Physicists have used machine learning approaches to uncover unexpected details about how particles interact within complex systems. Their research focuses on non-reciprocal forces, where one particle affects another particle differently and vice versa.
The survey results are PNASthe result of a collaboration between experimental and theoretical physicists at Emory University. By combining a custom neural network with laboratory data from dusty plasma, the research team showed that artificial intelligence can do more than analyze data and make predictions. It helps you discover completely new laws of physics.
“We showed that AI can be used to discover new physics,” says Justin Barton, Emory Professor of Experimental Physics and senior co-author of the paper. “Our AI method is not a black box. We understand how it works and why it works. The framework that AI provides is also universal. It has the potential to be applied to other many-body systems, opening new avenues of discovery.”
Precision insights into dust plasma forces
This study provides one of the most detailed explanations to date of the physics governing dusty plasmas. This system consists of an ionized gas filled with interacting charged particles, including small dust particles.
Using an AI model, the researchers were able to describe non-reciprocal forces with more than 99% accuracy. These forces are notoriously difficult to measure and model.
“We can describe these forces with more than 99% accuracy,” said Ilya Nemenman, professor of theoretical physics at Emory and co-senior author of the paper. “What’s even more interesting is that we showed that some common theoretical assumptions about these forces are not completely accurate. We can now see what’s going on in great detail, so we can correct these inaccuracies.”
The research team believes the method is broadly applicable to systems consisting of many interacting components. These range from industrial materials such as paints and inks to groups of living cells.
The study’s lead author is Wentao Yu, who worked on the project as an Emory doctoral student and is now a postdoctoral fellow at the California Institute of Technology. Co-author Eslam Abdelalim also contributed as an Emory graduate student and is currently a postdoctoral fellow at Georgia Tech.
This research was primarily supported by the National Science Foundation, with additional funding from the Simons Foundation.
“This project is a great example of interdisciplinary collaboration where the development of new knowledge in plasma physics and AI can lead to further advances in the study of living systems,” said Vyacheslav (Slava) Lukin, program director of the NSF Plasma Physics Program. “The dynamics of these complex systems are governed by collective interactions, and emerging AI techniques can help us better describe, perceive, understand, and even control them.”
Description of the fourth state of matter
Plasma is often referred to as the fourth state of matter. In this state, the gas is ionized and electrons and ions move freely, creating unique properties such as electrical conductivity. Plasma makes up about 99.9% of the visible universe, from the solar wind flowing from the sun to lightning strikes on Earth.
Dust plasma contains additional electrically charged dust particles and appears in many environments, from Saturn’s rings to Earth’s ionosphere.
On the Moon, the weak gravity causes electrically charged dust to float on the surface. “That’s why when astronauts walk on the moon, their suits get covered in dust,” Burton explains.
On Earth, soot can mix with smoke and form dusty plasma during wildfires. These charged particles can interfere with radio signals, making it more difficult for firefighters to communicate.
Tracking particle movement in 3D
Burton’s lab studies dust plasmas and similar materials by recreating them in controlled experiments. Researchers suspend tiny plastic particles inside a vacuum chamber filled with plasma to simulate more complex systems. By adjusting the gas pressure, you can mimic real-world situations and observe how particles respond to different forces.
For this project, Burton and Yu developed a tomographic imaging method that captures the three-dimensional (3D) motion of particles. A laser sheet moves through the chamber and a high-speed camera records images. These snapshots can be combined to reconstruct the positions of dozens of particles over time, allowing researchers to track their movements in great detail.
Use AI to understand collective movement
A theoretical biophysicist, Nemenmann studies how complex systems emerge from simple interactions. He is particularly interested in collective movements, such as how cells move within the human body.
“The general question of how whole systems emerge from the interactions of small parts is very important,” Nemenman explains. “For example, in cancer, we want to understand how cell interactions are involved in some cells leaving the tumor and migrating to new locations and metastasizing.”
Compared to living systems, dust plasmas provide an easier environment to test new ideas. This made it an ideal case to explore whether AI could discover new physical principles.
“There is a lot of talk about how AI is revolutionizing science, but there are very few examples of something fundamentally new being directly discovered by an AI system,” Nemenman says.
Designing neural networks for discovery
Building the AI model required careful planning. Unlike systems trained on large datasets, this project had limited experimental data.
“When you’re investigating something new, there’s not a lot of data to train the AI on,” Nemenman explains. “That means we need to design neural networks that can learn new things while training on small amounts of data.”
The team spent more than a year refining the design through weekly meetings.
“We needed to build a network that followed the necessary rules while allowing us to explore and reason about unknown physics,” Barton explains.
“It took over a year of back and forth discussions in weekly meetings,” Nemenman added. “Once we came up with the correct structure of the network we needed for training, it turned out to be very simple.”
The final model separated particle motion into three main influences: drag due to velocity, environmental forces such as gravity, and forces between particles.
Amazing results and new insights
After training on 3D particle trajectories, the AI was able to capture complex interactions involving asymmetric forces between particles.
Researchers liken this behavior to two boats crossing a lake. Each boat creates waves that affect the other boats. Depending on the wave’s position, these waves push or pull the boat in different ways.
“In a dusty plasma, we described how leading particles attract trailing particles, but trailing particles always repel leading particles,” Nemenmann explains. “This phenomenon was predicted by some, but we now have an accurate approximation that didn’t exist before.”
This result also calls into question previous theories. One long-standing idea suggested that a particle’s charge increases in direct proportion to its size. The new findings show that larger particles carry more charge, but the relationship is more complex and depends on factors such as plasma density and temperature.
Another assumption was that the forces between particles were independent of particle size and decreased exponentially with distance. The AI model revealed that particle size affects how quickly these forces weaken.
The research team confirmed these conclusions through additional experiments.
New tools for exploring complex systems
Researchers have developed a physically-based neural network that can run on a standard desktop computer. They believe this provides a flexible framework for studying many-body systems across a variety of disciplines.
Nemenmann will soon be teaching at the Konstanz School of Collective Behavior in Germany. There, scientists study systems ranging from flocks of birds to crowds of humans.
“I will be teaching students from all over the world how to use AI to infer the physics of collective motion – not in dusty plasmas, but in living systems,” he says.
Even with these advances, human expertise remains essential. Scientists must carefully design models and interpret results.
“Developing and using AI tools in ways that truly advance science, technology, and the humanities requires critical thinking,” Barton says.
He remains optimistic about the future.
“I think it’s like the Star Trek motto: boldly go where no one has gone before,” Burton says. “When used properly, AI can open the door to entirely new realms.”
