Theoretical physicists use imagination and a deep understanding of mathematics to decipher the fundamental laws of the universe that govern particles, forces, and everything in between. Increasingly, theorists are doing that work with the help of machine learning.
As expected, the group of theorists who use machine learning includes those classified as “computational” theorists. But it also includes “formal” theorists, those concerned with the self-contradictions of theoretical frameworks such as string theory and quantum gravity. And that includes “phenomenologists,'' or theorists who sit next to experimentalists and formulate hypotheses about new particles and interactions that can be tested experimentally. Analyze the data collected in the experiment. We then figure out how to use those results to build new models and test them experimentally.
In all areas of theory, machine learning algorithms are speeding up processes, performing calculations that were previously impossible, and even forcing theorists to rethink the way they study theoretical physics.
“We're really nearing the beginning of what, to me, is a clear revolution in the use of computers in scientific discovery,” said Jim Halverson, a professor of physics at Northeastern University. “It's like being within 50 years of Galileo first pointing a telescope to the sky. Many of the current advances utilize machine learning.”
David See, a professor in Rutgers University's Department of Physics and Astronomy, thinks similarly about machine learning. “Instead of looking at the sky, you're looking at the data,” he says. “So now we can look deeper into our data. We're breaking new ground.”
Machine learning for theoretical physics
“The kinds of theoretical calculations I perform don't have datasets, and machine learning is used to speed up first-principles theoretical calculations in a mathematically precise way,” said MIT Physics Associate. says professor and researcher Fiala Shanahan. Theoretical nuclear and particle physics.
Based on a fundamental understanding of particle physics, Shanahan uses lattice field theory to calculate the structure of protons, neutrons, and atomic nuclei. She uses her machine learning to calculate calculations that are “much faster than they could have been calculated, or perhaps couldn't be calculated any other way,” and guaranteed to be accurate. He says he is doing it.
Lattice field theory calculations are computationally intensive and take a very long time on conventional computers, leading to what Shanahan calls “large computational programs” in her field. Machine learning algorithms make calculations faster and more doable, but theorists still need to use supercomputers to perform the calculations.
Shanahan and his collaborators recently demonstrated that machine learning can generate samples from the fundamental probability distributions associated with lattice field theory without using any “training data.”
The hope is that machine learning algorithms will allow physicists like Shanahan to directly calculate the properties of atomic nuclei, such as argon and xenon, that are too large to study using traditional approaches. This research will help future experiments such as deep underground neutrino experiments and various dark matter searches that use such nuclei as targets in experimental instruments.
As Shanahan shows, physicists don't necessarily need to have data to use machine learning algorithms. But when you have data, especially large amounts of data, machine learning is a very powerful tool for processing it.
Halverson incorporates machine learning into his string theory research. The equations that underpin string theory contain many possible solutions that theorists must sort through. “But the numbers are astronomical, so a brute force scan is simply not possible.”
For example, Halverson has worked on theoretical datasets containing more than 10^755 elements, far exceeding the number of particles in the universe. “With this huge data set, you might imagine performing some kind of search problem according to the rules of a well-defined game,” he says. “We're looking for specific things, but we also have to meet certain constraints.”
In this way, conducting theoretical research on string theory can be similar to playing games such as chess or Go. So Halverson and his colleagues used a machine learning approach called reinforcement learning, often used in gameplay settings, to create algorithms that can explore astronomically large systems and pinpoint the data of interest. did.
Theorists also use machine learning for so-called discovery applications, which search for hidden correlations in structures or hidden relationships in raw information. For example, theorists working with data from particle colliders work with complex mathematical equations that physicists must simplify before calculating how particles scatter. Machine learning can help accelerate this process by suggesting possible solutions. It is much easier for theorists to adopt and test a proposal than to come up with it from scratch.
Shih also uses machine learning to classify data in his phenomenological research. He works with the European Space Agency's Gaia telescope, which catalogs the positions and velocities of every star in the Milky Way.
Recently, Shih and his colleagues combined theory and Gaia data to generate a three-dimensional map of the density of dark matter in galaxies. “All of this is made possible by relatively new machine learning techniques that were not available five years ago,” he says. “Until recently, it was unimaginable to do data analysis like this.”
Working with NANOGrav scientists working on pulsar timing arrays, Shih used another machine learning technique called simulation-based inference. NANOGrav has to perform calculations by inverting a huge matrix, a process that previously took about a week. Machine learning can perform these calculations using simulated samples of data. The process takes him nearly 24 hours and creates a database that astrophysicists can sample in seconds.
Some phenomenologists are using machine learning to redefine the way physicists explore new physics.
Traditionally, theorists formulate a hypothesis, define what it would be like for an experimenter to find evidence that the hypothesis is true, and then ask the experimenter to look for that evidence. But with machine learning, theorists can formulate a hypothesis, define what happens if we deviate from that hypothesis, and use algorithms to look for evidence that the hypothesis is incorrect.
“This is very controversial because what we typically do in science is hypothesis testing, A vs. B,” said MIT professor of physics and director of the National Science Foundation's Artificial Intelligence and Fundamentals Interactions. says Jesse Thaler, founding director of the Institute for Action Research (IAIFI). “The idea that now you might say, 'Let's look for unusual features,' without specifying what you're looking for,” is a different way of doing science.
Benefits and challenges
For many theorists, machine learning has already proven to be a promising tool for advancing research. “More classical or traditional approaches typically involve bending the data, reducing it to fewer dimensions, or fitting the data to a very simple model with a small number of parameters,” he said. says Mr. “Of course, that leads to a lot of biases and assumptions and information being lost along the way.
“With these modern machine learning techniques, you don't have to do any of that. You can use all your data with minimal assumptions.”
However, as Thaler notes, physicists continue to express concerns about using machine learning in theoretical physics. One problem is that some algorithms make predictions without uncertainty. And physicists have long worried that machine learning is too much of a “black box,” meaning it reaches decisions without showing how it works.
That's why Halverson and his colleagues are working to prove that machine learning algorithms can produce understandable results. “We are establishing valid and rigorous results that will pass the mathematician's sniff test, both in string theory and in a broader context,” Halverson said.
This effort will help physicists establish new standards not only in physics but also in machine learning. “We in particle physics have very high standards for what it means to discover something, or what it means to do a rigorous analysis, so in a sense, “We're at the forefront of transforming machine learning,” Thaler said.
“We are moving from off-the-shelf tools that may not incorporate all the best practices in physics to tools that not only incorporate the best practices in physics but can be exported to other fields.”
The future of machine learning and theory
The use of machine learning in both experiments and theory has blurred the lines between two traditionally distinct camps. In fact, some argue that a new type of physicist is emerging: the data physicist. Shih coined the term to describe scientists at the confluence of experiment, theory, and data science at the 2022 Snowmass U.S. High Energy Physics Community Planning Conference. Although the title data physicist is not yet commonly used, physicists who know how to analyze large amounts of data are in high demand. And machine learning is already deeply ingrained in this type of work.
Shih advocates recruiting and retaining more young people who know how to use machine learning in physics. “We lost a lot of people to industry,” he says. “We need jobs to build a robust pipeline of talent to keep them in the field.
“I think we're doing well at the postdoctoral level and, of course, at the graduate student level, but we need to create more faculty jobs in this interdisciplinary machine learning and data science area in physics and astronomy.”
Theorists say they believe these jobs are here to stay. And they're not afraid that machine learning will replace it.
Thaler acknowledges that machine learning may eventually be able to do what theoretical physicists can do, but only if physicists have a deep understanding of their science and are able to He states that this is limited to cases that can be explained to a computer.
“Practically expressing some aspects of the scientific process in rigorous algorithmic terms so that they can also be performed on computers is a rich scientific endeavor in itself, and has the potential to really accelerate the methods of scientific discovery.” “It’s a hidden endeavor,” he says.
Ultimately, Shi says, theorists see machine learning as a “hammer-like” tool. “You have a universal tool and you can apply it in many different places.”
“It's just a type of algorithm,” says Shanahan, research coordinator for theoretical physics at IAIFI. “We hope that machine learning, like any algorithm, has the advantage of allowing us to do things that we couldn't do otherwise.”
Machine learning, when used well, could make physicists' lives a little easier. It may also return the amount of time currently spent performing calculations and analyzing data.
“We are constantly learning about the enormous size of the world, which may be hiding fascinating phenomena, whether they are new phenomena beyond the standard model or phenomena within the standard model that we have not yet seen. We have the dataset,” Thaler said. “If you have to spend an entire PhD thesis studying a small possibility, you can't explore the vast space of possibilities fast enough, given the large amount of data coming in.
“Given the fact that each of us on earth has a limited amount of time and a limited number of people paying attention to our data, we want to maximize our ability to discover new phenomena. “Yes,” he says. “Collaborating with computers seems to be one way for him to do that.”
