Large-R Jet Tagging Algorithm (ATLAS GEP)

This repository contains code for the development and performance studies of a large-R jet tagging algorithm for the ATLAS Global Event Processor.

---

Directory Structure

algorithm/

Contains the HLS-based basic (v2) & advanced (v3) algorithms, their respective testbenches, code for reading and writing test vector files, and a standalone C++-based emulation which is synchronized to be bit-wise (input-output) accurate with the HLS algorithm.

analysis/

Contains the NTupling code (reads in TrigGepPerf/GEPOutputReader output & JETM42 DAOD) to generate algorithm input test vectors and ROOT NTuples, code for generating event displays, and code for analyzing the output of the jet tagger algorithm and performing rate estimates. Only LRJNTupler.C, largeRJetAnalysisAndRates.C, and makeEventDisplays.C are actively used and under-development.

config/

Contains .py scripts for running/synthesizing the HLS basic & advanced algorithms.

data/

Is a placeholder directory, though contains scripts for converting from jet tagger test vector formats --> GEP-FW test vector format, and generating random data in jet tagger test vector format.

project/

Contains a deprecated .tcl script for running/synthesizing the HLS algorithm.

---

Instructions for Different Scripts

NOTE: for all different scripts, most parameters are changed within the script rather than as arguments in the terminal, and all paths are configurable, but take default values of what the author uses on either the U Chicago Tier 3 Analaysis Facility (login.af.uchicago.edu) or the Miller group workstation (millerlabml01.uchicago.edu).

---

HLS Algorithm

To run the HLS algorithm or synthesis, simply do:

python run_hls(_adv).py

After configuring within the script which algorithm configurations to run over and the arguments for jet_tagger_hls(_adv).py, which actually calls HLS.

Use: - hls_exec = 0 for C-sim only

• hls_exec = 1 for synthesis only
• hls_exec = 2 for C-sim, synthesis & co-sim
• hls_exec = 3 for C-sim, synthesis, co-sim, & implementation

When running, it will output within config/w/ a directory which contains the output from running HLS, which is configured based on the algorithm parameters used for that run.

---

Standalone Emulation

The standalone emulation is run through a configuration script:

algorithm/emulation/jetTaggerConfigLocal.sh

Parameters

1. algorithm version (2, 3)

2. rMergeCuts (d_search): configures the distance away from the original seeds to search for proto-seeds to take the midpoint between and shift the final jet seed positions there. Typical values used include 0.001 (disabled), 1.5, 1.75 & 2. Values should not exceed 2 times the jet radius to prevent loss of E_T from the original jet.

3. rSquaredCuts: jet radius squared, typical values include 1, 1.21, 1.44

4. nIOs: number of considered input objects to be used for reclustering and E_T computation of each jet. Typical values when using towers as input objects: 64, 128, 256. Tower multiplicity when applying SoftKiller PU-suppression typically does not exceed 128. The value used when using WTA-cone jets as input objects is 8 (10 = max jets per event - 2 = fixed number of seeds).

5. nSeeds: number of seeds and consequently outputted large-R jets per event.

6. signals: use (true false) to process both signals and background, or only (true) or (false) to process only signals or backgrounds, respectively.

7. signalStrings: a list of signal strings to process, currently supported signal strings are: ("VBF_hh_bbbb_cvv0" "VBF_hh_bbbb_cvv1" "ggF_hh_bbbb" "ZvvHbb" "ttbar_had" "Zprime_ttbar")

8. inputObjects: types of input objects used for reclustering, currently supported are: ("gepWTAConeCellsTowersJets" "gepCellsTowers"). Note that the basic algorithm can only use gepWTAConeCellsTowersJets. Deprecated input objects include CaloTopoTowers, GEPBasicClusters (still functional), and 422 Topo Clusters.

9. seedObjects: type of objects used for seeding & subjet finding, currently supported are: ("gepWTAConeCellsTowersJets" "jFEXSRJ" "gFEXSRJ")

10. condor: whether to run on condor, currently unsupported and unnecessary since emulation runs very quickly.

---

LUT & Constants Generation

The emulation also relies on having already run:

root -b
.L writeLUTsConstantsEmulation.cc
main()

This writes the used constants and Look Up Table (LUT) header files for the algorithm provided the hard-coded constants for each algorithm version. Parameters here include the number of bits allocated to different data-types, total bits for jet tagger TOBs, & the rSquared and rMerge cuts used.

In order to run a certain configuration of the algorithm, the constants and LUTs must have been already generated by this script.

---

Emulation Workflow

• jetTaggerConfigLocal.sh routes the correct constants and LUT header files for each algorithm configuration to be included in algorithm/emulation/UsedConstantsLUTs/

• jetTaggerEmulation.cc routes the correct input NTuple based on the signal / background options, and sets the correct output file name based on the algorithm configuration processed.

---

NTupling Script

Running the standalone emulation requires having generated NTuples from combining a GEPOutputReader NTuple and JETM42 Derivation DAOD file. In the future, it should just take GEPOutputReader NTuple as input.

The NTupling script (LRJNTupler.C) uses constants (bit-widths, ranges) hard-coded in analysis/analyisHelperFunctions.h. These constants must be synchronized with what are used when running the emulation.

The script can be run through:

root -b -l -q 'LRJNTupler.C'

Paths used are hard-coded in functions: - makeFileDir

• makeMemPrintFileNames
• makeOutputFileName

---

Analysis Script

Running the analysis script requires having already run the standalone emulation, and providing the correct path to the signal and background used.

The script allows processing multiple signals and backgrounds in one run, and will output object property & performance plots for each signal and background, and all files processed overlaid.

Output Directories

• largeRJetPlots/: plots of object properties and other misc. plots
• rateVsEff/: performance plots evaluating algorithm by rate estimates and ROC curves for different selections applied to jet tagger TOBs
• overlayJZSliceHistograms/: plots showing reconstructed object E_T distributions for different JZ slices (0-9) which can be used to evaluate energy resolution and algorithm performance compared to truth & offline jets
• overlayMultipleFiles/: object property & performance plots for all different files looped through

Run with:

root -b -l -q 'largeRJetAnalysisAndRates.C'

---

Workflow Diagram

---

Contact

Any questions can be directed to the repository author:

mark.steven.larson@cern.ch