The Evolution and Performance of the ATLAS

Calorimeter-Based Triggers in 2011 and 2012Ivana Hristova, On behalf of the ATLAS Collaboration

Abstract—An overview is presented of the ATLAS triggers thatuse information from the calorimeter system. The identificationof high transverse-energy electrons, photons, τ leptons and jets,as well high missing and total transverse energy, was optimisedfor the high-luminosity and high-pile-up running conditions atLHC in 2011 and 2012. Results are shown that demonstratethe achieved trigger robustness: the selection of possible physicssignatures was performed with high efficiency while maintainingthe rates within their required limits.

Index Terms—ATLAS, Trigger, Calorimeter

I. INTRODUCTION

THE physics at the dawn of the tera-electronvolt scale

may remain within the scope of the Standard Model

and its extensions or require another theory to describe new

phenomena. Either scenario will have implications on our

understanding of matter. The ATLAS detector [1] has been

recording high-energy proton-proton (pp) collisions at the

Large Hadron Collider (LHC) [2] at CERN since 2009 [3]–

[5]. After minimum-bias events at a centre-of-mass energy

of 900 GeV were taken in 2009, throughout 2010 the num-

ber of bunches per beam increased stepwise up to 368

and the instantaneous (2010 integrated) luminosity reached

2.1×1032 cm−2 s−1 (45 pb−1) at a fixed centre-of-mass

energy of 7 TeV with 150 ns spacing and nominal intensities

of the bunches. The peak pile-up, i.e., the average number of

interactions per bunch crossing including the main collision

that triggered the recording at peak instantaneous luminosity,

was 3 at the maximum, well below the design value of 23. In

2011 the beam setup was upgraded: the number of bunches

per beam reached 1380, separated by 50 ns and had intensities

up to 40% above the nominal design value, which produced

instantaneous (2011 integrated) luminosity and peak pile-up

of 3.65×1033 cm−2 s−1 (5.61 fb−1) and 17. In 2012, as the

energy increased to 8 TeV and the beam parameters were

further changed to allow for faster accumulation of physics

data, the experiment operated at conditions with pile-up above

the design. The maximum peak luminosity and pile-up at the

start of an LHC beam fill increased to 6.7×1033 cm−2 s−1

and 32. The integrated luminosity was about 6.5 fb−1, as of

June 2012. This evolution presented a challenge to the trigger

and data acquisition systems which were continuously tuned

to the running conditions.

In conjunction with information from other sub-detectors,

the capability of the calorimeter system [6], [7] (Fig. 1)

to detect candidate electrons, photons, τ leptons, jets, and

missing and total transverse energy in collision events was

used at the trigger level to select in real time (online) and

Fig. 1. A schematic view of the ATLAS calorimeter system, subdivided intoLAr and Tile calorimeters [16]. The electromagnetic part consists of LAr inthe barrel and end-cap regions. The hadronic part is comprised of Tile in thebarrel and LAr in the end-cap region. The forward region is covered by LAr.

store to permanent storage only data of interest for further

(offline) physics analysis. An introduction to the trigger and

data acquisition (DAQ) systems (TDAQ) [8], [9] and their

operation during 2011 and the first half of 2012 are presented

in Section II, followed by highlights from the implementation

and performance of the first-level calorimeter trigger [15] in

Section III. The results from the optimisations of the high-level

triggers and their evolution are shown in Section IV.

II. TRIGGER AND DATA ACQUISITION

The purpose of the trigger system is to reduce the input

collision rate from a maximum of 40 MHz to an output rate

which can be sustained by the DAQ system with minimum

deadtime [17]. Its core functionality is the processing of

detector information using fast algorithms that select a variety

of physics events with high efficiency and reject background.

The ATLAS trigger consists of three levels. At the first level

(L1), which is based on custom hardware, the trigger decision

is made within less than ∼2.5 µs and the rate is reduced to

about 60 kHz. Only coarse-granularity information from the

calorimeter and muon detectors is used. The higher software-

based trigger levels, together denoted as HLT, further reduce

the rate to 5 kHz within 60 ms at level-2 (L2) and 0.4 kHz

within 1000 ms at the event filter (EF). The L2 algorithms use

fine-granularity data only within the regions of interest (RoI)

identified at L1 which correspond to about 2% of the whole

detector. The complete event information is available at the

978-1-4673-1084-0/12/$31.00 ©2012 IEEE

EF. The given numbers are representative of the 2011-2012

operation.

The 2011-2012 data-taking period is characterised by steady

increase of the peak pile-up, up to 17 in 2011 and in 2012

up to 32, i.e., well above the design value of 23. Most of the

trigger rates increase linearly with instantaneous luminosity as

expected, whereas certain trigger selections reveal strong pile-

up dependence and their rates rise much faster than linear.

The trigger system is configured on a run-by-run basis (where

a run typically spans an LHC fill) via a trigger menu which

consists of logical combinations of trigger objects (items) at

L1 that seed the sequences of algorithms (chains) at HLT.

To keep the various rates within their allocated budgets the

trigger items and chains are prescaled during a run, i.e.,

only a predefined fraction of the input bunch crossings are

accepted by applying prescale factors to each of the triggers.

For adapting to foreseen changes in the operating conditions

various optimisations of the software and hardware resources

are incorporated during regular interventions to the TDAQ

system. For example, in 2011 the requests to the readout

system PCs were optimised both in software and hadrware

to maximise the amount of data-taking rate [18].

The 2011-2012 evolution of the trigger menu itself and the

tuning of the selection criteria to the LHC conditions require

a detailed analysis of data and simulation samples. Such

trigger analyses are continuously carried out and the required

changes are tested [19] and implemented during technical

stops between the data-taking periods. Particular results of

those studies are described in more detail in the following

sections.

III. LEVEL-1 CALORIMETER TRIGGER

The L1 calorimeter trigger (L1Calo) is a hardware-based

digital system which uses custom electronics to process the

analog trigger-tower signals from the calorimeter. The signals

are first digitised, then a look-up table (LUT) performs the

final calibration [22], [23] of the transverse energy, ET, as

well as additional signal transformations such as pedestal sub-

traction, treatment of saturated pulses, and noise suppression

(Section III-D). The LUT output values are the inputs to the

L1Calo algorithmic processing (Section III-A) and they are

also stored in a FIFO for readout to the DAQ system and—

since end of 2011—usage in the HLT trigger (Section IV-C).

A. L1Calo Trigger Signatures

The L1Calo system searches for signatures with high ET

values, above a set of configurable thresholds, that can be

classified as isolated electrons or photons (EM), τ ’s (TAU), jets

(J), and missing transverse energy (XE). Jets in the forward

calorimeter region (FJ) can also be identified. Table I lists the

thresholds for single-, double- (2EM) and multi- (4J, 5J) object

L1 primary trigger items and their evolution with pile-up and

luminosity.

The identification of these objects is based on FPGA

algorithms which process the coarse-granularity information

organised in 5,248 LAr and 1,920 Tile projective trigger

towers covering the electromagnetic (Em) and hadronic (Had)

Fig. 2. Evolution of the LHC peak luminosity in 2011 (top) and 2012(bottom) [20].

calorimeters. As input to the sliding-window algorithm, a

2×2 trigger-tower grid in η × φ is used to find EM RoIs

and determine the sets of ET thresholds they passed. The

presudorapidity is defined as η = − ln tan θ

2with θ and φ

the polar and azimuthal angles, respectively, while an RoI is a

cone-shaped portion of the detector uniquely specified by its

(η, φ) axis coordinates and vertex at the interaction point. In

addition to the Em cluster, a group of 2×2 Had trigger towers

is used for searches of TAU RoIs. The J RoIs are similarly

identified, however using an 8×8 grid in both the Em and

Had layers, while the XE trigger relies on the full sum of all

trigger-tower energies.

To enhance the selection and control the rates of objects

with low unprescaled thresholds, the L1Calo algorithms allow

for additional requirements, e.g., that the transverse energy in

an isolation (I) region in the electromagnetic calorimeter is less

that 5 GeV and the electromagnetic-shower energy deposited

in the hadronic (H) calorimeter is less than 1 GeV. Also the

actual threshold values are allowed to vary (V) with η by a few

GeV around the central threshold value (see [10] for details).

This η-dependence is introduced to account for the varying

amount of material in front of the calorimeter which is not

corrected for in the raw L1Calo energy.

The number of different types of signatures passing the

corresponding thresholds are counted and these multiplicities

(e.g. 2EM, 5J) are sent to the central trigger processor for use

in the L1-accept decision. The overall trend with increasing

luminosity and pile-up (Table I) is to raise the lowest thresh-

TABLE IEVOLUTION OF THE L1CALO PRIMARY TRIGGER ITEMS IN 2011-2012.

L1 ET threshold, max. max. lumi.,

item [GeV] pile-up 1033 cm−2

s−1

EM 16 11 2.416VH 30 17 3.918VH 30 32 6.7

2EM 10 14 11 2.410VH 12 17 3.910VH 12 32 6.7

TAU 30 11 2.430 17 3.940 32 6.7

2TAU 11 11 2.411 17 3.920 32 6.7

J 75 11 2.475 17 3.975 32 6.7

4J 10 11 2.410 17 3.915 32 6.7

5J 10 20 17 3.910 15 20 32 6.7

XE 40 11 2.450 17 1.340 BGRP7 32 6.7

FJ 50 11 2.475 17 3.975 32 6.7

olds and make use of the isolation criteria. To control the XE

rates, as well as the rates of J and XE combined triggers, and

to keep their thresholds low L1Calo pile-up noise suppression

is implemented in the forward calorimeter (FCal in Fig. 1,

Section III-D). Also a special bunch group (BGRP7) is used

that vetoes triggers in the first few beam bunches per train

where the calorimeter response remains elevated due to out-

of-time pile-up.

The performance of selected L1Calo main triggers which

seed the lowest-threshold unprescaled HLT triggers and their

improvements are presented in the following subsections.

B. Electron/Photon Trigger

With increasing luminosity the EM thresholds are raised,

however, not significantly in order to preserve the acceptance

for physics analyses. For the optimisation of the trigger rates

the hadronic leakage requirement and the η-variable thresholds

are used, e.g., to define the EM16VH item. For that trigger

the ET threshold is increased up to 18 GeV in those η

regions where the L1 efficiency is sufficiently high to have

no effect on the event acceptance at the HLT. As a result the

L1 EM16VH efficiency curve is shifted towards higher values

of the transverse momentum, pT compared to EM16, as shown

in Fig. 3 (top). Although here the hadronic veto requirement

(H) has no effect on the L1 efficiency normalised to offline

identified electrons as a hadron leakage cut is also applied

offline relative to the electron pT value, it has a significant

effect on the rate reduction.

C. Tau/Hadron Trigger

All the single-object L1 TAU triggers were well behaved in

2011-2012. They scale linearly with luminosity and show no

Fig. 3. Top: Efficiencies of the L1 EM16 and EM16VH triggers as a functionof the offline transverse momentum, pT [10]. Bottom: Isolation energy ofL1 objects associated to the EF tau20 medium1 (Section IV-B) chain andmatched to offline probe τ ’s within ∆R < 0.2. The offline candidates areselected using a tag-and-probe analysis with Z → ττ events collected byATLAS during Summer 2011 [12].

pile-up dependence. Similar to the EM cluster selection, the

TAU cluster algorithm evolution is characterised by adjustment

of the lowest threshold values and tune of the prescale factors.

Unlike for the EM algorithm, an EM isolation requirement is

added for certain threshold values in order to further reduce

the L1 output rate of those items without increasing their

threshold. The energy in the EM isolation ring is required

to be less than a fixed value of 5 GeV. The choice of this

value is based on the analysis of 2011-data and Z → ττ MC

samples using the tag-and-probe method. The good data-to-

MC agreement, shown in Fig. 3 (bottom), implies a sufficiently

reasonable description and understanding of the underlying τ -

shower shape variables in MC which is used to tune the algo-

rithm. The shower shape variables are used for τ identification

and they describe the longitudinal and lateral distribution of

the τ hadronic decay particles as they develop into a shower

by further decays, propagation and energy depositions in the

calorimeter.

D. Jet/Missing-Transverse Energy Trigger and Pile-up Noise

Suppression in FCal

Jets are abundant in collision events and a characteristic

feature of most physics as well as background processes. The

high-pT jet threshold, J75 in Table I, is kept unmodified,

however a larger number of multi-jet trigger items is added

in the menu as luminosity increases.

For trigger items based on ET-sums from all trigger towers

including FCal (Fig. 1) the L1 rates increase faster than lin-

early with luminosity. In addition to including special trigger

items in the menu in 2012 (XE40 BGRP7 in Table I), L1Calo

noise cuts are applied to minimise the pile-up dependence of

the XE triggers and reduce their rates. The optimisation of the

L1Calo pile-up noise cuts and their effect on the XE trigger

performance are described below.

In addition to the ET calibration, the L1Calo LUT allows

several non-linear signal operations to be carried out simulta-

neously, one of which is to suppress the contribution to the

trigger from towers with very small signals (typically ∼1 GeV)

that are most likely due to noise by setting their output

ADC values to zero. The noise thresholds are configurable

and different values can be set in different regions of the

detector. Initially, in the 2009-2010 runs the energy is required

to be positive and above 1.0-1.3 GeV. FCal is particularly

sensitive to pile-up as the increase of the latter implies higher

occupancies and larger deposited energies which in addition

scale with the larger size, i.e., ∆η × ∆φ = 0.4 × 0.4,

of the trigger towers in this forward calorimeter region at

3.1 < |η| < 4.9 compared with the barrel region at |η| < 2.5.

The pile-up noise distribution in FCal is studied using a

sample of 2011-2012 collision data taken with an unbiased

random trigger. The ADC values of the FCal trigger towers,

after pedestal subtraction, are expected to vary around zero

and the width of their distribution can be used as a measure

of the noise. Figure 4 (top) shows the noise distributions in

ADC counts for four η regions between 3.1 and 4.9 of the FCal

electromagnetic layer (FCAL1). As expected, the width of the

noise distribution increases with η and, almost proportionally,

with the η-bin size. The noise width as a function of the

number of interactions per bunch crossing (pile-up) is shown

in Fig. 4 (bottom) for the four η regions and pile-up values

between 8 and 25. A linear dependence of the noise vs. the

pile-up is observed, with different slopes for the different η

regions. While in the most central η-bin the noise is found to

be pile-up independent, the slopes increase the more forward

are the regions that are covered by the corresponding bins.

These results are used to derive η- and pile-up-dependent noise

threshold cuts which are first implemented at the beginning

of 2012 and adjusted to the evolving luminosity and pile-up

conditions. The pile-up noise cuts are increased in two steps

from the nominal of ∼1 GeV up to 6.5 GeV and then up to

10.9 GeV for average pile-up values of 15 and 25, respectively.

Figure 5 shows the resulting XE rate reduction based on

studies performed at the end of 2011: to reduce the maximum

input rate to, e.g., 10 kHz the threshold can be lowered from

∼40 GeV without noise cuts to ∼26 GeV with noise cuts.

For the XE50 trigger four different choices of noise cuts are

simulated and the efficiency is found to remain essentially

unchanged, as presented in Figure 6. Here, the tested FCal

noise cuts are optimised for pile-up of 15 and 20, and in

addition to FCal, cuts are applied to the trigger towers of the

electromagnetic end-cap inner wheel for pile-up of 20 and 25,

while the data used for the analysis are taken at pile-up of 23.

IV. HIGH LEVEL TRIGGER

At the HLT, sophisticated offline-type event selection algo-

rithms running on commercial computer processor farms allow

for a more flexible control over the output trigger rates at a

Fig. 4. Distribution of the noise in the FCal calorimeter for nominal trigger-tower sizes, ∆η×∆φ = 0.4× 0.4, in regions of η, i.e., bin1: 3.1-3.2, bin2:3.2-3.5, bin3: 3.5-4.2, bin4: 4.2-4.9 (top) and the pile-up dependence of thewidth (or RMS denoted as σ) of the noise distributions for the same η bins(bottom) [11].

Fig. 5. The XE trigger rates as a function of threshold for several pile-up noise cut scenarios. The rates are estimated by applying noise cuts toZeroBias events at luminosity ∼3.2×1033 cm−2 s−1 and pile-up of 15. The2011 configuration corresponds to noise cuts of ∼1 GeV in all trigger towers.The loose forward noise cut applies cuts of 6.5, 5.5 and 2.5 GeV in the firstFCal layer and 4.5 GeV in the second layer at |η| > 3.5. The tighter noisecuts raise these by 1 GeV, and also raise the noise cuts in all other towersbeyond |η| = 2.5 by 0.5 GeV. The final case removes FCal entirely from theXE calculation. At this luminosity, the bulk of the fake XE rate is achievedwith the loose cuts [11].

minimum loss of efficiency with respect to L1. The difference

between L2 and EF refers mostly to the degree of complexity

and precision of the algorithmic processing. The following

description of the HLT triggers relies on the adopted trigger-

chain notation, e.g. 2g20X medium, is used to identify two

photons (γ’s) with each photon passing a threshold of 20 GeV

and additional requirement(s) X. The label extension which

can be ”loose”, ”medium” or ”tight” denotes the operating

point of the trigger, i.e., the physics performance in terms of

Fig. 6. The L1 XE50 efficiency as a function of the offline topologicalcluster-based missing ET for a sample of candidate W → eν events (singleelectron with ET above 25 GeV, passing tight identification and other qualityrequirements, with the candidate W transverse mass greater than 40 GeV)for four different choices of the FCal noise cuts (see text) [11].

the efficiency of the affected signal channel and the rate for

the chosen selection.

Several basic strategies can be used to improve the rejection

performance while maintaining the signal performance of the

HLT trigger chains:

• Less separation between L1 and HLT thresholds, while

trying to preserve the same plateau of the efficiency turn-

on curve.

• Tighter selection criteria on the shower shape variables.

• Higher ET thresholds.

• Possibility to introduce isolation on single electron trig-

gers to keep rates below the allowed average bandwidth.

Optimisation results and evolution for the different signa-

tures are described in the following subsections. A table for

each signature lists the evolution of the lowest unprescaled

trigger chains in 2011-2012. It is worth underlining that

during this period of time the load on the TDAQ system was

constantly increasing and a tremendous amount of work was

carried out to achieve an optimal performance of the high

level trigger that in turn ensured a high efficiency with which

physics data were collected.

A. Electron/Photon Signature

Each EM RoI provided by L1 is refined at the HLT in

order to identify either an electron or a photon signature

(with possible overlap between the two signatures). The HLT

selection is based on shower shape variables reconstructed

from the fine-granularity information of the Em calorimeters.

In addition, only for electron identification, a search for high-

pT tracks in the tracking detectors is performed, followed by

a match between calorimeter showers and charged tracks.

During the 2011-2012 operation, the difference between the

EF and L1 thresholds is progressively reduced in order to keep

the L1 rate within the allowed bandwidth without raising the

threshold at the HLT and hence at the offline analysis. For

example, the seed of the 2e12 medium trigger is changed from

2EM7 to 2EM10, i.e, using a tighter (T) L1 threshold. The

electron identification criteria at L2 are brought closer to those

at EF, while at EF the cuts are re-optimised to avoid raising

Fig. 7. EF trigger rates as function of instantaneous luminosity for singleelectron (top) and photon (bottom) triggers.

the threshold further. For example, the change from medium

to medium1 (Figs. 8, 9) requirements includes tighter cuts on

the shower shapes. Finally, at the EF the energy thresholds are

raised in steps from 20 to 22 to 24 GeV for electrons and from

60 to 80 to 120 GeV for photons, with increasing luminosity.

Figure 7 shows the EF rates of the single electron (top) and

photon (bottom) triggers as a function of the luminosity and

illustrates the achieved rate reduction.

The performance of the electron signature is evaluated

with a data-driven tag-and-probe method on Z → ee events.

Figure 8 (top) shows the efficiencies for the lowest unprescaled

single electron triggers. The efficiency of e20 medium is

computed relative to offline electrons with pT>21 GeV and the

efficiencies e22 medium and e22vh medium1 are computed

relative to offline electrons with pT>23 GeV. Between 25 and

35 GeV the efficiency is slowly increasing before finally reach-

ing the plateau value at about 35 GeV. Inefficiencies of these

triggers mainly arise from the resolution of reconstruction and

identification variables at the HLT with respect to offline.

Photon trigger efficiencies were computed for offline photon

candidates satisfying tight identification criteria and results are

shown in Fig. 8 (bottom) as a function of the offline photon

pT. Rapid turn-on curves are observed with no significant

dependence on the offline value and efficiency values close

to 100% are measured in the plateau.

The electron and photon (γ) trigger menu and its evolution

with luminosity and pile-up is shown in Table II.

B. Tau/Hadron Signature

The L2 algorithms unpack the calorimeter cells contained

within a region ∆η × ∆φ = 0.8 × 0.8 centred around the

(η, φ) position of the TAU-cluster RoI identified at L1. This

full-granularity information is used to refine the RoI position

and ET, and to compute various shower shape variables as

Fig. 8. Efficiencies as function of the offline pT for the electron (top) andphoton (bottom) triggers [10].

TABLE IIEVOLUTION OF THE ELECTRON AND PHOTON HLT PRIMARY SIGNATURES

IN 2011-2012.

L1 EF signature yearseed

EM14 e20 medium 2011EM16 e22 mediumEM16VH e22vh medium2EM5 2e10 medium2EM7 2e12 medium2EM10 2e12T medium2EM10VH 2e12Tvh mediumEM30 g60 looseEM30 g80 loose2EM14 2g20 loose

EM18VH e24vhi medium 20122EM10VH 2e12Tvh looseEM18VH e24vh medium e7 mediumEM30 g120 looseEM30 2g40 loose2EM10VH 2g20vh medium2EM12 EM16V g35 g25 medium

well as track-based parameters which are used as selection

criteria to differentiate between τ ’s and jets similarly as done

in the offline analysis. In 2011 the output rates saturate

with increasing luminosity (Fig. 9, top) and the efficiencies

significantly decrease as a function of pile-up (Fig. 10, top)

for the main τ triggers (single, double and combined). To

avoid further performance degradation in 2012, the cut on the

ET-weighted shower radius in the electromagnetic calorimeter

is reduced from 0.4 to 0.2 thus significantly reducing the

amount of pile-up picked up by the selection algorithm. Also,

instead of using all tracks only those with compatible impact

parameters (|∆Z| < 2 mm) with the leading (i.e., highest-pT)

track are considered, where ∆Z is with respect to the primary

(i.e., highest pT-sum) vertex. In addition to the track-based

isolation (i) (i.e. a cut on the ratio of the tracks pT-sum in

Fig. 9. EF trigger rates as function of the instantaneous luminosity during2011 (top) and 2012 (bottom).

0.1 < ∆R < 0.3 to that in ∆R < 0.1 from the leading track),

a cut is introduced on a calorimeter-based isolation variable

defined as the ratio of ET in ∆R < 0.1 to that in ∆R < 0.2.

Separate cuts are used for the 1-prong and multi-prong τ -

decays.

For the τ identification in 2012 variables and algorithms

more similar to those in the offline reconstruction are used

at the EF. The selection is changed from the cut-based to the

multivariate techniques developed offline, i.e. boosted decision

trees (BDT) and log-likelihood. Figures 9, 10 (bottom) show

the improved performance of the τ trigger signatures selected

with the modified algorithms. The rates (Fig. 9) remain linear

up to the highest luminosity despite the harsher data-taking

conditions in 2012 relative to 2011. The loss of efficiency

at high pile-up is recovered (Fig. 10) after re-processing the

2011 data with the BDT-based HLT event selection (shown

here with respect to the offline BDT analysis).

The τ trigger menu and its evolution with luminosity and

pile-up is shown in Table III.

C. Jet Signature

The jet selection is based mostly on information from the

calorimeter. The coarser ∆η × ∆φ granularity of 0.2 × 0.2(compared to 0.1× 0.1 for EM and TAU) used by the L1Calo

sliding-window algorithm [15] for the search of jet clusters

leads to a poor jet position resolution, particularly for low ET

jets. To improve this, a new jet algorithm referred to as L2

full scan, L2FS, is developed and implemented in 2011.

The L2FS jet algorithm is a L2 software trigger, however

instead of being seeded by L1 RoIs it uses all L1Calo trigger

Fig. 10. Trigger efficiencies as function of the number of interaction vertices(pile-up) for 2011 data (top) and for the same data re-processed with improved2012 HLT algorithms (bottom) [12].

TABLE IIIEVOLUTION OF THE τ HLT PRIMARY SIGNATURES IN 2011-2012.

L1 EF signature yearseed

TAU50 tau100 medium 2011TAU50 tau125 medium(1)2TAU8 TAU11 tau29 medium tau20 medium2TAU11 TAU15 tau29T medium tau20T medium2TAU6 EM10 tau16 loose e15 medium2TAU8 EM10 tau20 medium e15 medium2TAU8I EM10VH tau20i medium e15vh mediumTAU11 XE20 tau29 medium xe35TAU15 XE25 3J10 tau29T medium xe35 3L1J10

TAU40 tau125 medium 20122TAU11I TAU15 tau29Ti medium tau20Ti medium2TAU11I EM14VH tau20Ti medium e18vh mediumTAU15I XE35 tau29Ti medium xe40 tight

towers as input. This allows a more sophisticated jet search

with an unseeded anti-kT cluster algorithm to be carried

out. Furthermore, instead of L1 RoIs, the L2FS-identified

jets can be used to seed an anti-kT algorithm at L2 that

uses the associated jet calorimeter cells. Such selection is

most beneficial for broad jets and events with complex event

topologies.

Figure 11 (top) presents a comparison of the jet position

resolution for the different jet trigger algorithms (online) using

the position identified by the reconstruction software (offline)

as a reference. The resolution at L2FS is comparable to that

at L2 and is significantly improved with respect to that at

L1. In addition Fig. 11 (bottom) shows that a ∼ 10% gain in

efficiency is achieved for the selection of six-jet events when

using the L2FS anti-kT algorithm based on the L1Calo trigger

towers instead of the default L1 sliding-window algorithm.

Fig. 11. Top: The jet position resolution (in η, evaluated with respect tooffline anti-kT R=0.4 jets) of the L1, L2FS (denoted as L1.5) and L2 jettriggers in 2011 pp collisions. The jet finding algorithm for L1 is a 0.8×0.8sliding window, for L2FS (denoted as L1.5) it is anti-kT with R=0.4 and forL2 a three-iteration cone R=0.4 seeded by a L1 jet. Bottom: The efficiencyfor L1 (sliding window) and L2FS (anti-kT R=0.4) jets to satisfy a six jettrigger in events where at least six jets have been identified offline. The gainin efficiency is achieved due to the different jet algorithms used [13].

In comparison with the 2010 jet EF algorithms, running in

pass-through mode, i.e. rejecting no events, a full scan (EFFS)

of the RoIs is implemented as of 2011. From the unpacked

RoI information topological clusters (tc) of calorimeter cells

calibrated at the electromagnetic (em) scale are used as input to

the anti-kT jet finding algorithm with radius R = 0.4, 1.0 (a4,

a10). Also as of 2011, removal of noisy calorimeter cells and

pile-up noise suppression at L2 are used. The jet clusters are

calibrated at the em-scale at L2 and EF. Calibration to a local

weight (lcw) scale for selected EF chains and jet calibration

at the hadronic (had) scale at L2 and EF are used as of 2012.

The jet trigger menu and its evolution are shown in Table V.

D. Missing Transverse-Energy Signature

The reconstruction of the missing transverse energy, here

denoted as Emiss

Tfor brevity, relies entirely on data from the

full calorimeter. The 2011 Emiss

Tvalue at L2 is equivalent to

that computed by L1Calo at L1 using the RoI information

and the trigger-tower energies. After a significant upgrade in

2012 of the LAr, Tile and the DAQ systems it is possible to

access the cell-based ET-sums provided by the calorimeter

front-end boards and implement a computation of the Emiss

T

at the L2 trigger using this information. Figure 12 (top) shows

TABLE IVEVOLUTION OF THE JET HLT PRIMARY SIGNATURES IN 2011-2012.

L1 EF signature yearseed

J75 j180 a4tc 2011J75 j240 a10tcJ75 j240 a4tcJ75 j240 a10tc5J10 5j30 a4tc5J10 j80 a4tc

J75 j360 a4tchad 2012J75 j360 a10tcemJ75 j360 a4tclcwJ75 j360 a10tclcw4J15 5j55 a4tchad L2FS

the residual distributions (σ) of the Emiss

Tvalues at L2 with

respect to those at EF, denoted as EF−L1 in 2011 and EF−L2

in 2012. The resolution improves by about 50% which leads

to an increase of the rejection rate (at a given L2 threshold)

by a factor of 5-6. Thus in 2012 lower L1 XE thresholds are

used in comparison with 2011.

At the EF, the following cell-based algorithm to compute

the Emiss

Tis of limited applicability and used only in 2011. It

is based on ET-sums from the ∼182,468 LAr and ∼5,192 Tile

calorimeter readout cells whose ET values are above a certain

noise threshold, σ. Cells are selected with a one-sided noise

cut, i.e. the cell ET value is required to be larger than 3 ×σ. However, the method is susceptible to pile-up because the

large ET deposited in the calorimeter gives rise to fluctuations

which in turn produce large fake Emiss

Tmeasurements.

To overcome this, in 2012 a topological cluster algorithm

at the EF is used instead. Three-dimensional clusters are

built around cell seeds with ET above 4 × σ by iteratively

collecting cells with ET above 2 × σ and finally adding

all immediate neighbouring cells of those gathered at the

previous step, thus the last step effectively adds cells with

ET above 0× σ. This resembles the offline algorithm except

that the local weight calibration (tclcw) for the cell energy

corrections is not dependent on the bunch crossing. Despite the

more severe conditions in 2012, this change results in about

25% improvement of the resolution in comparison with 2011

when only uncalibrated cell-based clusters are used. Figure 12

(bottom) shows the overall improvement of the Emiss

Ttrigger in

terms of the efficiency turn-on curve for the lowest unprescaled

trigger chain used in 2011 and 2012. The efficiencies below

the plateau increase by up to 50%. The turn-on point of the

curve at which the 99% efficiency is reached improves by

∼20 GeV.

The Emiss

Tand jet+Emiss

Ttrigger menus and their evolution

are shown in Tables V and VI.

V. SUMMARY

In this document an overview is given of the evolution

and optimisation studies performed for the calorimeter-based

triggers. At L1 η-dependent noise cuts were implemented in

the forward calorimeter region for high pile-up conditions.

At HLT more advanced techniques and improved algorithms

compared to those initially used at the start of data-taking

were applied in order to perform more offline-type selection,

Fig. 12. Top: The resolution of the x component of Emiss

Tfor 2012 data

compared with the corresponding 2011 algorithm: the EF−L2 and EF−L1residuals are shown, where the latter corresponds to the L2 resolution in 2011.Bottom: The improvement on the turn-on curve for simulated ZH → ννbb(using Pythia and mH =120 GeV) using either the lowest unprescaled triggerchain in 2012 (L1 XE50 → L2 xe55 → EF xe60 verytight) or in 2012(L1 XE40 BGRP7 → L2 xe45T → EF xe80T tclcw loose) [14].

TABLE VEVOLUTION OF THE Emiss

THLT PRIMARY SIGNATURES IN 2011-2012.

L1 EF signature yearseed

XE40 xe60 2011XE50 xe60XE50 xe70

XE40 BGRP7 xe80T tclcw 2012XE50 xe80 tclcwXE40 xe80 tclcw

TABLE VIEVOLUTION OF THE JET+Emiss

THLT PRIMARY SIGNATURES IN

2011-2012.

L1 EF signature yearseed

J50 XE20 j75 a4tc EFFS xe45 loose 2011J50 XE35 j75 a4tc EFFS xe55

J30 XE40 j80 a4tchad xe100 tclcw veryloose 2012J75 j170 a4tchad EFxe80 tclcw

i.e. using larger amounts, finer-granularity and more precise

readout information. The presented results demonstrate the

robustness of the trigger system in the high luminosity, high

pile-up environment of the LHC in 2011 and 2012.

ACKNOWLEDGMENT

The author thanks the ATLAS Collaboration and Humboldt

Universitat zu Berlin for supporting the participation at the

18-th Real Time Conference, June 11-15, 2012, Berkeley,

CA, as well as Ricardo Goncalo, David Francis, Dora Merelli

and the conference organisers for help with the preparation.

Martin Wessels, Stephen Hillier, and the HLT signature group

convenors gave comments on the talk. Brian Petersen gave

most elaborate comments that shaped the present document.

Thanks to Hans Peter Beck for a very detailed review.

REFERENCES

[1] ATLAS Collaboration, The ATLAS Experiment at the

CERN Large Hadron Collider, 2008 JINST 3 S08003,doi:10.1088/1748-0221/3/08/S08003.

[2] L. Evans and P. Bryant (editors), LHC Machine, 2008 JINST 3 S08001,doi:10.1088/1748-0221/3/08/S08001.

[3] ATLAS Collaboration (Ed.), The Performance of the ATLAS Detector,Springer-Verlag, ISBN 978-3-642-22115-6, Reprinted from Eur.Phys. J.C, Volume 70 (2010) and 71 (2011).

[4] ATLAS Collaboration, Performance of the ATLAS Detector using First

Collision Data, JHEP 1009 (2010) 056, doi:10.1007/JHEP09(2010)056.[5] ATLAS Collaboration, Performance of the ATLAS Trig-

ger System in 2010, Eur.Phys.J. C72 (2012) 1849,doi:10.1140/epjc/s10052-011-1849-1.

[6] ATLAS Collaboration, Liquid Argon Calorimeter Technical Design

Report, CERN-LHCC-96-041, http://cdsweb.cern.ch/record/331061.[7] ATLAS Collaboration, Tile Calorimeter Technical Design Report,

CERN-LHCC-96-042, http://cdsweb.cern.ch/record/331062.[8] ATLAS Collaboration, First-Level Trigger Technical Design Report,

CERN-LHCC-98-014, http://cdsweb.cern.ch/record/381429.[9] ATLAS Collaboration, ATLAS High-Level Trigger, Data Acquisi-

tion and Controls Technical Design Report, CERN-LHCC-2003-022,http://cdsweb.cern.ch/record/616089.

[10] ATLAS Collaboration, Performance of the Electron and Pho-

ton Trigger in pp Collisions at√s=7 TeV with the AT-

LAS Detector at the LHC in 2011, ATL-COM-DAQ-2012-008,https://cdsweb.cern.ch/record/1426717.

[11] ATLAS Collaboration,https://twiki.cern.ch/twiki/bin/view/AtlasPublic/L1CaloTriggerPublicResults.

[12] ATLAS Collaboration,https://twiki.cern.ch/twiki/bin/view/AtlasPublic/TauTriggerPublicResults.

[13] ATLAS Collaboration,https://twiki.cern.ch/twiki/bin/view/AtlasPublic/JetTriggerPublicResults.

[14] ATLAS Collaboration,https://twiki.cern.ch/twiki/bin/view/AtlasPublic/MissingEtTriggerPublicResults.

[15] R. Achenbach et al, The ATLAS Level-1 Calorimeter Trigger, 2008JINST 3 P03001, doi:10.1088/1748-0221/3/03/P03001.

[16] ATLAS Collaboration,http://www.atlas.ch/photos/calorimeters-combined-barrel.html.

[17] T. Kono, on behalf of the ATLAS Collaboration, Per-

formance and Evolution of the ATLAS TDAQ System

with pp Collisions at 7 TeV, These proceedings,https://cdsweb.cern.ch/record/1455265/files/ATL-DAQ-PROC-2012-011.pdf.

[18] E. Lipeles, R. Ospanov and D. Schaefer, Proceedings, Computing inHigh Energy and Nuclear Physics, New York, NY, USA, 21-25 May2012, https://cdsweb.cern.ch/record/1458520.

[19] M. zur Nedden, on behalf of the ATLAS Collaboration, Experience

with the Custom-Developed ATLAS Offline Trigger Monitoring

Framework and Reprocessing Infrastructure, These proceedings,https://cdsweb.cern.ch/record/1455670/files/ATL-DAQ-PROC-2012-019.pdf.

[20] ATLAS Collaboration,https://twiki.cern.ch/twiki/bin/view/AtlasPublic/LuminosityPublicResults.

[21] C. Boulahouache, The ATLAS LAr Calorimeter Level 1 Trigger Signal

Pre-Processing System: Installation, Commissioning and Calibration

Results, IEEE Transactions on Nuclear Science, Vol. 57, pp. 642-649,2010, doi:10.1109/TNS.2009.2035118.

[22] J. T. Childers (for ATLAS Collaboration), ATLAS Level-1 Calorimeter

Trigger Hardware: Initial Timing and Energy Calibration, 2011 J.Phys.: Conf. Ser. 293 012061, doi:10.1088/1742-6596/293/1/012061.

[23] M. Wessels (for ATLAS Collaboration), Calibration and Performance

of the ATLAS Level-1 Calorimeter Trigger with LHC Collision Data,Proceedings, 2nd International Conference on Technology and Instru-mentation in Particle Physics, Chicago, IL, USA, 9-14 Jun 2011,http://cdsweb.cern.ch/record/1382111/.