A symphony of data

Article title: “MUSiC: a model unspecific search for new physics in
proton-proton collisions at \sqrt{s} = 13 TeV”

Authors: The CMS Collaboration

Reference: https://arxiv.org/abs/2010.02984

First of all, let us take care of the spoilers: no new particles or phenomena have been found… Having taken this concern away, let us focus on the important concept behind MUSiC.

ATLAS and CMS, the two largest experiments using collisions at the LHC, are known as “general purpose experiments” for a good reason. They were built to look at a wide variety of physical processes and, up to now, each has checked dozens of proposed theoretical extensions of the Standard Model, in addition to checking the Model itself. However, in almost all cases their searches rely on definite theory predictions and focus on very specific combinations of particles and their kinematic properties. In this way, the experiments may still be far from utilizing their full potential. But now an algorithm named MUSiC is here to help.

MUSiC takes all events recorded by CMS that comprise of clean-cut particles and compares them against the expectations from the Standard Model, untethering itself from narrow definitions for the search conditions.

We should clarify here that an “event” is the result of an individual proton-proton collision (among the many happening each time the proton bunches cross), consisting of a bouquet of particles. First of all, MUSiC needs to work with events with particles that are well-recognized by the experiment’s detectors, to cut down on uncertainty. It must also use particles that are well-modeled, because it will rely on the comparison of data to simulation and, so, wants to be sure about the accuracy of the latter.

Display of an event with two muons at CMS. (Source: CMS experiment)

All this boils down to working with events with combinations of specific, but several, particles: electrons, muons, photons, hadronic jets from light-flavour (=up, down, strange) quarks or gluons and from bottom quarks, and deficits in the total transverse momentum (typically the signature of the uncatchable neutrinos or perhaps of unknown exotic particles). And to make things even more clean-cut, it keeps only events that include either an electron or a muon, both being well-understood characters.

These particles’ combinations result in hundreds of different “final states” caught by the detectors. However, they all correspond to only a dozen combos of particles created in the collisions according to the Standard Model, before some of them decay to lighter ones. For them, we know and simulate pretty well what we expect the experiment to measure.

MUSiC proceeded by comparing three kinematic quantities of these final states, as measured by CMS during the year 2016, to their simulated values. The three quantities of interest are the combined mass, combined transverse momentum and combined missing transverse momentum. It’s in their distributions that new particles would most probably show up, regardless of which theoretical model they follow. The range of values covered is pretty wide. All in all, the method extends the kinematic reach of usual searches, as it also does with the collection of final states.

An example distribution from MUSiC: Transverse mass for the final state comprising of one muon and missing transverse momentum. Color histograms: Simulated Standard Model processes. Red line: Signal from a hypothetical W’ boson with mass of 3TeV. (Source: paper)

So the kinematic distributions are checked against the simulated expectations in an automatized way, with MUSiC looking for every physicist’s dream: deviations. Any deviation from the simulation, meaning either fewer or more recorded events, is quantified by getting a probability value. This probability is calculated by also taking into account the much dreaded “look elsewhere effect”. (Which comes from the fact that, statistically, in a large number of distributions a random fluctuation that will mimic a genuine deviation is bound to appear sooner or later.)

When all’s said and done the collection of probabilities is overviewed. The MUSiC protocol says that any significant deviation will be scrutinized with more traditional methods – only that this need never actually arose in the 2016 data: all the data played along with the Standard Model, in all 1,069 examined final states and their kinematic ranges.

For the record, the largest deviation was spotted in the final state comprising three electrons, two generic hadronic jets and one jet coming from a bottom quark. Seven events were counted whereas the simulation gave 2.7±1.8 events (mostly coming from the production of a top plus an anti-top quark plus an intermediate vector boson from the collision; the fractional values are due to extrapolating to the amount of collected data). This excess was not seen in other related final states, “related” in that they also either include the same particles or have one less. Everything pointed to a fluctuation and the case was closed.

However, the goal of MUSiC was not strictly to find something new, but rather to demonstrate a method for model un-specific searches with collisions data. The mission seems to be accomplished, with CMS becoming even more general-purpose.

Read more:

Another generic search method in ATLAS: Going Rogue: The Search for Anything (and Everything) with ATLAS

And a take with machine learning: Letting the Machines Seach for New Physics

Fancy checking a good old model-specific search? Uncovering a Higgs Hiding Behind Backgrounds

Machine Learning The LHC ABC’s

Article Title: ABCDisCo: Automating the ABCD Method with Machine Learning

Authors: Gregor Kasieczka, Benjamin Nachman, Matthew D. Schwartz, David Shih

Reference: arxiv:2007.14400

When LHC experiments try to look for the signatures of new particles in their data they always apply a series of selection criteria to the recorded collisions. The selections pick out events that look similar to the sought after signal. Often they then compare the observed number of events passing these criteria to the number they would expect to be there from ‘background’ processes. If they see many more events in real data than the predicted background that is evidence of the sought after signal. Crucial to whole endeavor is being able to accurately estimate the number of events background processes would produce. Underestimate it and you may incorrectly claim evidence of a signal, overestimate it and you may miss the chance to find a highly sought after signal.

However it is not always so easy to estimate the expected number of background events. While LHC experiments do have high quality simulations of the Standard Model processes that produce these backgrounds they aren’t perfect. Particularly processes involving the strong force (aka Quantum Chromodynamics, QCD) are very difficult to simulate, and refining these simulations is an active area of research. Because of these deficiencies we don’t always trust background estimates based solely on these simulations, especially when applying very specific selection criteria.

Therefore experiments often employ ‘data-driven’ methods where they estimate the amount background events by using control regions in the data. One of the most widely used techniques is called the ABCD method.

An illustration of the ABCD method. The signal region, A, is defined as the region in which f and g are greater than some value. The amount of background in region A is estimated using regions B C and D which are dominated by background.

The ABCD method can applied if the selection of signal-like events involves two independent variables f and g. If one defines the ‘signal region’, A,  (the part of the data in which we are looking for a signal) as having f  and g each greater than some amount, then one can use the neighboring regions B, C, and D to estimate the amount of background in region A. If the number of signal events outside region A is small, the number of background events in region A can be estimated as N_A = N_B * (N_C/N_D).

In modern analyses often one of these selection requirements involves the score of a neural network trained to identify the sought after signal. Because neural networks are powerful learners one often has to be careful that they don’t accidentally learn about the other variable that will be used in the ABCD method, such as the mass of the signal particle. If two variables become correlated, a background estimate with the ABCD method will not be possible. This often means augmenting the neural network either during training or after the fact so that it is intentionally ‘de-correlated’ with respect to the other variable. While there are several known techniques to do this, it is still a tricky process and often good background estimates come with a trade off of reduced classification performance.

In this latest work the authors devise a way to have the neural networks help with the background estimate rather than hindering it. The idea is rather than training a single network to classify signal-like events, they simultaneously train two networks both trying to identify the signal. But during this training they use a groovy technique called ‘DisCo’ (short for Distance Correlation) to ensure that these two networks output is independent from each other. This forces the networks to learn to use independent information to identify the signal. This then allows these networks to be used in an ABCD background estimate quite easily.

The authors try out this new technique, dubbed ‘Double DisCo’, on several examples. They demonstrate they are able to have quality background estimates using the ABCD method while achieving great classification performance. They show that this method improves upon the previous state of the art technique of decorrelating a single network from a fixed variable like mass and using cuts on the mass and classifier to define the ABCD regions (called ‘Single Disco’ here).

Using the task of identifying jets containing boosted top quarks, they compare the classification performance (x-axis) and quality of the ABCD background estimate (y-axis) achievable with the new Double DisCo technique (yellow points) and previously state of the art Single DisCo (blue points). One can see the Double DisCo method is able to achieve higher background rejection with a similar or better amount of ABCD closure.

While there have been many papers over the last few years about applying neural networks to classification tasks in high energy physics, not many have thought about how to use them to improve background estimates as well. Because of their importance, background estimates are often the most time consuming part of a search for new physics. So this technique is both interesting and immediately practical to searches done with LHC data. Hopefully it will be put to use in the near future!

Further Reading:

Quanta Magazine Article “How Artificial Intelligence Can Supercharge the Search for New Particles

Recent ATLAS Summary on New Machine Learning Techniques “Machine learning qualitatively changes the search for new particles

CERN Tutorial on “Background Estimation with the ABCD Method

Summary of Paper of Previous Decorrelation Techniques used in ATLAS “Performance of mass-decorrelated jet substructure observables for hadronic two-body decay tagging in ATLAS

A shortcut to truth

Article title: “Automated detector simulation and reconstruction
parametrization using machine learning”

Authors: D. Benjamin, S.V. Chekanov, W. Hopkins, Y. Li, J.R. Love

Reference: https://arxiv.org/abs/2002.11516 (https://iopscience.iop.org/article/10.1088/1748-0221/15/05/P05025)

Demonstration of probability density function as the output of a neural network. (Source: paper)

The simulation of particle collisions at the LHC is a pharaonic task. The messy chromodynamics of protons must be modeled; the statistics of the collision products must reflect the Standard Model; each particle has to travel through the detectors and interact with all the elements in its path. Its presence will eventually be reduced to electronic measurements, which, after all, is all we know about it.

The work of the simulation ends somewhere here, and that of the reconstruction starts; namely to go from electronic signals to particles. Reconstruction is a process common to simulation and to the real world. Starting from the tangle of statistical and detector effects that the actual measurements include, the goal is to divine the properties of the initial collision products.

Now, researchers at the Argonne National Laboratory looked into going from the simulated particles as produced in the collisions (aka “truth objects”) directly to the reconstructed ones (aka “reco objects”): bypassing the steps of the detailed interaction with the detectors and of the reconstruction algorithm could make the studies that use simulations much more speedy and efficient.

Display of a collision event involving hadronic jets at ATLAS. Each colored block corresponds to interaction with a detector element. (Source: ATLAS experiment)

The team used a neural network which it trained on simulations of the full set. The goal was to have the network learn to produce the properties of the reco objects when given only the truth objects. The process succeeded in producing the transverse momenta of hadronic jets, and looks suitable for any kind of particle and for other kinematic quantities.

More specifically, the researchers began with two million simulated jet events, fully passed through the ATLAS experiment and the reconstruction algorithm. For each of them, the network took the kinematic properties of the truth jet as input and was trained to achieve the reconstructed transverse momentum.

The network was taught to perform multi-categorization: its output didn’t consist of a single node giving the momentum value, but of 400 nodes, each corresponding to a different range of values. The output of each node was the probability for that particular range. In other words, the result was a probability density function for the reconstructed momentum of a given jet.

The final step was to select the momentum randomly from this distribution. For half a million of test jets, all this resulted in good agreement with the actual reconstructed momenta, specifically within 5% for values above 20 GeV. In addition, it seems that the training was sensitive to the effects of quantities other than the target one (e.g. the effects of the position in the detector), as the neural network was able to pick up on the dependencies between the input variables. Also, hadronic jets are complicated animals, so it is expected that the method will work on other objects just as well.

Comparison of the reconstructed transverse momentum between the full simulation and reconstruction (“Delphes”) and the neural net output. (Source: paper)

All in all, this work showed the perspective for neural networks to imitate successfully the effects of the detector and the reconstruction. Simulations in large experiments typically take up loads of time and resources due to their size, intricacy and frequent need for updates in the hardware conditions. Such a shortcut, needing only small numbers of fully processed events, would speed up studies such as optimization of the reconstruction and detector upgrades.

More reading:

Argonne Lab press release: https://www.anl.gov/article/learning-more-about-particle-collisions-with-machine-learning

Intro to neural networks: https://physicsworld.com/a/neural-networks-explained/

Crystals are dark matter’s best friends

Article title: “Development of ultra-pure NaI(Tl) detector for COSINE-200 experiment”

Authors: B.J. Park et el.

Reference: arxiv:2004.06287

The landscape of direct detection of dark matter is a perplexing one; all experiments have so far come up with deafening silence, except for a single one which promises a symphony. This is the DAMA/LIBRA experiment in Gran Sasso, Italy, which has been seeing an annual modulation in its signal for two decades now.

Such an annual modulation is as dark-matter-like as it gets. First proposed by Katherine Freese in 1987, it would be the result of earth’s motion inside the galactic halo of dark matter in the same direction as the sun for half of the year and in the opposite direction during the other half. However, DAMA/LIBRA’s results are in conflict with other experiments – but with the catch that none of those used the same setup. The way to settle this is obviously to build more experiments with the DAMA/LIBRA setup. This is an ongoing effort which ultimately focuses on the crystals at its heart.

Cylindrical crystals wrapped in reflector, bounded by photomultipliers (PMTs) and surrounded by scintillators. (COSINE-100)

The specific crystals are made of the scintillating material thallium-doped sodium iodide, NaI(Tl). Dark matter particles, and particularly WIMPs, would collide elastically with atomic nuclei and the recoil would give off photons, which would eventually be captured by photomultiplier tubes at the ends of each crystal.

Right now a number of NaI(Tl)-based experiments are at various stages of preparation around the world, with COSINE-100 at the Yangyang mountain, S.Korea, already producing negative results. However, these are still not on equal footing with DAMA/LIBRA’s because of higher backgrounds at COSINE-100. What is the collaboration to do, then? The answer is focus even more on the crystals and how they are prepared.

Setup of the COSINE-100 experiment. (COSINE-100)

Over the last couple of years some serious R&D went into growing better crystals for COSINE-200, the planned upgrade of COSINE-100. Yes, a crystal is something that can and does grow. A seed placed inside the raw material, in this case NaI(Tl) powder, leads it to organize itself around the seed’s structure over the next hours or days.

In COSINE-100 the most annoying backgrounds came from within the crystals themselves because of the production process, because of natural radioactivity, and because of cosmogenically induced isotopes. Let’s see how each of these was tackled during the experiment’s mission towards a radiopure upgrade.

Improved techniques of growing and preparing the crystals reduced contamination from the materials of the grower device and from the ambient environment. At the same time different raw materials were tried out to put the inherent contamination under control.

Among a handful of naturally present radioactive isotopes particular care was given to 40K. 40K can decay characteristically to an X-ray of 3.2keV and a γ-ray of 1,460keV, a combination convenient for tagging it to a large extent. The tagging is done with the help of 2,000 liters of liquid scintillator surrounding the crystals. However, if the γ-ray escapes the crystal then the left-behind X-ray will mimic the expected signal from WIMPs… Eventually the dangerous 40K was brought down to levels comparable to those in DAMA/LIBRA through the investigation of various techniques and first materials.

But the main source of radioactive background in COSINE-100 was isotopes such as 3H or 22Na created inside the crystals by cosmic ray muons, after their production. Now, their abundance was reduced significantly by two simple moves: the crystals were grown locally at a very low altitude and installed underground within a few weeks (instead of being transported from a lab at 1,400 meters above sea in Colorado). Moreover, most of the remaining cosmogenic background is to decay away within a couple of years.

Components of the background, and temporal evolution of the cosmogenic radioactivity. (Source)

Where are these efforts standing? The energy range of interest for testing the DAMA/LIBRA signal is 1-6keV. This corresponds to a background target of 1 count/kg/day/keV. After the crystals R&D, the achieved contamination was less than about 0.34 counts. In short, everything is ready for COSINE-100 to upgrade to COSINE-200 and test the annual modulation without the previous ambiguities that stood in the way.

Learn more:

More on DAMA/LIBRA in ParticleBites.

Cross-checking the modulation.

The COSINE-100 experiment.

First COSINE-100 results.

Listening for axions

If dark matter actually consists of a new kind of particle, then the most up-and-coming candidate is the axion. The axion is a consequence of the Peccei-Quinn mechanism, a plausible solution to the “strong CP problem,” or why the strong nuclear force conserves the CP-symmetry although there are no reasons for it to. It is a very light neutral boson, named by Frank Wilczek after a detergent brand (in a move that obviously dates its introduction in the ’70s).

Axion decay in a magnetic field: the result is a photon. (Source.)

Most experiments that try to directly detect dark matter have looked for WIMPs (weakly interacting massive particles). However, as those searches have not borne fruit, the focus started turning to axions, which make for good candidates given their properties and the fact that if they exist, then they exist in multitudes throughout the galaxies. Axions “speak” to the QCD part of the Standard Model, so they can appear in interaction vertices with hadronic loops. The end result is that axions passing through a magnetic field will convert to photons.

In practical terms, their detection boils down to having strong magnets, sensitive electronics and an electromagnetically very quiet place at one’s disposal. One can then sit back and wait for the hypothesized axions to pass through the detector as earth moves through the dark matter halo surrounding the Milky Way. Which is precisely why such experiments are known as “haloscopes.”

Now, the most veteran haloscope of all published significant new results. Alas, it is still empty-handed, but we can look at why its update is important and how it was reached.

ADMX (Axion Dark Matter eXperiment) of the University of Washington has been around for a quarter-century. By listening for signals from axions, it progressively gnaws away at the space of allowed values for their mass and coupling to photons, focusing on an area of interest:

Latest exclusion limits on the axion mass and coupling to photons.

Unlike higher values, this area is not excluded by astrophysical considerations (e.g. stars cooling off through axion emission) and other types of experiments (such as looking for axions from the sun). In addition, the bands above the lines denoted “KSVZ” and “DFSZ” are special. They correspond to the predictions of two models with favorable theoretical properties. So, ADMX is dedicated to scanning this parameter space. And the new analysis added one more year of data-taking, making a significant dent in this ballpark.

As mentioned, the presence of axions would be inferred from a stream of photons in the detector. The excluded mass range was scanned by “tuning” the experiment to different frequencies, while at each frequency step longer observation times probed smaller values for the axion-photon coupling.

Two things that this search needs is a lot of quiet and some good amplification, as the signal from a typical axion is expected to be as weak as the signal from a mobile phone left on the surface of Mars (around 10-23W). The setup is indeed stripped of noise by being placed in a dilution refrigerator, which keeps its temperature at a few tenths of a degree above absolute zero. This is practically the domain governed by quantum noise, so advantage can be taken of the finesse of quantum technology: for the first time ADMX used SQUIDs, superconducting quantum interference devices, for the amplification of the signal.

The heart of the experiment inside the refrigerator. The resonant frequency of the cavity is tuned to match the photons -hopefully- given off by axions. (Source.)

In the end, a good chunk of the parameter space which is favored by the theory might have been excluded, but the haloscope is ready to look at the rest of it. Just think of how, one day, a pulse inside a small device in a university lab might be a messenger of the mysteries unfolding across the cosmos.


Publication by the ADMX collaboration. (arXiv)

Learn more:

  1. The theory behind axions.
  2. The hitchhiker’s guide to the dilution refrigerator.
  3. Intro to KSVZ and DFSZ axions (and more).
  4. Resonant cavities.

Making Smarter Snap Judgments at the LHC

Collisions at the Large Hadron Collider happen fast. 40 million times a second, bunches of 1011 protons are smashed together. The rate of these collisions is so fast that the computing infrastructure of the experiments can’t keep up with all of them. We are not able to read out and store the result of every collision that happens, so we have to ‘throw out’ nearly all of them. Luckily most of these collisions are not very interesting anyways. Most of them are low energy interactions of quarks and gluons via the strong force that have been already been studied at previous colliders. In fact, the interesting processes, like ones that create a Higgs boson, can happen billions of times less often than the uninteresting ones.

The LHC experiments are thus faced with a very interesting challenge, how do you decide extremely quickly whether an event is interesting and worth keeping or not? This what the ‘trigger’ system, the Marie Kondo of LHC experiments, are designed to do. CMS for example has a two-tiered trigger system. The first level has 4 microseconds to make a decision and must reduce the event rate from 40 millions events per second to 100,000. This speed requirement means the decision has to be made using at the hardware level, requiring the use of specialized electronics to quickly to synthesize the raw information from the detector into a rough idea of what happened in the event. Selected events are then passed to the High Level Trigger (HLT), which has 150 milliseconds to run versions of the CMS reconstruction algorithms to further reduce the event rate to a thousand per second.

While this system works very well for most uses of the data, like measuring the decay of Higgs bosons, sometimes it can be a significant obstacle. If you want to look through the data for evidence of a new particle that is relatively light, it can be difficult to prevent the trigger from throwing out possible signal events. This is because one of the most basic criteria the trigger uses to select ‘interesting’ events is that they leave a significant amount of energy in the detector. But the decay products of a new particle that is relatively light won’t have a substantial amount of energy and thus may look ‘uninteresting’ to the trigger.

In order to get the most out of their collisions, experimenters are thinking hard about these problems and devising new ways to look for signals the triggers might be missing. One idea is to save additional events from the HLT in a substantially reduced size. Rather than saving the raw information from the event, that can be fully processed at a later time, instead the only the output of the quick reconstruction done by the trigger is saved. At the cost of some precision, this can reduce the size of each event by roughly two orders of magnitude, allowing events with significantly lower energy to be stored. CMS and ATLAS have used this technique to look for new particles decaying to two jets and LHCb has used it to look for dark photons. The use of these fast reconstruction techniques allows them to search for, and rule out the existence of, particles with much lower masses than otherwise possible. As experiments explore new computing infrastructures (like GPU’s) to speed up their high level triggers, they may try to do even more sophisticated analyses using these techniques. 

But experimenters aren’t just satisfied with getting more out of their high level triggers, they want to revamp the low-level ones as well. In order to get these hardware-level triggers to make smarter decisions, experimenters are trying get them to run machine learning models. Machine learning has become very popular tool to look for rare signals in LHC data. One of the advantages of machine learning models is that once they have been trained, they can make complex inferences in a very short amount of time. Perfect for a trigger! Now a group of experimentalists have developed a library that can translate the most popular types machine learning models into a format that can be run on the Field Programmable Gate Arrays used in lowest level triggers. This would allow experiments to quickly identify events from rare signals that have complex signatures that the current low-level triggers don’t have time to look for. 

The LHC experiments are working hard to get the most out their collisions. There could be particles being produced in LHC collisions already but we haven’t been able to see them because of our current triggers, but these new techniques are trying to cover our current blind spots. Look out for new ideas on how to quickly search for interesting signatures, especially as we get closer the high luminosity upgrade of the LHC.

Read More:

CERN Courier article on programming FPGA’s

IRIS HEP Article on a recent workshop on Fast ML techniques

CERN Courier article on older CMS search for low mass dijet resonances

ATLAS Search using ‘trigger-level’ jets

LHCb Search for Dark Photons using fast reconstruction based on a high level trigger

Paper demonstrating the feasibility of running ML models for jet tagging on FPGA’s

Letting the Machines Search for New Physics

Article: “Anomaly Detection for Resonant New Physics with Machine Learning”

Authors: Jack H. Collins, Kiel Howe, Benjamin Nachman

Reference : https://arxiv.org/abs/1805.02664

One of the main goals of LHC experiments is to look for signals of physics beyond the Standard Model; new particles that may explain some of the mysteries the Standard Model doesn’t answer. The typical way this works is that theorists come up with a new particle that would solve some mystery and they spell out how it interacts with the particles we already know about. Then experimentalists design a strategy of how to search for evidence of that particle in the mountains of data that the LHC produces. So far none of the searches performed in this way have seen any definitive evidence of new particles, leading experimentalists to rule out a lot of the parameter space of theorists favorite models.

A summary of searches the ATLAS collaboration has performed. The left columns show model being searched for, what experimental signature was looked at and how much data has been analyzed so far. The color bars show the regions that have been ruled out based on the null result of the search. As you can see, we have already covered a lot of territory.

Despite this extensive program of searches, one might wonder if we are still missing something. What if there was a new particle in the data, waiting to be discovered, but theorists haven’t thought of it yet so it hasn’t been looked for? This gives experimentalists a very interesting challenge, how do you look for something new, when you don’t know what you are looking for? One approach, which Particle Bites has talked about before, is to look at as many final states as possible and compare what you see in data to simulation and look for any large deviations. This is a good approach, but may be limited in its sensitivity to small signals. When a normal search for a specific model is performed one usually makes a series of selection requirements on the data, that are chosen to remove background events and keep signal events. Nowadays, these selection requirements are getting more complex, often using neural networks, a common type of machine learning model, trained to discriminate signal versus background. Without some sort of selection like this you may miss a smaller signal within the large amount of background events.

This new approach lets the neural network itself decide what signal to  look for. It uses part of the data itself to train a neural network to find a signal, and then uses the rest of the data to actually look for that signal. This lets you search for many different kinds of models at the same time!

If that sounds like magic, lets try to break it down. You have to assume something about the new particle you are looking for, and the technique here assumes it forms a resonant peak. This is a common assumption of searches. If a new particle were being produced in LHC collisions and then decaying, then you would get an excess of events where the invariant mass of its decay products have a particular value. So if you plotted the number of events in bins of invariant mass you would expect a new particle to show up as a nice peak on top of a relatively smooth background distribution. This is a very common search strategy, and often colloquially referred to as a ‘bump hunt’. This strategy was how the Higgs boson was discovered in 2012.

A histogram showing the invariant mass of photon pairs. The Higgs boson shows up as a bump at 125 GeV. Plot from here

The other secret ingredient we need is the idea of Classification Without Labels (abbreviated CWoLa, pronounced like koala). The way neural networks are usually trained in high energy physics is using fully labeled simulated examples. The network is shown a set of examples and then guesses which are signal and which are background. Using the true label of the event, the network is told which of the examples it got wrong, its parameters are updated accordingly, and it slowly improves. The crucial challenge when trying to train using real data is that we don’t know the true label of any of data, so its hard to tell the network how to improve. Rather than trying to use the true labels of any of the events, the CWoLA technique uses mixtures of events. Lets say you have 2 mixed samples of events, sample A and sample B, but you know that sample A has more signal events in it than sample B. Then, instead of trying to classify signal versus background directly, you can train a classifier to distinguish between events from sample A and events from sample B and what that network will learn to do is distinguish between signal and background. You can actually show that the optimal classifier for distinguishing the two mixed samples is the same as the optimal classifier of signal versus background. Even more amazing, this technique actually works quite well in practice, achieving good results even when there is only a few percent of signal in one of the samples.

An illustration of the CWoLa method. A classifier trained to distinguish between two mixed samples of signal and background events learns can learn to classify signal versus background. Taken from here

The technique described in the paper combines these two ideas in a clever way. Because we expect the new particle to show up in a narrow region of invariant mass, you can use some of your data to train a classifier to distinguish between events in a given slice of invariant mass from other events. If there is no signal with a mass in that region then the classifier should essentially learn nothing, but if there was a signal in that region that the classifier should learn to separate signal and background. Then one can apply that classifier to select events in the rest of your data (which hasn’t been used in the training) and look for a peak that would indicate a new particle. Because you don’t know ahead of time what mass any new particle should have, you scan over the whole range you have sufficient data for, looking for a new particle in each slice.

The specific case that they use to demonstrate the power of this technique is for new particles decaying to pairs of jets. On the surface, jets, the large sprays of particles produced when quark or gluon is made in a LHC collision, all look the same. But actually the insides of jets, their sub-structure, can contain very useful information about what kind of particle produced it. If a new particle that is produced decays into other particles, like top quarks, W bosons or some a new BSM particle, before decaying into quarks then there will be a lot of interesting sub-structure to the resulting jet, which can be used to distinguish it from regular jets. In this paper the neural network uses information about the sub-structure for both of the jets in event to determine if the event is signal-like or background-like.

The authors test out their new technique on a simulated dataset, containing some events where a new particle is produced and a large number of QCD background events. They train a neural network to distinguish events in a window of invariant mass of the jet pair from other events. With no selection applied there is no visible bump in the dijet invariant mass spectrum. With their technique they are able to train a classifier that can reject enough background such that a clear mass peak of the new particle shows up. This shows that you can find a new particle without relying on searching for a particular model, allowing you to be sensitive to particles overlooked by existing searches.

Demonstration of the bump hunt search. The shaded histogram is the amount of signal in the dataset. The different levels of blue points show the data remaining after applying tighter and tighter selection based on the neural network classifier score. The red line is the predicted amount of background events based on fitting the sideband regions. One can see that for the tightest selection (bottom set of points), the data forms a clear bump over the background estimate, indicating the presence of a new particle

This paper was one of the first to really demonstrate the power of machine-learning based searches. There is actually a competition being held to inspire researchers to try out other techniques on a mock dataset. So expect to see more new search strategies utilizing machine learning being released soon. Of course the real excitement will be when a search like this is applied to real data and we can see if machines can find new physics that us humans have overlooked!

Read More:

  1. Quanta Magazine Article “How Artificial Intelligence Can Supercharge the Search for New Particles”
  2. Blog Post on the CWoLa Method “Training Collider Classifiers on Real Data”
  3. Particle Bites Post “Going Rogue: The Search for Anything (and Everything) with ATLAS”
  4. Blog Post on applying ML to top quark decays “What does Bidirectional LSTM Neural Networks has to do with Top Quarks?”
  5. Extended Version of Original Paper “Extending the Bump Hunt with Machine Learning”

CMS catches the top quark running

CMS catches the top quark running

Article : “Running of the top quark mass from proton-proton collisions at √ s = 13 TeV“

Authors: The CMS Collaboration

Reference: https://arxiv.org/abs/1909.09193

When theorists were first developing quantum field theory in the 1940’s they quickly ran into a problem. Some of their calculations kept producing infinities which didn’t make physical sense. After scratching their heads for a while they eventually came up with a procedure known as renormalization to solve the problem.  Renormalization neatly hid away the infinities that were plaguing their calculations by absorbing them into the constants (like masses and couplings) in the theory, but it also produced some surprising predictions. Renormalization said that all these ‘constants’ weren’t actually constant at all! The value of these ‘constants’ depended on the energy scale at which you probed the theory.

One of the most famous realizations of this phenomena is the ‘running’ of the strong coupling constant. The value of a coupling encodes the strength of a force. The strong nuclear force, responsible for holding protons and neutrons together, is actually so strong at low energies our normal techniques for calculation don’t work. But in 1973, Gross, Wilczek and Politzer realized that in quantum chromodynamics (QCD), the quantum field theory describing the strong force, renormalization would make the strong coupling constant ‘run’ smaller at high energies. This meant at higher energies one could use normal perturbative techniques to do calculations. This behavior of the strong force is called ‘asymptotic freedom’ and earned them a Nobel prize. Thanks to asymptotic freedom, it is actually much easier for us to understand what QCD predicts for high energy LHC collisions than for the properties of bound states like the proton.  

Figure 1: The value of the strong coupling constant (α_s) is plotted as a function of the energy scale. Data from multiple experiments at different energies are compared to the prediction from QCD of how it should run.  From [5]
Now for the first time, CMS has measured the running of a new fundamental parameter, the mass of the top quark. More than just being a cool thing to see, measuring how the top quark mass runs tests our understanding of QCD and can also be sensitive to physics beyond the Standard Model. The top quark is the heaviest fundamental particle we know about, and many think that it has a key role to play in solving some puzzles of the Standard Model. In order to measure the top quark mass at different energies, CMS used the fact that the rate of producing a top quark-antiquark pair depends on the mass of the top quark. So by measuring this rate at different energies they can extract the top quark mass at different scales. 

Top quarks nearly always decay into W-bosons and b quarks. Like all quarks, the b quarks then create a large shower of particles before they reach the detector called a jet. The W-bosons can decay either into a lepton and a neutrino or two quarks. The CMS detector is very good at reconstructing leptons and jets, but neutrinos escape undetected. However one can infer the presence of neutrinos in an event because we know energy must be conserved in the collision, so if neutrinos are produced we will see ‘missing’ energy in the event. The CMS analyzers looked for top anti-top pairs where one W-boson decayed to an electron and a neutrino and the other decayed to a muon and a neutrino. By using information about the electron, muon, missing energy, and jets in an event, the kinematics of the top and anti-top pair can be reconstructed. 

The measured running of the top quark mass is shown in Figure 2. The data agree with the predicted running from QCD at the level of 1.1 sigma, and the no-running hypothesis is excluded at above 95% confidence level. Rather than being limited by the amount of data, the main uncertainties in this result come from the theoretical understanding of the top quark production and decay, which the analyzers need to model very precisely in order to extract the top quark mass. So CMS will need some help from theorists if they want to improve this result in the future. 

Figure 2: The ratio of the top quark mass compared to its mass at a reference scale (476 GeV) is plotted as a function of energy. The red line is the theoretical prediction of how the mass should run in QCD.

Read More:

  1. “The Strengths of Known Forces” https://profmattstrassler.com/articles-and-posts/particle-physics-basics/the-known-forces-of-nature/the-strength-of-the-known-forces/
  2. “Renormalization Made Easy” http://math.ucr.edu/home/baez/renormalization.html
  3. “Studying the Higgs via Top Quark Couplings” https://particlebites.com/?p=4718
  4. “The QCD Running Coupling” https://arxiv.org/abs/1604.08082
  5. CMS Measurement of QCD Running Coupling https://arxiv.org/abs/1609.05331

Going Rogue: The Search for Anything (and Everything) with ATLAS

Title: “A model-independent general search for new phenomena with the ATLAS detector at √s=13 TeV”

Author: The ATLAS Collaboration

Reference: ATLAS-PHYS-CONF-2017-001


When a single experimental collaboration has a few thousand contributors (and even more opinions), there are a lot of rules. These rules dictate everything from how you get authorship rights to how you get chosen to give a conference talk. In fact, this rulebook is so thorough that it could be the topic of a whole other post. But for now, I want to focus on one rule in particular, a rule that has only been around for a few decades in particle physics but is considered one of the most important practices of good science: blinding.

In brief, blinding is the notion that it’s experimentally compromising for a scientist to look at the data before finalizing the analysis. As much as we like to think of ourselves as perfectly objective observers, the truth is, when we really really want a particular result (let’s say a SUSY discovery), that desire can bias our work. For instance, imagine you were looking at actual collision data while you were designing a signal region. You might unconsciously craft your selection in such a way to force an excess of data over background prediction. To avoid such human influences, particle physics experiments “blind” their analyses while they are under construction, and only look at the data once everything else is in place and validated.

Figure 1: “Blind analysis: Hide results to seek the truth”, R. MacCounor & S. Perlmutter for Nature.com

This technique has kept the field of particle physics in rigorous shape for quite a while. But there’s always been a subtle downside to this practice. If we only ever look at the data after we finalize an analysis, we are trapped within the confines of theoretically motivated signatures. In this blinding paradigm, we’ll look at all the places that theory has shone a spotlight on, but we won’t look everywhere. Our whole game is to search for new physics. But what if amongst all our signal regions and hypothesis testing and neural net classifications… we’ve simply missed something?

It is this nagging question that motivates a specific method of combing the LHC datasets for new physics, one that the authors of this paper call a “structured, global and automated way to search for new physics.” With this proposal, we can let the data itself tell us where to look and throw unblinding caution to the winds.

The idea is simple: scan the whole ATLAS dataset for discrepancies, setting a threshold for what defines a feature as “interesting”. If this preliminary scan stumbles upon a mysterious excess of data over Standard Model background, don’t just run straight to Stockholm proclaiming a discovery. Instead, simply remember to look at this area again once more data is collected. If your feature of interest is a fluctuation, it will wash out and go away. If not, you can keep watching it until you collect enough statistics to do the running to Stockholm bit. Essentially, you let a first scan of the data rather than theory define your signal regions of interest. In fact, all the cool kids are doing it: H1, CDF, D0, and even ATLAS and CMS have performed earlier versions of this general search.

The nuts and bolts of this particular paper include 3.2 fb-1 of 2015 13 TeV LHC data to try out. Since the whole goal of this strategy is to be as general as possible, we might as well go big or go home with potential topologies. To that end, the authors comb through all the data and select any event “involving high pT isolated leptons (electrons and muons), photons, jets, b-tagged jets and missing transverse momentum”. All of the backgrounds are simply modeled with Monte Carlo simulation.

Once we have all these events, we need to sort them. Here, “the classification includes all possible final state configurations and object multiplicities, e.g. if a data event with seven reconstructed muons is found it is classified in a ‘7- muon’ event class (7μ).” When you add up all the possible permutations of objects and multiplicities, you come up with a cool 639 event classes with at least 1 data event and a Standard Model expectation of at least 0.1.

From here, it’s just a matter of checking data vs. MC agreement and the pulls for each event class. The authors also apply some measures to weed out the low stat or otherwise sketchy regions; for instance, 1 electron + many jets is more likely to be multijet faking a lepton and shouldn’t necessarily be considered as a good event category. Once this logic applied, you can plot all of your SRs together grouped by category; Figure 2 shows an example for the multijet events. The paper includes 10 of these plots in total, with regions ranging in complexity from nothing but 1μ1j to more complicated final states like ETmiss2μ1γ4j (say that five times fast.)

Figure 2: The number of events in data and for the different SM background predictions considered. The classes are labeled according to the multiplicity and type (e, μ, γ, j, b, ETmiss) of the reconstructed objects for this event class. The hatched bands indicate the total uncertainty of the SM prediction.


Once we can see data next to Standard Model prediction for all these categories, it’s necessary to have a way to measure just how unusual an excess may be. The authors of this paper implement an algorithm that searches for the region of largest deviation in the distributions of two variables that are good at discriminating background from new physics. These are the effective massthe sum of all jet and missing momenta, and the invariant mass, computed with all visible objects and no missing energy.

For each deviation found, a simple likelihood function is built as the convolution of probability density functions (pdfs): one Poissonian pdf to describe the event yields, and Gaussian pdfs for each systematic uncertainty. The integral of this function, p0, is the probability that the Standard Model expectation fluctuated to the observed yield. This p0 value is an industry standard in particle physics: a value of p0 < 3e-7 is our threshold for discovery.

Sadly (or reassuringly), the smallest p0 value found in this scan is 3e-04 (in the 1m1e4b2j event class). To figure out precisely how significant this value is, the authors ran a series of pseudoexperiments for each event class and applied the same scanning algorithm to them, to determine how often such a deviation would occur in a wholly different fake dataset. In fact, a p0 of 3e-04 was expected 70% of the pseudoexperiments.

So the excesses that were observed are not (so far) significant enough to focus on. But the beauty of this analysis strategy is that this deviation can be easily followed up with the addition of a newer dataset. Think of these general searches as the sidekick of the superheros that are our flagship SUSY, exotics, and dark matter searches. They can help us dot i’s and cross t’s, make sure nothing falls through the cracks— and eventually, just maybe, make a discovery.

Why Electroweak SUSY is the Next Big Thing

Title: “Search for new physics in events with two low momentum opposite-sign leptons and missing transverse energy at s = 13 TeV”

Author: CMS Collaboration

Reference: CMS-PAS-SUS-16-048


March is an exciting month for high energy physicists. Every year at this time, scientists from all over the world gather for the annual Moriond Conference, where all of the latest results are shown and discussed. Now that this physics Christmas season is over, I, like many other physicists, am sifting through the proceedings, trying to get a hint of what is the new cool physics to be chasing after. My conclusions? The Higgsino search is high on this list.

Physicists chatting at the 2017 Moriond Conference. Image credit ATLAS-PHOTO-2017-009-1.

The search for Higgsinos falls under the broad and complex umbrella of searches for supersymmetry (SUSY). We’ve talked about SUSY on Particlebites in the past; see a recent post on the stop search for reference. Recall that the basic prediction of SUSY is that every boson in the Standard Model has a fermionic supersymmetric partner, and every fermion gets a bosonic partner.

So then what exactly is a Higgsino? The naming convention of SUSY would indicate that the –ino suffix means that a Higgsino is the supersymmetric partner of the Higgs boson. This is partly true, but the whole story is a bit more complicated, and requires some understanding of the Higgs mechanism.

To summarize, in our Standard Model, the photon carries the electromagnetic force, and the W and Z carry the weak force. But before electroweak symmetry breaking, these bosons did not have such distinct tasks. Rather, there were three massless bosons, the B, W, and Higgs, which together all carried the electroweak force. It is the supersymmetric partners of these three bosons that mix to form new mass eigenstates, which we call simply charginos or neutralinos, depending on their charge. When we search for new particles, we are searching for these mass eigenstates, and then interpreting our results in the context of electroweak-inos.

SUSY searches can be broken into many different analyses, each targeting a particular particle or group of particles in this new sector. Starting with the particles that are suspected to have low mass is a good idea, since we’re more likely to observe these at the current LHC collision energy. If we begin with these light particles, and add in the popular theory of naturalness, we conclude that Higgsinos will be the easiest to find of all the new SUSY particles. More specifically, the theory predicts three Higgsinos that mix into two neutralinos and a chargino, each with a mass around 200-300 GeV, but with a very small mass splitting between the three. See Figure 1 for a sample mass spectra of all these particles, where N and C indicate neutralino or chargino respectively (keep in mind this is just a possibility; in principle, any bino/wino/higgsino mass hierarchy is allowed.)

Figure 1: Sample electroweak SUSY mass spectrum. Image credit: T. Lari, INFN Milano

This is both good news and bad news. The good part is that we have reason to think that there are three Higgsinos with masses that are well within our reach at the LHC. The bad news is that this mass spectrum is very compressed, making the Higgsinos extremely difficult to detect experimentally. This is due to the fact that when C1 or N2 decays to N1 (the lightest neutralino), there is very little mass difference leftover to supply energy to the decay products. As a result, all of the final state objects (two N1s plus a W or a Z as a byproduct, see Figure 2) will have very low momentum and thus are very difficult to detect.

Figure 2: Electroweakino pair production and decay (CMS-PAS-SUS-16-048).

The CMS collaboration Higgsino analysis documented here uses a clever analysis strategy for such compressed decay scenarios. Since initial state radiation (ISR) jets occur often in proton-proton collisions, you can ask for your event to have one. This jet radiating from the collision will give the system a kick in the opposite direction, providing enough energy to those soft particles for them to be detectable. At the end of the day, the analysis team looks for events with ISR, missing transverse energy (MET), and two soft opposite sign leptons from the Z decay (to distinguish from hadronic SM-like backgrounds). Figure 3 shows a basic diagram of what these signal events would look like.

Figure 3: Signal event vector diagram. Image credit C. Botta, CERN

In order to conduct this search, several new analysis techniques were employed. Reconstruction of leptons at low pT becomes extremely important in this regime, and the standard cone isolation of the lepton and impact parameter cuts are used to ensure proper lepton identification. New discriminating variables are also added, which exploit kinematic information about the lepton and the soft particles around it, in order to distinguish “prompt” (signal) leptons from those that may have come from a jet and are thus “non prompt” (background.)

In addition, the analysis team paid special attention to the triggers that could be used to select signal events from the immense number of collisions, creating a new “compressed” trigger that uses combined information from both soft muons (pT > 5 GeV) and missing energy ( > 125 GeV).

With all of this effort, the group is able to probe down to a mass splitting between Higgsinos of 20 GeV, excluding N2 masses up to 230 GeV. This is an especially remarkable result because the current strongest limit on Higgsinos comes from the LEP experiment, a result that is over ten years old! Because the Higgsino searches are strongly limited by the low cross section of electroweak SUSY, additional data will certainly mean that these searches will proceed quickly, and more stringent bounds will be placed (or, perhaps, a discovery is in store!)

Figure 4: Figure 5: The observed exclusion contours (black) with the corresponding 1 standard deviation uncertainties. The dashed red curves present the expected limits with 1 SD experimental uncertainties (CMS-PAS-SUS-16-048).


Further Reading: 

  1. “Natural SUSY Endures”, Michele Papucci, Joshua T. Ruderman, Andreas Weiler.  arXiv [hep-ph] 1110.6926
  2. “Cornering electroweakinos at the LHC”, Stefania Gori, Sunghoon Jung, Lian-Tao Wang. arXiv [hep-ph] 1307.5952