan inT eventsbe) and done with easily minimum-bias a large in events multiplicity events with a large of highas multiplicity events pbackgroundT tracks (see of highas Section pbackgroundT tracks 3.2).. (seeT Section 3.2).. pT

CMS Phase-2 SimulationCMS Phase-2CMS,=140Phase-2 Simulation SimulationCMS,=140Phase-2,=140 Simulation ,=140

O(10) Tbps O(10) Tbps ~ 2.5 μs ~ 2.5 μs Semileptonic tt SemileptonicSemileptonic tt tt Semileptonic tt correlatorcorrelator 10 10 10 10 Rate [MHz] Rate [MHz] Rate [MHz] Rate [MHz] global trigger global trigger 1 1 1 1

The CorrelatorThe Correlator combines combines all of this all information of this information into trigger into objects, trigger whichobjects, are which are MET (Calo) MET (Calo) MET (Calo) MET (Calo) sent to thesent Global to the Trigger Global to Trigger make theto make Level-1 the trigger Level-1 decision. trigger decision. MET (TK Δz) MET (TK Δz) MET (TK Δz) MET (TK Δz) −1 −1 −1 −1 10 10 10 MET (PF+PUPPI) 10 MET (PF+PUPPI)MET (PF+PUPPI) MET (PF+PUPPI)

Particle-LevelParticle-Level Reconstruction Reconstruction 0.1 0.2 0.3 0.4 0.50.1 0.60.2 0.7 0.30.1 0.8 0.40.2 0.9 0.5 0.31 0.6 0.4 0.7 0.50.1 0.8 0.60.2 0.9 0.7 0.31 0.8 0.4 0.9 0.51 0.6 0.7 0.8 0.9 1 Signal effciencySignal effciencySignal effciencySignal effciency Figure 3.8: (Left) efficiencyFigure 3.8: for (Left) selecting efficiency signal for and selecting background signal for and three background different forEmiss threetrigger different Emiss triggermiss miss Particle FlowParticle reconstruction Flow reconstruction correlates correlates detector detector signals to signals identify to final-state identify final-state Figure 3.8: (Left) efficiencyFigure 3.8: for (Left) selecting efficiency signal for and selecting backgroundT signal forand three backgroundT different forET threetrigger different ET trigger 34 34 5 Performance in simulation5 Performance in simulation algorithms and (Right)algorithms efficiencyalgorithms and (Right)turn-on and efficiency curves (Right)algorithms for efficiencyturn-on three and different curves (Right) turn-on forH efficiencyT curves threetrigger different for working turn-on threeH differentT curves points:trigger forH working threetrigger different points: working points:trigger working points: particles particles and measure and Secondly, measure their in low properties. Secondly, pileup their environments, in low properties. pileup This there environments, approach is less This information there approach is less provides available information locally provides available better just due locally to better just due to T HT FPGA aImplementation calorimeter-onlyFPGAFigure aImplementation algorithmcalorimeter-only 4.aMET calorimeter-only (purple); trigger algorithm a track-only ratea calorimeter-onlyalgorithmversus (purple); trigger signal a (purple); track-only algorithm effi algorithmciency a track-only trigger using (left) (purple); tracks algorithm and triggerH consistent a track-onlytrigger algorithmusing tracks turn-on trigger using consistent curve tracksalgorithm (right) consistent using for the tracks consistent the lack of pileup.the This lack ofmeans pileup. the This↵ distribution means the is↵ notdistribution as well populated is not as andwell the populated uncertainty and the uncertaintywith the primarywith vertex the (red); primary and vertex a PF+PUPPI (red); and trigger a PF+PUPPI algorithm trigger (blue). algorithm The H thresholds (blue).T The H thresholds performanceperformance than single-detector than single-detector reconstruction reconstruction and is already and is used already in the used in the with the primarywith vertex the (red); primary and vertex a PF+PUPPI (red); and trigger aT PF+PUPPI algorithm trigger (blue).T algorithm The HT thresholds (blue). The HT thresholds were chosen so thatparticle-flowwere each chosenH trigger andso that PUPPI path each based correspondsH trigger trigger path to (blue), a corresponds rate a ofcalorimeter-only 20 kHz. to a The rate study of trigger 20 kHz. was (purple), The study and a was trigger based on on PU is larger.on PU is larger. Particle flowParticle and flow PUPPI wereT and chosen are PUPPI so inherentlythatwereT each chosen areHT trigger so inherentlythat regional path each correspondsHT trigger regional algorithms.path to a corresponds rate of 20 kHz.algorithms. to a AsTherate studyof a 20 kHz. was proof-of- AsThe study a was proof-of- High LevelHigh Trigger Level and Trigger offline and reconstruction. offline reconstruction. conducted usingtracks aconducted background consistentconducted using sample a with background using of the minimum-bias primary aconducted background sample vertex using of collisions sample minimum-bias (red) a background [4]. ofand minimum-bias a signal collisions sample sample ofand collisions minimum-bias from a signal t¯t and sample a signal collisions from sample t¯t and from a signal t¯t sample from t¯t Anti-kT (R=0.7), n = 80 Anti-kT (R=0.7), n = 80 Anti-kT (R=0.7) Anti-kT (R=0.7)JHEP 10 (2014) 59 JHEPprinciple, 10 (2014) 59 simulated we divide eventssimulated decaying the semileptonically, events detector decaying correspondingsemileptonically, into η tox corresponding an ɸ average ~ 0.55 PUη to of an 140. ɸ averagex 0.55 PU ofregions 140. accepting up PU CMS-PRF-14-001 principle, we divide the detector into x ~ 0.55 x 0.55 regions accepting up PU CMS-PRF-14-001 simulated eventssimulated decaying semileptonically, events decaying semileptonically,corresponding to corresponding an average PU to of an 140. average PU of 140. 0.3 0.3 0.2 0.2 LV LV LV Anti-k , R = 0.4 Anti-kCalo , R = 0.4 Calo Anti-k LV , R = 0.4 Anti-kCalo , R = 0.4 Calo |η0.6| < 2.5 CMS |η0.6|

LV PFlowLV PFlow LV LV neutral neutral 0.15 0.15 trigger algorithms5trigger based Conclusions algorithms on a single based detector and on (e.g. a single Outlook either detector calorimeter-only (e.g. either or calorimeter-only tracker-only). As or tracker-only). As hadron hadron HCAL HCAL PFlowCHS PFlowCHS PFlowCHS PFlowCHS ImplementedImplementedtrigger in algorithms Xilinxtrigger based Vivado in algorithms on aXilinx single basedHLS detector Vivado on v2016.4 (e.g.a single either HLS detector calorimeter-only forv2016.4 (e.g. an either Ulstrascale+ orcalorimeter-only tracker-only). for an Ulstrascale+ Asor tracker-only). VU9P As VU9P clusters 0.2 clusters 0.2 an early proxy ofan the early potential proxy gains of the expected potential by gains this approach,expected by we this study approach, the missing we study trans- the missing trans- 0.4 0.4 PUPPI PUPPI0.4 PUPPI0.4 PUPPI an early proxy ofan the early potential proxy gains of the expected potential by gains this expectedapproach, by we this study approach, the missing we study trans- the missing trans- miss miss miss verse momentumFPGA,verse (E T four momentum)FPGA, and the regions scalar (E T four summed) and the can regionspT scalarover summedbeall jets in processedmiss canp anT over event, beall typically jets in processed an referred with event, typically ~40% referred with of ~40% the FPGA of the FPGA We exploreverse the momentum feasibilityverse (E ofT momentumusing) and the particle-flow scalar (ET summed) and reconstruction thepT scalarover summed all jets and inp an PUPPIT over event, all pileup typically jets in anmitigation, referred event, typically referred , (pT - pT , (pT - pT Detector Detector , (pT - pT 0.1 , (pT - pT to as H . Gains are also expected for jet substructure studies for heavy-particle tagging, and Energy resolution Energy resolution 0.1 to as H . Gains are also expected for jet substructure studies for heavy-particle tagging, and photon Energy resolution Energy resolution T photon σ σ T σ σ the best performingto as HT. Gains algorithms areto as alsoHT expected. in Gains current are for also o jetffl substructureine expected processing, for studies jet substructure in for the heavy-particle CMS studies Level-1 for tagging, heavy-particle trigger and cor- tagging, and lepton isolation,resources. thoughlepton isolation, weresources. leave The those though topics welatencies leave for The future those studies. topics latencies for for An future example Particle studies. of thefor An early example Particle per- Flow of the early and per- Flow PUPPI and are PUPPI are 0.1 lepton isolation, thoughlepton isolation, we leave thoughthose topics we leave for future those studies. topics for An future example studies. of the An early example per- of the early per- Particle Flow Particle Flow 0.2 0.1 0.2 0.2 0.2 formance of the L1relatorformance PF+PUPPI to of maintain the algorithm, L1 PF+PUPPI performance using a algorithm, Phase-2 in detector theusing high a simulation Phase-2 pileup detector but conditions using simulation tracks of the but usingHL-LHC. tracks We present fitted fitted fitted fitted approximatelyapproximatelyformance 500 of the L1nsformance PF+PUPPI and 500 of the100 algorithm, L1ns PF+PUPPI ns,and using respectively. 100 algorithm,a Phase-2 ns, detector using respectively. a simulationPhase-2 detector but using simulation tracks but using tracks ECAL ECAL 0.05 0.05 reconstructed withproof-of-principlereconstructed an offline algorithm with an implementations offline and mocked-up algorithm to and of have thesemocked-up similar algorithms performance to havefor similar (resolu- FPGAs performance and find (resolu- that they signif- clusters clusters reconstructed withreconstructed an offline algorithm with an offline and mocked-up algorithm to and have mocked-up similar performance to have similar (resolu- performance (resolu- tion, pT thresholdicantlytion, of 2 GeV,pT threshold improvetion, etc) as thatthreshold of the 2 expected GeV, physics etc) of from2 as GeV, performance that the etc) expected L1 as track that fromfinder, of expected the the is Level-1 L1provided from track the finder,trigger in L1 Fig. track is 3.8. while provided finder, being is in provided Fig. feasible 3.8. in Fig. in terms 3.8. pT tion, pT threshold of 2 GeV, etc) asmiss that expected frommiss the L1 track finder, is provided in Fig. 3.8. The figures showThe the figuressignal vs show background the signal selection vs background efficiency selection for different efficiencyET fortriggers, different as ET triggers,miss as miss 0 0 0 0 0 0 of FPGAThe resource figures usage showThe the and figures signal latency. vs show background the signal selection vs background efficiency selection for different efficiencyET fortriggers, different as ET triggers, as 0 10020 2000100 200 300 10020 1000 400 200100 200 500 3001000 40020 500 10050200 20 1000 10010050200 1000 100 OutlookwellOutlook as the efficiencywell turn-on as the efficiency curves for turn-on different curvesHT triggers, for different usingH aT semi-leptonictriggers, using t¯t a signal semi-leptonic t¯t signal well as the efficiencywell turn-onas the efficiency curves for turn-on different curvesHT triggers, for different usingHT atriggers, semi-leptonic using t¯t a signal semi-leptonic t¯t signal RefpT (GeV) RefpT (GeV) Ref nRefPU nPU miss miss p (GeV) p (GeV) p (GeV) p (GeV) having true ET having, andThe they true concept compareET , of and three adapting theymiss different compare algorithms algorithms threemiss different used based in algorithmso onffline standalone processing based calori- on to standalone the FPGAs calori- of the Level-1 T T T T having true ET having, and they true compareET , and three they different compare algorithms three different based algorithms on standalone based calori- on standalone calori- Ref Ref meter, standalonetriggermeter, track, standalone mayas well give as PF+PUPPI.track, rise to as additional well The asH PF+PUPPI.T efficiency opportunities The turn-onHT toefficiency curves improve show turn-on physics three curves performance. show three We note in Figure 13: Jet energyFigure resolution 13: Jet asenergy a function resolution of p as ain function the barrel of (left)p in and the in barrel the endcap (left) and in the endcapThese proof-of-principleThese proof-of-principlemeter, studies standalonemeter, track, indicate studies standalone as well as track, PF+PUPPI. the indicate as wellfeasibility The asH PF+PUPPI.T theefficiency feasibility ofThe turn-on H performingT efficiency curves of show turn-on performing three curves Particle show three Particle Figure 7:JetpFigureT resolution 7:Jet vs.p presolutionT (left) for vs.nPUpT=(left) 80 for for⌘n< T2=.5 80 and for jet⌘pT

Figure 8: TheFigure single jet 8: mass The resolution single jet mass for nPU resolution= 80 for for jetsnPU with= 80100 for GeV jets

First we look atFirst jet mass we look for central at jet mass jets with for central 100 GeV jets

– 15 – – 15 – [9] J. Duarte, et al., “Fast inference of deep neural networks in FPGAs for particle physics”, arXiv:1804:06913 [physics.ins-det] (2018).