Proc. Natl. Acad. Sci. USA Vol. 90, pp. 9754-9757, November 1993 Colloquium Paper

This paper was presented at a coUoquium entled "Images of Science: Science ofImages," organized by Albert V. Crewe, held January 13 and 14, 1992, at the National Academy of Sciences, Washington, DC. Pattern recognition at the Fermilab collider and Superconducting Supercollider HENRY J. FRISCH The Enrico Fermi Institute and Physics Department, University of Chicago, Chicago, IL 60637-1433

ABSTRACT In a colliding beam accelerator such as Fer- particles such as protons and neutrons. Strongly interacting milab or the Superconducting Supercollider (SSC) protons, or particles are called hadrons generically. antiprotons, collide at a rate between 105 (Fermilab) and 108 The Standard Model agrees with experimental data to (SSC) collisions per second. In real time experimentalists have exquisite precision. So what is the problem? First, we to select those events which are candidates for exploring the haven't seen the top quark yet (we haven't directly detected limit of known phenomena at a much lower rate, 1-100 per the tau neutrino either but seem more content about it). Also, second, for recording on permanent media. The rate of events there are indications from experiments detecting neutrinos from new physics sources is expected to be much lower, as low from the sun and from cosmic rays that we may not under- as a few per year. This is a severe problem in pattern stand the relationships among the three types of neutrinos. recognition: with an input data stream of up to 1015 potential However there are much more fundamental problems in bits per second in its images, we have to pick out those images the Standard Model that force us to the conclusion that it is that are potentially interesting in real time at a discrimination internally inconsistent and, hence, incomplete. Calculations level of 1 part in 106, with a known efficiency. I will describe of very high energy scattering of the W bosons give diver- the overall filtering strategies and the custom hardware to do gences-unphysical answers. Our experience has been that this event selection (a.k.a. pattern recognition). when this problem occurs in a theory the solution is that there are new effects at the energy regime where the problem is and that these new effects cancel or otherwise ameliorate the What Physics Problem Are We Trying to Solve? unphysical calculated effects. We therefore believe there is "new physics" in the TeV energy region; this is the basis for High-energy physics has evolved a fairly complete picture of the Superconducting Supercollider (SSC). In addition to this the fundamental forces, the underlying natural symmetries, inconsistency, there is the disturbing ad hoc nature of the and the fundamental particles. We call it "The Standard Standard Model-so much is "put in by hand." Why leptons Model" (model because much of it is ad hoc, and standard and quarks? Why the parallel structure ofthree pairs ofeach? slightly facetiously, but implying that it is the common core Why three? Why is the W boson left-handed, and why is there of belief that will be embellished as we theorize). [For an no corresponding right-handed boson? And so forth. introduction for the scientist not in high-energy physics, see, The strategy we have adopted, such as it is, is to probe to for example, Gottfried and Weisskopf(1). A detailed descrip- shorter and shorter distances by building bigger and bigger tion at the advanced undergraduate physics level is given in accelerators (microscopes). To search for the rare interac- one of the standard tions that may be the indication of "new physics" we go to textbooks used in the field (2).] higher interaction rates. And amid events that are more and The ingredients are the particles we consider elementary- more complicated we have to recognize the patterns ofthese i.e., not composite and hence point-like-which come in two new categories. Fermions have values of intrinsic spin that are interesting events. This is the subject of this talk. half-integer multiples of Fl, Planck's constant; bosons have Searching for a Few Good Events integer values. The elementary bosons are the carriers of the forces: the photon is the "carrier" of the electromagnetic Each collision of a proton in one beam with an antiproton force, the W and Z bosons are carriers ofthe weak force, and (Fermilab) or proton (SSC) in the other counter-rotating the gluon is the carrier of the strong force. beam converts the collision energy into particle production, The elementary fermions are themselves divided into two with typically 50-100 charged particles and 25-50 neutral classes with very different properties. One class is the particles produced initially. What kinds of particles (pions, leptons, consisting of the electron, the muon, the tau lepton, kaons, protons, charmed particles, top, electrons, neutrinos, each with its associated neutral partner neutrino. The second etc.) are produced, and what their distributions in angle and is the quarks, consisting also (we believe) of three pairs: the momentum are, vary from collision to collision with proba- up and down quarks, the charm and strange quarks, and the bilities determined by the underlying dynamics (forces) and top and bottom quarks. Two differences between quarks and kinematics (momentum and energy conservation, impact leptons are that leptons do not interact via the strong inter- parameter, etc.). Each collision thus has an equal a priori action while quarks do, and leptons have integral values of probability of producing something interesting; we therefore electric charge (0 or 1) while quarks are fractionally charged need to inspect as many as possible per unit time. (2/3 or -1/3!). Leptons exist freely in nature, but quarks Each collision is recorded as an electronic "snapshot" by seem to be required to be confined inside strongly interacting a custom-built detector. These detectors have now evolved to be mammoth objects, with many individual recording chan- The publication costs of this article were defrayed in part by page charge payment. This article must therefore be hereby marked "advertisement" Abbreviations: SSC, Superconducting Supercollider; CDF, Collider in accordance with 18 U.S.C. §1734 solely to indicate this fact. Detector at Fermilab. 9754 Downloaded by guest on October 2, 2021 Colloquium Paper: Frisch Proc. Natl. Acad. Sci. USA 90 (1993) 9755 nels that measure such quantities as the positions ofparticles Each time the bunches cross there is about a 20% chance that in space as they traverse the detector, or energies deposited a proton in the proton bunch hits an antiproton in the in calorimeters by particles that enter and interact in them. counter-rotating antiproton bunch at present intensities. At Fig. 1 shows the Collider Detector at Fermilab (CDF). The the SSC the bunches both will contain protons, and are only detector has approximately 150,000 channels of high-gain 16 nsec apart, with typically 1.6 collisions per bunch-crossing amplification and data recording. It stands 30 ft tall and and a total of 108 proton-proton collisions per second. weighs many thousands of tons. A collision recorded by the Each collision produces particles which leave tracks or detector is known as an "event." It is the images of billions deposit energy in the detector. The detector has to register ofevents per second, reconstructed from the electronic data, and store this information while a decision is made if the that we have to inspect to see ifwe have found anything new, event is of sufficient interest to keep; at this stage in the expected or unexpected. Fig. 2 shows a typical garden- selection only one event in 105 or so can be kept. If an event variety event, as reconstructed into an image by computer is kept several milliseconds are spent digitizing all of the from a fraction of the data in the event. Many different event information and transferring the information from the pictures can be made using different kinds ofdata supplied by detector elements into buffers in a farm of computers con- the detector; inferring what was the underlying physics (top quark? Higgs boson?) from the data is the art and science of nected to the data acquisition system. the analysis. The technical problem is that the detector must be ready for the next crossing in general; otherwise one is wasting Description of the Technical Problem beam time and sensitivity to rare events by being "dead" while dealing with an event one most likely will not keep. The The beams of protons (antiprotons) are bunched so that the situation is analogous to a camera: we take a snapshot of two beams collide at fixed places around the ring. At the every crossing with the detector, and have to be ready to take Tevatron at Fermilab, for example, there are six bunches of the next picture 3.5 usec (Fermilab) or 16 nsec (SSC) later. protons going clockwise (looking down) in the accelerator The "picture" generated by the detector has 105 separate and six bunches of antiprotons going counterclockwise. The pieces of information at Fermilab and about 106 at the SSC. bunches collide in six places in the ring, and those places are To summarize the technical problem, taking Fermilab as where one locates detectors (actually at Fermilab only three the example: of the six are used for detectors; the other three are used for (i) There is a new snapshot captured 286,000 times per accelerator-related purposes). second. At Fermilab the bunches are 3.5 ,usec apart, and there are (ii) Each snapshot has approximately 100,000 individual typically 105 to 106 proton-antiproton collisions per second. measurements of either energy or spatial information.

FIG. 1. CDF particle detector (at Fermilab). Downloaded by guest on October 2, 2021 9756 Colloquium Paper: Frisch Proc. Natl. Acad. Sci. USA 90 (1993) of selection decisions, the initial ones being very fast but rather coarse, with more and more information being used in the later decisions. At Fermilab the very first decision, which rejects the largest fraction ofevents as not interesting, is done in the 3.5 ,sec between beam crossings. This decision, called Level 1, is done by asking for an indication only of a collision that has produced a local pattern ofinterest-either a modest amount ofenergy deposited in a single element ofthe detector or a crude indication that a muon has penetrated to the outside layers of the detector. The next higher decision process, Level 2, uses custom hardware to find patterns of energy in the calorimeter characteristic ofjets of particles or of electrons and uses custom track-finding pattern recogni- B tion to find tracks of particles in the tracking chambers. The tracks are linked to both the calorimeter clusters and to the muon chambers. Finally, a list ofjets, electrons, and muons is presented to a very fast custom processor that makes the Level 2 decision based on comparing the list of what is found to a predetermined list of patterns that are considered inter- esting. For example, the decays of a pair of top quarks can result in two very energetic charged leptons. Any event in which two charged leptons (electrons, muons, or taus) with high momentum are found is kept. To summarize the strategy: (i) Make a fast decision based on a subset of all the information. This information is at coarser granularity and comes directly along dedicated cables to the trigger proces- sors. (ii) The trigger decision is made in successive levels, with the first level being done (at CDF) between beam crossings so that no sensitivity to collisions is wasted. Each level uses more information, takes longer, and is more sophisticated in terms of pattern recognition. (iii) If the event is rejected as uninteresting by the trigger the detector is reset to be ready for the next crossing. (iv) The rest of the information is stored by sample-and- FIG. 2. A typical event from the CDF detector in which two hold circuits on the detector while the fast decision is being "jets" can be seen. (A) The "Lego" plot, in which the cylindrical made. On CDF the storage is analog; at the SSC the decision detector has been "unrolled" into a plane in which the energy is still to be made if it's to be analog or digital. deposited in the detector is displayed. (B) The tracking chamber as (v) The trigger uses known local signatures (patterns) of seen in a beam's-eye view. The chamber is imbedded in a magnetic electrons, muons, and jets as primitives in selecting events. field which causes the charged particle tracks to form arcs of circles, (vi) The Level 2 decision, which takes about 20 ,usec, is allowing momentum measurement. based on lists of such primitives. It uses programable custom and fast technical detail-it uses a (iii) The energy measurements require 16 bits of dynamic processors memory (one range and about 10-bit precision (linearity is the issue); the custom bus that is 2048 wide for speed). spatial information is contained in fewer bits, typically 10-12. (vii) If the coarser information in the trigger satisfies the (iv) The sources of data are diverse, with differing char- trigger criteria, the full data are then digitized to full precision acteristics in response, settling times, signal shape, etc. and are read out into a farm of commercial processors. At (some channels are photomultipliers, some are proportional CDF this farm has a processing power of about 1100 MIPS chambers, some are inherently digital, some are analog, etc.). (million instructions per second). Each of these computers (v) The data sources are spread out over a big physical area acts on one event at a time with the same analysis programs (many tens of meters), and are 200 ft from a decent room in that are used off-line to make a more detailed filtering. which one can work. (This implies possible ground loops, (viii) The filtering pipeline for CDF starts with 105 to 106 timing problems, signal settling time problems, etc.). events per second input to Level 1, which rejects all but about (vi) We can record only a few events per second perma- 5000 per second, which are input to Level 2. Level 2 has to nently (implying a rejection factor of approximately 105 to reduce the rate to less than 25 per second, which are then fully 106). digitized and read out into the computer farm, which is called (vii) Interesting events are rare: for example, we expect Level 3. Level 3 then analyzes the events and rejects all but fewer than 100 top quarks detected per year, hidden among 4-6 per second, which are written out to 8-mm tape. The 1012 non-top events. most interesting events are skimmed off at this time and At the SSC, the big difference from the above is that there written directly to disk. is only 16 nsec between beam crossings. The rejection factor required in real time is in fact not much different, as we Some Comments expect to be able to record about 100 events per second permanently. The strategy has some key underlying aspects: (i) We look for (at least some) unknown processes by Solutions to the Technical Problems searching for rare combinations of known local signatures- e.g., electron showers. (This is the weak spot philosophi- The solution we have evolved with the CDF detector, and cally, in that truly bizarre new phenomena could in principle which we will propagate to the SSC, is to make a succession be missed.) Downloaded by guest on October 2, 2021 Colloquium Paper: Frisch Proc. Natl. Acad. Sci. USA 90 (1993) 9757 (ii) We go directly to small data sets of candidate events. being about 250 kbytes, and to select between 2 and 6 events This emphasis on data-set-driven analysis is a very important per second to save for future analysis. The strategy depends ingredient in getting results out quickly. on looking at a coarser subset of the data for local signatures (iii) We try, albeit painfully, to get the entire collaboration ofelectrons, muons, andjets, whose patterns we understand, (372 physicists) to agree in advance on the selection criteria and forming lists of these as primitives. From the list we (be bold). (iv) We monitor the process intensively in real time (be select topologies and patterns that are possibly indicative of careful.'). interesting physics. The structure is a multilevel one, with (v) We overlap criteria to play against each other in order each level taking longer than the last, but with fewer events to measure efficiencies for the individual local signatures. input to that level. Real-time monitoring and feedback is (vi) We prescale certain selections (i.e., take 1 of N) with crucial to the strategy. Quick access to the data by a large some ofthe criteria removed, again to measure the efficiency number of people is thus critical. Finally, we have extended of those criteria. this strategy to the SSC, where the event rate is higher by a (vii) The whole process requires rapid access to the data by factor of 100 to 1000, and the events are expected to be larger the whole collaboration to monitor and correct the process as by approximately a factor of 4 on average. it continues. Summary 1. Gottfried, K. & Weisskopf, V. (1983) Concepts of Particle Physics (Clarendon, Oxford). We have developed a strategy and the custom hardware to 2. Griffiths, D. (1987) Introduction to Elementary Particles (Wi- look at 105 to 106 events per second at Fermilab, each event ley, New York). Downloaded by guest on October 2, 2021