How do we know what we see ?


Particle accelerators create collisions among high-energy particles in order to produce new particles. A large variety of particles may be produced in each collision, but only a few are of special interest. How do we recognize these? In the following, I'll explain the basics of particle detection and event selection.

Detecting particles

Typical modern multipurpose detectors surround the interaction point - where the accelerated beams collide - in successive layers of subdetectors.
Fig 1. View of the CMS detector, one of the two general-purpose detectors being constructed for the LHC accelerator.

Closest to the beam are tracking devices, which detect the passing of charged particles without severely modifying their course. It is essential that the amount of material in these detectors is small, so typically they use a gas mixture or thin layers of silicon. When a charged particle passes through the material, it will kick out electrons from the atoms. A high voltage applied to the material collects these free electrons, creating a small pulse which is amplified, digitized and ultimately read out by computers. A single particle passes through many layers of detection, leaving a track consisting of many points, for each of the many particles passing through the detector. To distinguish between tracks, a magnetic field is applied to the whole tracker, so that charged particles will curve in flight. While this complicates the reconstruction, eventually the charge and momentum of the particles can be determined from the bending of their tracks.

Fig 2. Simulated charged particle "tracks" as seen in the ATLAS detector.

After passing through the tracker, particles enter the second major type of subdetectors, the calorimeters. These are solid blocks or layers of heavy material that absorb the energy of the incoming particle and generate an electronic signal proportional to its total energy. There are two main types of calorimeters, electromagnetic and hadronic. An electromagnetic calorimeter might consist of blocks of some crystal, like lead glass, the same material that might be used to make a vase, but other materials can also be used. Incoming electrons and photons, that is electromagnetic particles, interact with the electrons of the crystal, producing more electrons, antielectrons and photons, leading to a shower or cascade of particles. In the end we are left with many photons that do not have enough energy to create new particles, but can be observed as a small pulse of light. This can be converted into an electric pulse and sent to computers after digitization. 

Fig 3. Simulated electromagnetic shower in the liquid Argon calorimeter of the Atlas detector at LHC.

Hadrons are more penetrating and can pass through the electromagnetic calorimeters with little loss of energy, since they don't interact so easily with electrons, they prefer the atomic nuclei. hadronic calorimeters therefore typically consist of some heavy material, like lead or steel. A hadron hitting the nucleus of an atom produces secondary particles, which in turn hit other nuclei, producing more particles, until all the incoming energy is spent. This shower is usually sampled by layers of scintillator or gas detectors that produce a small pulse proportional to the number of particles passing through.

Fig. 4. The hadronic calorimeter of CMS, a contribution from Fermilab.

Fig 5. Reconstructed event with an electron (red), two jets of particles (blue) and a neutrino (dashed) seen by the OPAL detector at LEP.

There is only one type of charged particles that can pass through all the material in the calorimeters, muons. They are very much like electrons, only heavier and much more penetrating. Their tracks are measured again in the muon chambers located outside the hadronic calorimeters, adding more points at a large distance, thereby improving the precision of the measurement. Because so few particles reach the muon chambers, the ones that do can be easily identified as muons, providing a very strong identification. When you look at a typical detector, what you see is actually the outside of the muon chambers. The only long-lived particles not seen in the detector are the neutrinos, neutral, weakly interacting particles. Their interactions are in fact so weak, that they easily pass through the Earth without notice, and they can only be detected by specialized detectors. But we still can recognize their presence indirectly, by exploiting the conservation of energy and momentum. Since the incoming beams are symmetric, the momentum of any particles going one way must be balanced by other particles going the other way. So if we find an imbalance, we know that the missing momentum was carried away by neutrinos.

Selecting interesting events

With the advance of technology particle accelerators get more and more powerful, and detectors get bigger and bigger with ever increasing resolution. This naturally leads to a rise in the amount of raw data produced, so the latest technology must also be used to read out and interpret this data, leading to new solutions being introduced at the last minute and upgrades of working systems to improve performance or delay costs. Below I describe a typical solution with a few numbers to give you a feeling of the scale of the problem, but remember that this is just an example. Particles in the accelerator collide thousands or millions of times each second, but most of the collisions are uninteresting. (Remember, we want to discover something new!) How to select the few events that we deem important? This is started by the trigger subsystem. For efficiency, it is always divided in two or more levels. At the first level trigger, in a simplified view, imagine that each time a subdetector sees something that might be part of an interesting event, it raises a flag. The central trigger processor looks at these flags, and when it sees a predefined pattern, like 'two high-momentum muons and two high-momentum electrons', it knows the event might be interesting and instructs all the subdetectors to send their data on, otherwise after a timeout each subdetector throws away the data. With millions of collisions each second, the first level trigger can only keep in the order of one in a hundred.

Fig 6. The high-level trigger farm of the CDF detector at the Tevatron (under construction).

While each subdetector sees only a small fragment of each event, we need to look at each event as a whole. It is the task of the event builder to combine the fragments into entire events. Since no single computer could cope with the data rate of the entire detector, we use many computers, each dealing with a small fraction of the events, but collecting data from all the parts of the detector. Think of this as a big network switch: the fragments of an event come from all over the detector and are sent to the one computer collecting this particular event, the next event is sent to the next available computer and so on. The assembled data from the event builder is passed on to the filter farm, where computers look at the event as a whole and apply the high-level trigger selection criteria to decide whether this event is worth storing. These criteria are established based on computer simulations of the 'interesting' processes and 'uninteresting' background events that might look very similar. At this stage a mini-analysis of the event may be performed to identify the physics process observed, but there is not much time because of the many incoming events waiting. Selected events are written to permanent storage for detailed analysis.

Reconstructing what happened

Raw data is obtained from the detector and then stored on tape at the level of individual channels, that is the position of all the hits in the tracker or the measured energy in each individual calorimeter crystal. There is no connection among the many hits and deposits caused by the same particle. To make sense of it, one needs to get back the original particles that made these hits and then find out what the original parameters of those particles was. This process is called the reconstruction, and it involves many iterations of pattern matching and function fitting, to get a list of particles that were observed in the detector. The reconstruction uses calibration constants and other parameters of the detector to get the best estimate of the original particles. However, these parameters can be later improved, often using the data itself to better understand the secondary effects in the detector. As the data are reprocessed with the improved parameters, their quality improves, they 'mature'.

Extracting results

Once the data has been fully processed, detected particles have been reconstructed, and the 'real' data analysis can start. Physicists, including graduate students, try to answer specific questions, like Is there a Higgs particle with a mass of x GeV? or What are the properties of the top quark?  The general method is to use computer simulation to find out what would be the effect of a given assumption on our data, then try to isolate that effect by selecting events that are similar to the signal we are looking for. The remaining fraction of background events can be estimated from the simulation of all other processes that might fake the signal, and these must be taken into account when estimating the error of the measurement. Once we have selected a sample of events that mostly contain the particle we are looking for, we can study its properties, or if we can not select such a sample, we can derive a negative result that excludes the given particle with a given probability. For example you might generate simulated samples of Higgs bosons assuming various masses, compare these signal samples to all the background samples that only contain the standard processes, and define a set of requirements or cuts that you can apply to the data to select Higgs-like events. From theoretical calculations you know how many Higgs particles you expect to be produced and from the signal simulation you know how many of these you would see in your selected sample. From the background simulation you know how many events you would find that look like Higgs but are not. Comparing these numbers you calculate a probability, and you either say that there is a 99.9% chance that the experiment saw Higgs particles, in this case you can calculate their mass and estimate the error of the measurement, or you say that there is 95% chance that no Higgs particles were produced, therefore the Higgs particles must be heavier than a given mass.

For more information

For a gentle introduction to particle physics in general, including detectors, try

Akos Csilling
Research Fellow at CERN.
On leave form KFKI Research Institute for Particle and Nuclear Physics
Opal, L3, CMS experiments