UVA Researchers Built an AI Algorithm That Understands Physics

Normally, when testing the behavior of materials under high heat or explosive conditions, researchers have to run simulation after simulation, a data-intensive process that can take days even on a supercomputer. However, with a deep learning algorithm created by Stephen Baek, Phong Nguyen and their research team, the process takes less than a second on a laptop. 

Baek and Nguyen’s latest findings, part of research sponsored by the Air Force Office of Scientific Research and the Designing Materials to Revolutionize and Engineer our Future program within the National Science Foundation, were published in the most recent volume of the Science Advances journal. Collaborators include mechanical engineering professor H.S. Udaykumar and a team of computational mechanics at the University of Iowa. 

Baek and Nguyen, both faculty members in the UVA School of Data Science, believe the algorithm they developed, called physics-aware recurrent convolutions, or PARC, has profound implications in materials science and in other areas governed by complex physical processes, ranging from climate change models to the growth dynamics of some cancers and other diseases to atmospheric conditions on Mars. 

“The most exciting thing about PARC is its ability to learn physics purely from data without the need for an explicitly defined governing equation, which is a unique feature compared to other AI models for physics,” said Nguyen, who is an assistant professor of data science. “This capability enables PARC not only to solve known physics problems but also to discover unknown physics from data.”

Most AI algorithms, the co-authors explained, are “physics-naïve,” meaning that the algorithms account for physics variables but might do so arbitrarily, rather than according to the laws of physics we already know.

Often, “the way that the mathematical terms and variables interact with each other are not constrained,” said Baek, an associate professor of data science and mechanical and aerospace engineering. “For example, if you are interested in learning the dynamics of a car from data, physics will tell you that the acceleration of the vehicle must be a function of the thrust, or force, aerodynamic resistance, or drag, and friction. However, AI models don't know such facts, so they will just mix up those variables arbitrarily, in whatever way that makes sense of data.”

Physics-aware models like PARC are designed to associate different variables with physics laws. Essentially, the team taught PARC physics, enabling the algorithm to mirror how supercomputers solve complex physics equations and solve them in a fraction of the time.

“PARC can look at data containing complex patterns of violent changes in temperature and pressure during the detonation of explosives and propellants, and tell what is the equation that governs the dynamics,” Baek said.

In materials science, this has implications for material design, which is typically “a lengthy trial-and-error process,” according to Nguyen. 

“Computer simulation can take days even with supercomputing facilities; meanwhile, physical experiments can be even worse. This bottleneck limits the number of designs that can be explored and slows down the material development process,” he said. “With PARC, we can now drastically speed up the search process by reducing simulation time from days to milliseconds, even on a normal laptop. This means that we can explore and test many more designs and converge on the desired design much faster, leading to faster pace of innovation in material science.”

Much of the paper focuses on using PARC to determine how different materials respond during the detonation of explosives and propellants, with the goal of making them safer. However, both Baek and Nguyen envision PARC having a much broader impact. 

“The same mathematical and computational principles of PARC could also be generalized to other areas of sciences where the interest is to understand some complex dynamics from observational data,” Baek said. “For example, climate change could be modeled and predicted by learning the governing equation of atmospheric dynamics using PARC; growth dynamics of some diseases, such as cancer, could be modeled and predicted by learning the governing equation of biological growth using PARC; or complex quantum behavior could be modeled and understood by learning the governing equation of a complex quantum device using PARC.”

Or, Nguyen added, “imagine exploring a place like Mars, where we have limited knowledge about physics, and all we have is data. In such a scenario, an AI tool like PARC would be incredibly helpful.”

The group has already started working with, for example, UVA associate professor of physics Gia-Wei Chern on applying PARC to quantum physics. They are also working to better extract and present PARC’s findings in what Nguyen calls a more “human-interpretable format” – clear, digestible and able to inform decision-making in a variety of ways.