Physicist Huilin Qu vividly remembers the 2012 announcement of the discovery of the Higgs boson, a fundamental particle that helps explain the origin of mass.
The spokespeople of the ATLAS and CMS experiments at the Large Hadron Collider gave detailed presentations on their experiments’ findings. Qu was an undergrad studying physics at the time.
“We watched the webcast from a big auditorium at my university,” Qu says. “I arrived late and was sitting in the back. I didn’t understand much, but I was still very amazed.”
During the presentation by the spokesperson of the CMS experiment, Qu had an idea. “I thought, ‘Oh, this is the person I should apply to work with. Let's give it a try,’” he says.
Two years later, Qu was a PhD student at the University of California, Santa Barbara, doing research with postdoc Loukas Gouskos, under the supervision of the (now former) CMS spokesperson Qu had watched on the webcast that day, Professor Joe Incandela.
Their team’s goal was to look for new particles at the LHC, the world’s most powerful particle accelerator. But this time Gouskos was the one struck with an idea: Instead of performing traditional searches, he proposed taking a step back from physics analyses for a year to develop a new machine-learning tool.
“It was Mission: Impossible. Our goal was to make it possible.”
His hope was that the new tool would allow them to narrow in on properties of the Higgs boson—ones that many physicists thought could not be tested at the LHC.
“It was Mission: Impossible,” Gouskos says. “Our goal was to make it possible.”
Gouskos won the support of Qu and the rest of the team, including Incandela. “Joe was always very open to new ideas,” says Gouskos, who is now a professor at Brown University. “He didn’t need much convincing.”
The origin of mass
The Higgs boson is a rare and short-lived fundamental particle. Scientists can momentarily produce Higgs bosons by using a particle accelerator to inject a huge amount of concentrated energy into the Higgs field, a substance that fills the entire universe and through which all particles must swim. The particles that interact the most with the Higgs field swell with mass, and when a Higgs boson reaches the end of its lifetime, it preferentially transforms into these mass-loving children in a process called particle decay.
One way to figure out if a particle interacts with the Higgs field (and to figure out the strength of this interaction) is to look at how frequently a Higgs boson decays into it. For instance, the majority of Higgs bosons decay into bottom quarks, fundamental particles that are like the quarks that make up protons and neutrons, but significantly heavier. This tells physicists that bottom quarks gain mass by interacting with the Higgs field and have a close relationship with the Higgs boson.
The Standard Model of particle physics divides particles into three generations. After more than a decade of studying the Higgs boson, scientists have confirmed that the Higgs field interacts with the heaviest, third-generation particles (the top quark, the bottom quark, and the tau lepton). But does the Higgs also have a relationship with the two lighter generations of particles, including the first-generation quarks that make up the nuclei of all of the matter that we can see and touch? The theory says yes, but this has never been confirmed experimentally.
Qu and Gouskos wanted to see if machine learning could get them closer to the next step: observing a Higgs boson decaying into a quark in the second generation. “We have no evidence that the Higgs interacts with second-generation quarks,” Gouskos says.
According to the theory, the Higgs boson should decay into charm quarks, which are second-generation particles, about 3% of the time. “It’s 20 times smaller than the decay rate of Higgs bosons into bottom quarks,” says Qu, who is now a research physicist at CERN. “Discovering the Higgs boson decaying into bottom quarks was already enormously complicated, and here all the same challenges still apply, with even more obstacles.”
The Goldilocks quark
Even though Higgs bosons predominantly decay into bottom quarks, it took scientists an additional six years after the original discovery of the Higgs to pin down this decay channel. In fact, the Higgs boson was discovered through much rarer decay channels, such as one in which the Higgs boson morphs into a short-lived virtual particle and then transforms into two particles of light called photons. This process happens only 0.2% of the time, but because photons are very distinct, finding them in the data is relatively straightforward.
Quarks, on the other hand, are very messy particles.
"We never see the quarks,” Gouskos says. “Instead, they turn into something we call jets; each one is a spray of about 50 to 100 particles.”
Scientists were able to finally pin down the decay of Higgs bosons into bottom quarks in 2018, in part thanks to machine-learning algorithms that used the bottom quark’s large mass and decay patterns to separate the particle jets it created from similar-looking jets instigated by lighter quarks.
But this method of sifting the heavy from the light doesn’t work for charm quarks, Qu says: “The properties of the charm quarks are in between.”
A new tool
According to Qu, identifying the original quark that instigated a particle jet is a nearly impossible task for humans. “I never tried, but I would imagine that if I had to identify different types of jets, I would not do much better than a 50-50 guess,” he says.
That’s because humans are not good at multidimensional problems. "Each jet could have up to 100 particles in it, and for each particle we measure properties like its displacement, charge and momentum. This means that for each jet, we can have a few hundred, or even a few thousand input features.”
That’s why Qu and Gouskos decided to try machine learning.
Most physics machine-learning models borrow from image recognition (they treat each particle like a pixel in an image) or natural language processing (they treat each particle like a word in a sentence). This is how Qu, Gouskos and their team also started, but they realized that they would have much more sensitivity if they borrowed machine learning techniques from an emerging technology: self-driving cars.
“Self-driving cars use sensors to create a collection of points in space, and each point is assigned spatial coordinates and other properties,” Qu says. “The key turning point was when we realized that we shouldn’t represent a jet as a sequence like a sentence, but as a cloud of points.”
The team also programmed physics-specific rules into the algorithm to help it make sense of the data. “If you inject physics knowledge, it helps the algorithm learn,” Gouskos says. “For instance, we injected information about how to combine the masses and momenta of secondary particles to reconstruct the mass of the parent particle that produced them.”
Processing power
But as the team made their tool more and more powerful, they found that they no longer had the computing power to support it. Luckily, they were able to get gear from a related industry: PC gaming.
“Gamers need really powerful GPUs,” Qu says. “At the time, the CERN computing cluster didn’t have such powerful GPUs, and so we thought, why don't we just build something ourselves? It's not so expensive.”
After ordering four GPUs, Qu watched a YouTube video on how to build a computer.
“I was very scared. I had never built anything before, but in the end, it turned out to be very easy,” Qu says. “It would have been a great computer for gaming, except we installed Linux, which doesn't support any games, and it doesn’t have a screen.”
The group used their custom computer to train their machine-learning algorithms to look for charm-quark jets that might have originated from Higgs bosons. When they finally tested their method, they found that they had improved on the traditional jet classification techniques by an order of magnitude. They were able to classify a jet in less than 30 milliseconds and misidentified them less than 1% of the time.
“The first time we presented this improvement, the reaction we got from the CMS collaboration was, ‘There must be something wrong. It cannot be true,’” Qu says.
Let the search begin
The team spent the next year checking for any potential errors and validating their model. Their next step was to use their machine-learning tool on a real physics analysis.
Instead of starting with the Higgs, they looked for the similar, but much more common, process of Z bosons decaying into pairs of charm quarks.
Their analysis worked. “The agreement between the observed data and the predictions was spot on,” Gouskos says. “It was textbook agreement.”
With a successful demonstration under their belts, they put their machine-learning tool to the test, attempting to observe Higgs bosons decaying to charm quarks.
The team knew that if this process was truly as rare as the Standard Model predicted, it would be like trying to see a molecule with a microscope. They wouldn’t have enough data to get a clear view of it until after the LHC’s upgrade to the High-Luminosity LHC. But they wanted to give it a try.
Previous limits confirmed that the rate of Higgs boson decays into charm quarks could not be more than 10,000 times the expected rate predicted by the Standard Model. Using their tool, Gouskos, Qu and the rest of their team improved that limit by four orders of magnitude. “Now we know that it must be below four or five times the expected value,” Gouskos says.
Even though the Higgs-to-charm decay remains elusive, the results showed that—with more data—observing it might be within reach.
“We're doing absolutely incredible things,” Qu says. “People thought we wouldn’t have this level of sensitivity until the end of the HL-LHC, and we have already surpassed those projections.”
