In a peer-reviewed study published in Proceedings of the National Academy of Sciences (PNAS), physicists at Emory University have built a custom artificial intelligence system that did more than analyze data—it revealed entirely new physical behaviors in a complex state of matter known as dusty plasma.
The researchers—led by Ilya Nemenman and Justin Burton—designed the AI to learn from a small set of 3D experimental data. What it found was unexpected: non-reciprocal interactions between particles and structural corrections to decades-old scientific assumptions.
What Is Dusty Plasma—And Why Is It So Weird?
Dusty plasma, or complex plasma, is a high-temperature ionized gas that contains tiny dust particles, and it shows up in places ranging from Saturn’s rings and lunar dust clouds to wildfire smoke on Earth. While this material is common in space and astrophysical environments, it behaves in unusual ways that have long puzzled researchers.
In typical systems, forces between particles are reciprocal—if one particle pushes or pulls on another, the reverse is also true. But in dusty plasma, that rule breaks down. Forces become asymmetric: a leading particle can pull the one behind it, while the trailing particle repels the leader. This concept, known as non-reciprocal interaction, had been theorized but was never experimentally confirmed—until now.
The research team developed a custom 3D imaging setup using a high-speed camera and laser sheets to track the motion of plastic dust particles in a plasma-filled chamber. From these motion trails, they trained a purpose-built neural network, embedding known physical rules like gravity, drag, and interaction forces directly into the model.
AI Built for Discovery—Not Just Prediction
What sets this work apart is how the AI was used. Most machine learning tools in science are designed to predict outcomes or clean up noisy datasets. This system was trained to discover new rules.
“We showed that we can use AI to discover new physics,” said Justin Burton, a physicist and co-author of the study. “Our AI method is not a black box: we understand how and why it works.” Instead of working with millions of datapoints, the model relied on a smaller but highly detailed dataset, structured to allow for physical interpretation.
The neural network broke down particle motion into three components: drag-based velocity, environmental forces like gravity, and particle-to-particle interactions. Using this structure, it identified and described non-reciprocal forces with over 99% accuracy—a level of precision typically out of reach for traditional experimental methods.
Correcting Long-Standing Assumptions in Plasma Physics
Beyond mapping new interactions, the AI helped refine or overturn some long-accepted ideas in plasma physics. One assumption was that a particle’s electric charge increased directly with its size. The AI found that this wasn’t always true—the relationship also depends on the plasma’s density and temperature.
Another textbook belief was that the inter-particle force weakens exponentially with distance, regardless of size. In fact, the model showed that the rate of decay changes depending on particle size, suggesting more nuanced dynamics at play.
“Some common theoretical assumptions about these forces are not quite accurate,” said Nemenman. “We’re able to correct these inaccuracies because we can now see what’s occurring in such exquisite detail.”
Implications Far Beyond Plasma Research
Perhaps just as noteworthy as the findings themselves is the tool used to uncover them. This AI model didn’t require a supercomputer or cloud infrastructure—it ran on a standard desktop computer. Its accessibility makes the approach appealing for a wide range of scientific fields beyond plasma physics.
The researchers believe their method can be adapted to study many-body systems in biology (like cell migration) or industry (such as paint and material mixtures), where complex interactions are difficult to map using classical equations. Because the AI was designed with explainability in mind, it avoids the “black box” issue that plagues many machine learning approaches.
“For all the talk about how AI is revolutionizing science,” Nemenman said, “there are very few examples where something fundamentally new has been found directly by an AI system.” This work, he hopes, will be the start of more.