Big Data & Machine Learning in HEP

Transcription

1 Big Data & Machine Learning in HEP Mike Williams Department of Physics & Laboratory for Nuclear Science Massachusetts Institute of Technology March 17, 216

2 The Large Hadron Collider Outline { Big Data Real-Time Analysis ML Propaganda

3 Big Data & Real-Time Processing

4 Big Data LHC detectors are BIG in both size and data output. E.g., ATLAS weighs over 7 tonnes and has million channels. Figure 1: Schematic diagram of the ATLAS detector at the LHC, which weighs over 7 tons The beam-crossing rate at the LHC is sensor 4 MHz, in each and has about million electronic channels in its arrayswhich that areresults used to record data detector from multiple proton-proton collisions that occur at a rate of 4 MHz. per year! producing ~petabyte / second of data or zettabytes

5 Big Data Figure 1: Schematic diagram of the ATLAS detector at the LHC, which weighs over 7 ton Custom-built electronics used to read out the sensor arrays all perform zero and has about million electronic channels in its sensor arrays that are used to record data from suppression to drastically reduce the data rates. multiple proton-proton collisions that occur at a rate of 4 MHz post zero-suppression data rates per experiment (ATLAS & CMS) data [EB] data rate [TB/s] 18 1 year By the HL-LHC era, each experiment will need to process in real time roughly 3 Figure 2: Data volumes that must be processed by the data-ingestion system of the CMS experi 15 TB/s or 6 EB/year (comparable to largest-scale industrial data sets). ment: (left axis, red circles) peak data rate in terabytes per second and (right axis, blue solid line 5

6 Real-Time Processing Each experiment currently implements both hardware and software triggering stages (cascading approach). I ll focus on LHCb (most advanced trigger). ~TB/s (post zero suppression) ~5 GB/s Feature-building in custom electronics (e.g. FPGAs) used to greatly reduce rate, but online computing systems still have herculean tasks. 6

7 Real-Time Processing 5 GB/s Real-time reconstruction for all tracks with pt >.5 GeV. 27, physical CPU cores (>5k logical cores) are used to run the full offline reconstruction in real time. 8 GB/s Data buffered on disk while alignment/calibration done. 8 GB/s Real-time reconstruction for all tracks, final PID, etc., all available to select events. PB of disk used for the buffering while final alignment/calibration done. Final event selection done with access to best-quality data (mostly done during down time between fills)..7 GB/s 7

8 Real-Time Processing In Run 3, LHCb will increase the post-zero-suppression data rate to 5 TB/s, but also remove the hardware trigger. The full reconstruction will be done in real time on all events. 5 TB/s (post zero suppression) ATLAS/CMS will go to 15 TB/s in Run 4, but for hermetic detectors it s not feasible to go straight to software (wifi data transfer being investigated). 8

9 Offline Processing The bulk of offline computing is done using the Worldwide LHC Computing Grid (WLCG) which has 17 computing centers around the world and runs over 2M jobs / day. Currently, the LHC experiments persist about 3 PB / year of data -- but this still isn t enough. Many strategies being explored (scouting, partial-event storage, data compression, etc) to increase the physics output. 9

10 Machine Learning in HEP

11 ML in LHCb Trigger 5 GB/s Real-time reconstruction for all tracks with pt >.5 GeV. Two ML algorithms used in this stage that select about 7% of the output BW. 8 GB/s Data buffered on disk while alignment/calibration done. 8 GB/s Final event selection is currently roughly 3% ML based (first introduced into the main trigger algorithm for the start of 211 data taking). LHCb trigger using a custom-built ML algorithm. Real-time reconstruction for all tracks, final PID, etc., all available to select events..7 GB/s V.Gligorov, MW, JINST 8 (212) P

12 Ghost Killers & PID Ghost-killing neural networks do an excellent job of removing fake tracks (which is a big problem with the occupancies of the LHC). ) 2 candidates / (4 MeV/c LHCb preliminary all D Kπ fakes Fraction of candidates per bin LHCb Simulated τ µ µ + µ Calibrated τ µ µ + µ Data sidebands M PID response m Kπ 2 [MeV/c ] Particle ID is now commonly done using neural networks as well. These combine the info from all subsystems and exploit t h e m a n y ( o f t e n o b s c u r e ) correlations between features to ID particles. 12

13 Higgs to CMS When the LHC started, people were still worried about using ML. CMS actually published both cut-based and BDT-based Higgs exclusions for the WW decay mode -- so we can see directly what is gained: 95% C.L. Limit on / SM 2 1 CMS Preliminary s=7 TeV, L = 4.9 fb -1 s=8 TeV, L = 19.5 fb H WW 2l2 /1-jet cut based -1 Observed Median Expected Expected ± 1 Expected ± 2 Injection m =125 GeV ± 1 H m H [GeV] 13

14 Higgs to CMS When the LHC started, people were still worried about using ML. CMS actually published both cut-based and BDT-based Higgs exclusions for the WW decay mode -- so we can see directly what is gained: 95% C.L. Limit on / SM 2 1 CMS Preliminary -1 s=7 TeV, L = 4.9 fb s=8 TeV, L = 19.5 fb H WW 2l2 /1-jet -1 Observed Median Expected Expected ± 1 Expected ± 2 Injection m =125 GeV ± 1 H m H [GeV] 14

15 Using BDTs There are 3 standard uses for classification: (1) place a cut on the BDT response; (2) fit the BDT response; or (3) analyze the data in bins of BDT response. For example, the recent LHCb pentaquark discovery analysis trained a BDT to identify the decay Λb J/ψpK then placed a cut on the BDT response for each candidate which gives the mass plot below. The BDT information was then not used further in the analysis (it just selects events to study). Events / ( 4 MeV) LHCb 26k Λb 5% BKGD in signal region LHCb-PAPER [MeV] m J/ψ K p 15

16 Entries /.25 Tag jets using a SV then feed properties to 2 BDTs. Fit the 2-D BDT response distribution in data to extract the b-jet, c-jet and lightparton-jet yields. Each BDT uses features as input (ML as dimensional-reduction). Data MC(B) Data MC(S+B) 7 Data χ 2 /dof -3.5= CMS -1 s = 7TeV, L = 5. fb -1 s = 8TeV, L = 18.9 fb pp VH; H bb VH Fit BDT Response Background VH(bb) 125 GeV CMS, PRD 89 (214) 123 Background uncert. -4 χ 2 /dof -3.5 = log(s/b) BDT(b c) LHCb data 19 light parton JINST (215) P613 LHCb-PAPER beauty charm Initial (no-tagging) sample: D+jet 7% light parton, 22% charm, 8% beauty BDT(bc udsg) Figure 6: From a b-jet and c-je Most Higgs analyses now also work this way; i.e., they involve fitting BDT distributions. Essentially, one can reduce a large-n-d fitting problem into a 1-D fitting problem (while still using the N-D info). (left: VH(bb), 18 features input). responses observed in data (ann are found); (middle) projection BDT templates shown here wer were used to calibrate the BDT 16

17 Bs μμ Calibrate BDT to have uniform response on signal, bin data in BDT response and analyze simultaneously. ] 3 Candidates / (4 MeV/c 2 ) [ Candidates / (4 MeV/c 2 ) LHCb, BDT [.,.25) Data Signal and background Bs + B + Combinatorial background Semi-leptonic background Peaking background m [GeV/c 2 + ] LHCb, BDT [.6,.7) m [GeV/c 2 + ] Candidates / (4 MeV/c 2 ) Candidates / (4 MeV/c 2 ) LHCb, BDT [.25,.4) m [GeV/c 2 + ] LHCb, BDT [.7,.8) m [GeV/c 2 + ] Candidates / (4 MeV/c 2 ) Candidates / (4 MeV/c 2 ) LHCb, BDT [.4,.5) m [GeV/c 2 + ] LHCb, BDT [.8,.9) m [GeV/c 2 + ] Candidates / (4 MeV/c 2 ) Candidates / (4 MeV/c 2 ) andidates / (4 MeV/c 2 ) LHCb, BDT [.5,.6) m [GeV/c 2 + ] LHCb, BDT [.9,1.] m [GeV/c 2 + ] CMS, 7TeV, FR, BDT [.26,1.] Combining CMS&LHCb data gives > 5σ and a BF ~ 3x -9 7(consistent with the SM expected value). CMS&LHCb, Nature 522, 68 (215). 17

18 Regression ML is also now being used for regression. For example, CMS has moved to ML-based jet-energy corrections (e.g., for Higgs to bb). ) -1 (GeV 1/N dn/dm bb CMS Simulation m H = 125 GeV (set A) Regressed Raw PEAK = GeV FWHM = 27. GeV PEAK = GeV FWHM = 32.8 GeV CMS, PRD 92 (215) m bb (GeV) 18

19 Advanced Topics ML algorithms where the response is de-correlated from some set of userdefined features. See Stevens and Williams [ ]; Rogozhnikova, Bukva, Gligorov, Ustyuzhanin, Williams [14.414]. Used to search for a new boson in PRL 115 (215) 16182, and currently being used in many papers to come out soon. Deep learning is actively being explored as a way to remove the human-led feature-design stage. See Baldi, Sadowski, Whiteson [ ]. Parametrized classifiers to interpolate the optimal selection when only a discrete subset of possible signals has been simulated. See Baldi, Cranmer, Faucett, Sadowski, Whiteson [ ]. Hyper-parameter tuning using Bayesian global optimization. Blending classifiers, etc, etc... 19

20 Tools Most physicists are using either ROOT s TMVA package or the python scikitlearn (sklearn for short). Some limited usage of proprietary packages too. See also pypi.python.org/pypi/hep_ml/.2. for some custom HEP algorithms. 2

21 Propaganda

22 Collaboration The few of us (physicists) who do research in this area are convinced that direct collaboration/engagement with computer/data scientists is critical. We d like to open the box to make it easier for those communities to do research on our (awesome) problems (please contact me if interested). 22

23 Title xxx 23

24 Advertisement ML school for Ph.D. students in Lund (follows one last year in Moscow). Go to bit.ly/mlhep216, or me, or 24

25 Summary HEP experiments are VERY big data. The online systems must process some of the world s largest data sets. Amazingly, we re now seeing real-time reconstruction, calibration, and alignment -- and soon we ll see a triggerless-readout system at the LHC! Not long ago, using a ML algorithm was considered unsafe ; however, now it s rare to see a HEP data analysis that does not use ML algorithms. ML is used to kill fake tracks, do particle ID, classify events, regression, dimensional reduction, etc. ML algorithms are even starting to dominate real-time (trigger) selections. There are many great (free) tools out there that are pretty easy to use. It s never been easier to get started. For students, there are dedicated summer schools to help. For everyone, there are now dedicated workshops. 25