Researchers develop new machine-learning method

Two graduate students at the Facility for Rare Isotope Beams are exploring how models built for scientific computing could benefit artificial-intelligence applications for the nuclear physics community and beyond. 

A research team, led by two graduate students at the Facility for Rare Isotope Beams (FRIB) at Michigan State University (MSU), recently published a paper (“Parametric matrix models“) in Nature Communications showcasing a new way to approach algorithm development for machine learning and artificial intelligence (AI) applications. 

MSU doctoral students Patrick Cook and Danny Jammooa, both graduate assistants at FRIB, led the research. They worked with Dean Lee, professor of physics at FRIB and in MSU’s Department of Physics and Astronomy and Theoretical Nuclear Science Department Head at FRIB; Morten Hjorth-Jensen, professor of physics at the University of Oslo in Norway and former professor of physics at FRIB and in MSU’s Department of Physics & Astronomy; and Daniel Lee, professor of electrical and computer engineering at Cornell Tech.

The team focused on developing algorithms called parametric matrix models (PMMs). While scientists already use machine learning for scientific computing applications, PMMs are a new approach that has advantages over other methods for scientific computing. The results have implications for how to leverage AI technology for nuclear physics research and other fields. 

The group receives funding from the U.S. Department of Energy through the SmarT Reduction and Emulation Applying Machine Learning In Nuclear Environments (STREAMLINE) collaboration. Additional support has come from the National Science Foundation. 

“We were hoping to combine machine learning with another method we had been developing called eigenvector continuation,” said Dean Lee. “We had the good fortune of having two brilliant graduate students looking for related projects. So, from that point on, Patrick and Danny did the work in developing these PMMs.” 

The team found that these models excelled at learning the governing nuclear physics equations for a given problem, which corresponds closely to how scientists seek to solve physics problems—by using the underlying equations to inform modeling and simulations. The team’s proof of concept has implications for how to develop machine-learning methods tailored to the nuclear physics community.

PMMs are an efficient alternative to neural networks 

In the last decade, scientists have increasingly used a combination of experiment and simulation to make headway on research questions. As AI technologies rapidly advance, scientists can train models that further improve time-to-solution. The most common machine-learning method is a class of models called neural networks. As their name implies, neural networks act similarly to how neurons share information in the human bodies. Researchers train these models with large data volumes, providing feedback along that way that help guide the model in identifying complex connections or correlations of interest.

One challenge researchers encountered in nuclear physics applications was that neural networks may not reliably produce important constraints associated with a physical system. Additionally, they may not achieve the desired accuracy needed for scientific computing. They may also be difficult to interpret and difficult to extrapolate to new data. 

PMMs overcome these limitations by finding the governing equations for a nuclear physics model and producing a parameterized solution. Lee pointed out that physicists often solve equations that look simple but generate complex, intricate solutions. But by using PMMs in the context of machine learning, researchers can efficiently model the equations themselves rather than modelling the solutions directly. 

“We realized that PMMs were well-suited for applications where a researcher might be able to say, ‘If it were a smaller version of this problem, I could solve it,’ because PMMs ultimately are focused on finding what the smaller model is for a given problem,” Cook said. 

“Neural networks can be used for everything from large language models like ChatGPT to image identification and generation and a variety of challenges in between,” Jammooa said. “They are great universal approximators. However, they may not be the perfect tool for every job. I think of a neural network more like a Swiss Army knife, where you have a competent tool for any given type of job. PMMs are more like a 3D printer, where you can print the exact tool you need for the job.” 

Cook and Jammooa demonstrated PMMs’ superior performance for training applications based on physics principles in three different contexts—quantum-computing modeling, modeling quantum many-body systems, and general machine-learning problems (like multivariable regression and image classification). The team noted that in practice, PMMs were significantly smaller than neural networks because they were better able to fully incorporate important mathematical structures for a physical problem of interest, as a result requiring less computational resources. 

Dean Lee indicated that PMMs were already supporting practical applications. “We have started using PMMs in collaboration with FRIB accelerator scientists to help tune the particle beams for improved performance,” he said. “At FRIB, there are different collaborators to work with in different disciplines and contexts. That is a unique advantage of working at a U.S. Department of Energy Office of Science user facility like FRIB—PMMs can be used across all these other physics-related systems and support different research communities in different contexts.” 

Eric Gedenk is a freelance science writer. 

Michigan State University (MSU) operates the Facility for Rare Isotope Beams (FRIB) as a user facility for the U.S. Department of Energy Office of Science (DOE-SC), with financial support from and furthering the mission of the DOE-SC Office of Nuclear Physics. Hosting the most powerful heavy-ion accelerator, FRIB enables scientists to make discoveries about the properties of rare isotopes in order to better understand the physics of nuclei, nuclear astrophysics, fundamental interactions, and applications for society, including in medicine, homeland security, and industry. User facility operation is supported by the DOE-SC Office of Nuclear Physics as one of 28 DOE-SC user facilities.

The U.S. Department of Energy Office of Science is the single largest supporter of basic research in the physical sciences in the United States and is working to address some of today’s most pressing challenges. For more information, visit energy.gov/science.