Physicist Filin vividly remembers the 2012 announcement of the discovery of Higgs Boson, a basic particle that helps explain the origins of Mass.
A spokesperson for Atlas and CMS experiments at a large hadroncorider gave a detailed presentation on the discovery of the experiment. QU was an undergraduate studying physics at the time.
“We saw a webcast from my university's big auditorium,” says Qu. “I arrived late and sat in the back. I didn't really understand, but I was still very surprised.”
During the presentation by a spokesman for the CMS experiment, QU had an idea. “Yeah, this is the person I should apply to work with. Let's give it a try,” he says.
Two years later, QU was a doctoral student at the University of California, Santa Barbara, conducting research with postdoc Lucas Guscos under the supervision of (current) CMS spokesman QU, and under the supervision of Professor Joe Incandela, whom he had seen in the webcast that day.
Their team's goal was to search for new particles in LHC, the world's most powerful particle accelerator. However, this time Gouskos ran into an idea. Instead of performing traditional searches, he proposed to retreat a year from analyzing physics and develop new machine learning tools.
His hope was that new tools would allow them to narrow down the characteristics of Higgs Boson. This is something many physicists thought could not be tested with LHC.
“It was a mission: it was impossible,” says Gouskos. “Our goal was to make that possible.”
Gouskos has gained support for QU and the rest of the team, including Incandela. “Joe has always been very open to new ideas,” says Guscos, now a professor at Brown University. “He didn't need much persuasiveness.”
Origin of mass
Higgs Boson is a rare, short-lived basic particle. Scientists can instantly generate Higgs bosons by using particle accelerators to inject huge amounts of concentrated energy into the Higgs field, the material that fills the entire universe and requires all particles to swim. The particles that most interact with Higgsfield expand in mass, and when the Higgs boson reaches the end of its lifetime, they are preferentially converted to these mass-loving children in a process called particle decay.
One way to know if a particle interacts with the Higgs field (and how strong this interaction is) is to look at how often Higgs Boson falls into it. For example, the majority of Higgs bones decrease to base quarks, which constitute protons and neutrons but are significantly heavier quarks. This tells physicists that the bottom quark gains mass by interacting with Higgsfield and has a close relationship with Higgsboson.
The standard model of particle physics divides particles into three generations. After studying Higgs bosons for more than a decade, scientists have confirmed that Higgsfield interacts with the heaviest third-generation particles (upper quark, lower quark, and taurepton). But does Higgs also have a relationship with two light generation particles, including the first generation quarks that make up the core of all the problems we can see? The theory says yes, but this has never been confirmed experimentally.
QU and Gouskos wanted to see if machine learning could approach the next step. “There is no evidence that the Higgs interact with second-generation quarks,” Guscos says.
According to theory, Higgs bosons should collapse into Charm Quark, a second-generation particle, about 3% of the time. “It's 20 times smaller than the quark at the bottom of the attenuation rate of Higgs Bosons,” says QU, now a research physicist at CERN. “It's already very complicated to discover Higgs bosons collapsed in the bottom quarks, and here all the same challenges still apply and there are more obstacles.”
Goldilock Squark
The Higgs Bosons collapsed primarily into the bottom quark, but it took scientists six more years after Higgs' first discovery to anchor this collapse channel. In fact, Higgs Boson was discovered through a very rare attenuation channel, such as the Higgs Boson transforming into a short-lived virtual particle, converting it into two particles of light called photons. This process only occurs 0.2% of the time, but the photons are very clear, so finding them in the data is relatively easy.
Quarks, on the other hand, are very messy particles.
“We don't see quarks,” Guscos says. “Instead, they turn into what we call jets, each of which is a spray of about 50-100 particles.”
Scientists were able to finally fix the Higgs Boson collapse in the bottom quark in 2018. This is thanks to a machine learning algorithm that uses the large mass and decay pattern of the lower quark to separate similarly-looking jets stimulated by light quarks.
However, this method of sifting heavy objects from Light does not work with Charm Quark, QU states: “Charm Quark properties are in between.”
New tools
According to Qu, identifying the original quark that incited the particle jet is a near-impossible task for humans. “I've never tried it, but I imagine that if I had to identify different types of jets, it would be far better than the 50-50 guess,” he says.
This is because humans are not good at multidimensional problems. “Each jet can contain up to 100 particles, and for each particle we measure its properties such as its displacement, charge, momentum, etc., which means that each jet can have hundreds or even thousands of input capabilities.”
So QU and Gouskos decided to try out machine learning.
Most physics machine learning models borrow from image recognition (treating each particle like a pixel in an image) or from natural language processing (treating each particle like a word in a sentence). This is how QU, Gouskos and their teams also started, but they realized that borrowing machine learning technology from emerging technologies and self-driving cars is much more sensitive.
“Automated cars use sensors to create a collection of points in a space, with each point assigned spatial coordinates and other properties,” says Qu. “The important turning point was when I realized that the jet should be represented as a cloud of points, not as a sentence-like sequence.”
The team also programmed physics-specific rules into algorithms to help them understand the data. “Injecting physics knowledge helps algorithms learn,” says Gouskos. “For example, we injected information on how to combine the masses and momentum of secondary particles and reconstruct the masses of the parent particles that produced them.”
Processing capacity
However, as the teams made the tool more powerful, they discovered they no longer had the computing power to support it. Luckily they were able to get gear from related industries: PC games.
“Gamers need a really strong GPU,” says Qu. “At the time, CERN computing clusters didn't have such a powerful GPU, so why not build something on your own? It's not that expensive.”
After ordering four GPUs, QU watched a YouTube video on how to build a computer.
“I was very scared. I had never made anything before, but in the end it turned out to be very easy,” says Qu. “It would have been the best computer for the game, except that it had installed Linux that didn't support the game and had no screen.”
Using custom computers, the group trained machine learning algorithms to search for charm quarkjets that could have been born from Higgs Bosons. When they finally tested their methods, they discovered that they had improved their traditional jet classification techniques by orders of magnitude. They were able to classify the jets in under 30 ms, and could misidentify under 1% of time.
“When I first presented this improvement, the response I got from the CMS collaboration was, “I'm sure there's something wrong. That's not true,” QU says.
Start a search
The following year, the team checked for potential errors and validated the model. Their next step was to use machine learning tools in real physics analysis.
Instead of starting with the Higgs, they looked for a process that disintegrates a similar, but much more common process of Z-boson into a pair of charm quarks.
Their analysis worked. “The agreement between observed data and forecasts was a spot,” Gouskos said. “It was a textbook contract.”
A successful demonstration under their belt led them to the test with the machine learning tool in an attempt to break down Higgs' bones and captivate Quark.
The team knew that this process would be like trying to see molecules under a microscope if it was as rare as standard models predicted. Until after the upgrade of the LHC to high luminosity LHC, they didn't have enough data to see it clearly. But they wanted to give it a try.
Previous restrictions confirmed that the speed of Higgs Boson collapses into charm quarks cannot exceed 10,000 times the predicted rate predicted by the standard model. Using the tool, Gouskos, QU and other teams have improved their limits by four orders. “Now we know it must be less than four or five times the expected value,” Gouskos says.
The Higgs-to-Charm collapse remains elusive, but the results showed that, with the exception of more data, it could be within reach.
“We're absolutely incredible,” Qu says. “People thought they wouldn't have this level of sensitivity until the end of the HL-LHC, so they've already surpassed those predictions.”
