Tiny grains of dust floating inside a glowing plasma should, according to decades of theory, push and pull on each other in predictable ways. But when physicists at Emory University turned a machine-learning system loose on real particle trajectories captured in three dimensions, the AI returned force laws that textbooks do not contain.
“We designed the AI not just to predict what particles would do next, but to discover the rules governing their interactions,” said Ilya Nemenman, a professor of physics and biology at Emory and a senior author of the study. “What it found was that some of the most common assumptions about how charged dust grains interact in plasma are simply not correct.”
The results, published in the Proceedings of the National Academy of Sciences in early 2025 under the title “Physics-tailored machine learning reveals unexpected physics in dusty plasmas,” represent a rare instance in which a machine-learning tool did not merely replicate known physics but flagged previously unrecognized behavior. The paper is now drawing attention from researchers in fusion energy, semiconductor fabrication, and astrophysics, all fields where plasma interactions matter and where standard models may have been quietly missing non-linear effects.
Tracking dust in three dimensions
The experiment hinged on a technique called scanning laser sheet tomography. A thin laser sheet sweeps rapidly through a plasma chamber, illuminating individual dust particles suspended in the ionized gas. High-speed cameras record each sweep, and software stitches the snapshots into long, continuous three-dimensional trajectories for every grain in the field of view.
That measurement method was developed and validated in a separate technical paper cataloged by the U.S. Department of Energy’s Office of Scientific and Technical Information. The precision and duration of the resulting trajectory datasets are what made the next step possible: instead of assuming a force law and fitting parameters to the data, the team let a physics-constrained neural network infer the forces directly from observed motion.
“We gave the AI Newton’s second law and conservation principles, but we did not tell it what the force between two particles should look like,” said Justin Burton, an associate professor of physics at Emory and the study’s other senior author. “It had to figure that out from the trajectories alone.”
Where the textbooks fall short
For decades, physicists have modeled the interaction between charged dust grains in plasma using a Yukawa (screened Coulomb) potential: a simple formula in which the repulsive force between two grains drops off exponentially with distance, governed by a single screening length set by the surrounding plasma. The assumption works well enough for many calculations, and it has become a default starting point in simulations of dusty plasmas.
The AI disagreed. When the neural network inferred effective forces from the Emory trajectories, it returned curves that deviated from the Yukawa form in ways that were statistically significant and physically consistent across multiple experimental runs. The inferred forces showed non-linear features at intermediate distances, a regime where the standard model predicts a smooth, monotonic decline. According to the PNAS paper’s supplemental materials, the team ran extensive robustness checks and uncertainty analyses to rule out artifacts from noise, finite sample size, or model architecture choices.
The deviations are not small tweaks. They suggest that the effective interaction between dust grains depends on local plasma conditions in ways that a single screening parameter cannot capture. If confirmed, this would mean that simulations built on the Yukawa assumption have been systematically underestimating or mischaracterizing collective behavior in dusty plasmas.
What independent replication will need to show
The findings, while published in a top-tier peer-reviewed journal, still rest on a single experimental and analytical pipeline. The raw trajectory datasets have not been made publicly available, which means outside groups cannot yet re-derive the force laws or test whether the non-linear effects persist across different plasma pressures, gas compositions, or dust grain sizes.
No formal responses from plasma physics groups beyond Emory have appeared in the published literature as of June 2026. The absence of independent expert commentary leaves a gap between the novelty of the result and the community consensus that typically follows a major claim. Burton and Nemenman’s statements, while detailed, come from their own institution’s communications channels. Whether teams at national laboratories or other universities have begun replication efforts is not yet clear from available sources.
A broader survey-style preprint titled “DustNET: enabling machine learning and AI models of dusty plasmas” (arXiv: 2603.17493) provides useful context for understanding where this work fits within a growing push to apply AI to plasma science. But the preprint functions as intellectual scaffolding, not independent confirmation of the specific non-linear effects the Emory team reported.
Why it matters beyond the lab
Dusty plasmas are not exotic curiosities. They appear inside tokamak fusion reactors, where wall erosion sends microscopic particles into the plasma and alters confinement. They form in the plasma etching chambers used to manufacture semiconductor chips, where particle contamination can ruin wafers. And they exist in vast quantities in protoplanetary disks, Saturn’s rings, and comet tails, where dust-plasma interactions shape the structures astronomers observe.
If the standard Yukawa model has been missing non-linear force contributions, the downstream consequences could ripple through all of these domains. Fusion simulations might need revised transport coefficients. Semiconductor process models might require updated contamination predictions. Astrophysical dust simulations might produce different morphologies once the corrected forces are plugged in.
But those implications remain speculative until the new force laws are tested against macroscopic observables in real-world or realistic simulated systems. The PNAS paper does not include such cross-validation, and no follow-up study addressing it has appeared in the available literature.
A new role for AI in fundamental physics
Perhaps the most striking aspect of the Emory result is not the specific force law it uncovered but the method that uncovered it. Physics-informed machine learning, in which neural networks are constrained by known conservation laws but left free to discover unknown relationships, has been gaining traction across disciplines. Most applications to date have used AI to speed up simulations or interpolate between known data points. Cases in which the AI genuinely surprises its creators by finding physics they did not expect remain uncommon.
“This is what we hoped machine learning could do for physics,” Nemenman said. “Not just fit data faster, but show us something we were not looking for.”
Whether that something holds up under scrutiny from the broader community will determine whether this study becomes a footnote or a turning point. The data exist. The method is published. Now the field needs to decide whether to trust the machine.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
