When one thinks about CERN, the mind conjures visions of particles
colliding at near light speed, a sense of grandeur perhaps at the
towering reaches of intellectual curiosity, or the looming eight-story
tall ATLAS detector that is just one element of the 27km Large Hadron
Collider. These are the images that came into the zeitgeist of
popular culture after the monumental discovery of the Higgs Boson
in 2012. However it is, not the CERN that I have experienced.
Not many people are aware that CERN is involved with countless other
areas of research in particle physics. One of these areas is neutrino
physics, and the building of different types of neutrino detectors.
At the tip of the CERN site where it crosses the border into France,
tucked away between the
industrial warehouses on one side and an expanse of fields on the
other, is building 595 where I spent the summer of 2016, working
with other neutrino physicists at CERN.
Neutrino physics has its own LHC-scale testament to human innovation.
This is the Super-Kamiokande detector,located deep within a mountain
in Japan and filled with 50 kilotons of pure water.
Unlike accelerators, neutrino detectors are often containers
of immense amounts of fluid. There are lots of neutrinos showering the
Earth all the time, but because they interact so weakly you need
a lot of material to have any chance that some will produce
an observable reaction, which does occur every so often.
Shown above is one such event, a muon neutrino producing a shower
of light onto the cylindrical walls of the Super-K
detector. For a sense of scale, each pixel in the
image is a Photo-Multiplier Tube (PMT) which is about three feet in
Super-Kamiokande is a Cherenkov detector which means that it measures
the light (electromagnetic radiation) that is emitted when
charged particles move close to the speed of light through the water.
Charged particles like this appear when an
incoming muon neutrino (or electron neutrino) interacts with
one of the regular old electrons of the water molecules to
become an actual muon (or electron), suddenly finding itself
charged and emitting light in a cone shape which you see in the image.
Charged particles also interact electromagnetically with the water
(unlike the neutral neutrinos) causing them to scatter instead of propagating
in a straight line.
Since muons are much heavier than electrons the scattering is less
pronounced, which is why they project much sharper circles of light
onto the detector walls. One of the main functions of these detectors is to measure the rate of
incoming electron neutrinos versus muon neutrinos. Since each type of
neutrino only produces its corresponding charge particle, this
process allows us to differentiate between
electron rings and muon rings in the detector.
To the naked eye the two types of rings are easily differentiable, but
the software that exists to automate the task needs to be carefully
configured, and requires a great deal of computing resources.
It just so happens that Convolution Neural Networks (CNNs)
are great at visual tasks that are easily done by humans, so I was given the job to build one.
I started by generating a large amount of simulated data to populate
the training set. I then built a classification network on top of a
CNN that guessed the type of particle, which could be compared to the
actual type for training.
The standard algorithm has an extremely high accuracy for this type of
particle identification, about 99.85%, which was a daunting figure when
I was just starting out and my algorithm was coming in at 65%, tops. But
after four months of hard work, many optimisations, and a whole lot of
coffee, my algorithm finally made it past that mark, settling in at 99.9% --
the same in performance but taking a fraction of the time to run and
not requiring any technical knowledge of the Super-K detector to train.
There was still plenty to do before it could be deployed on actual data,
but the summer was over and I had to hand over the reigns to future
Thin Gap Chambers
CERN collaborates with hundreds of universities on their experiments.
Chief among them, the ATLAS particle detector which is an eight-story
high, 44m long piece of equipment that boggles the mind with its
sheer complexity. Its purpose is to measure the plethora of particles
that are produced when two protons collide at 99.999999% the speed of
light. It not only has to measure the existence of particles but also
their trajectory and energy in order to reconstruct the process that
All the particles in the standard model have vastly different properties,
and as such require different instrumentation to detect. The quarks,
for example exhibit confinement which means that a stable pair of quarks
that separate from each other in a high-energy process do not propagate
freely and instead quickly devolve into jets of hadrons and gluons. These
jets are measured by a series of calorimeters in the innermost part of the
detector. Electrons can also be measured in this region since they are
relatively light and charged. Neutrinos can not be detected
at all and are instead
inferred by subtracting the total energy of the detected particles from the
energy that went into the collision.
Muons, on the other hand, are much more massive than electrons and typically make
it out of the detector without being significantly slowed down. On the
outermost part of ATLAS are a set of Thin Gap Chambers (TGCs) that are made up
of four layers composed of a mesh of wires that measure the 2D position of muons as they pass through.
The positions of the muon in each layer can be combined to reconstruct its
In what was supposed to be the 2018 upgrade, the ATLAS detector needed better
TGCs to resolve the muon trajectories sufficiently for the new project
objectives. The job of commissioning and testing these new detectors was given
to the ATLAS group at McGill University where I was working in 2015.
Testing the resolution of these detectors was an integral part of the project
in order to make sure that they would meet the ATLAS specifications, however
without a particle accelerator on hand the detectors had to be profiled
by tracking cosmic ray muons.
This posed a problem because cosmic ray muons typically have a much lower
energy than muons produced in high-energy collisions, and as such have
a much higher potential to scatter randomly within the detector itself.
The output of the TGCs are four points in 3D space, which is reconstructed
to form a line. The line is not perfect, however, partially due to detector
error and partially from the scattering of the muons. If one only has
access to the output data, as in an experiment, it is unclear how much of the variation is
coming purely from the error in the detector itself and how much from the
That is where I came in. My job was to run simulations of the muons as they
passed through the experimental apparatus, and characterize the error
introduced by the scattering process. The simulation required information
on the spectrum of cosmic ray muons, taking into account that the concrete
in the physics building would attenuate many of the low-energy particles.
The four positions are used to reconstruct the trajectory of the muon
by doing a simple fit; a residual is the in-plane distance between the points
themselves and the reconstructed trajectory. Most of the project
was devoted to modelling the statistics of these residuals in the case of
pure scattering, detector resolution error, and the combination of the two.
The mathematical model of the residual distribution incorporated many concepts
from statistics, physics, and geometry. In these sorts of projects it is not
enough to run the simulate and measure the error that is introduced. In experimental
physics the idiom is often used: 'The devil is in the details'. Every aspect
of the physical process must be carefully understood because a small oversight
can easily propagate and derail the analysis of the experiment. That is the
beauty of creating a model, where each equation is justified
mathematically and physically and then shown to fit observations perfectly, thus
establishing the conceptual framework which is underlies the data.