Emory University physicists announced on April 23, 2026, a significant advancement in the application of artificial intelligence to fundamental physics, identifying previously unknown laws governing particle interactions in dusty plasma. By leveraging a physics-tailored neural network and high-precision 3D tracking, the research team successfully modeled non-reciprocal forces in the fourth state of matter with an accuracy exceeding 99%. The findings, published in the Proceedings of the National Academy of Sciences (PNAS), represent a shift from using AI for data processing toward its use as a tool for autonomous scientific discovery.
Dusty plasma is an ionized gas containing micron-sized solid particles, found naturally in planetary rings and comet tails, and terrestrially in environments such as wildfire smoke. Because these systems involve complex, many-body interactions, they have historically been difficult to model using standard analytical equations. The Emory team, led by professors Justin Burton and Ilya Nemenman, addressed this by developing a tomographic imaging system. This hardware setup utilizes a scanning laser sheet and a high-speed camera array to reconstruct the 3D trajectories of dozens of individual particles within a vacuum chamber at centimeter length scales.
The core of the discovery lies in the AI’s ability to decode non-reciprocal forces—interactions where the force exerted by one particle on another is not equal and opposite. The neural network, designated as a physics-informed model, was trained on experimental data to identify these one-way interactions. Technical performance data released on April 23, 2026, indicates the model achieved an R-squared value greater than 0.99 for predicting particle acceleration. Furthermore, the AI successfully corrected long-standing theoretical misconceptions, such as the assumption that a particle's electrical charge is strictly proportional to its radius. The model demonstrated that charge is also significantly influenced by local plasma density and temperature.
The researchers confirmed that the AI could infer particle mass through two independent computational pathways—force drag analysis and interparticle fitting—with near-perfect agreement. This validation reinforces the reliability of the discovered physical laws. Professor Justin Burton stated that the framework is universal and could be applied to other complex systems, including industrial colloids like paint and ink, as well as biological cell clusters.
The project, which included contributions from researchers Wentao Yu and Eslam Abdelaleem, utilized a neural network architecture that respects physical symmetries and constraints, ensuring the output is mathematically interpretable rather than a black box result. This development is expected to provide a new template for researchers across various disciplines to derive governing equations from high-dimensional experimental datasets. The specific DOI for the foundational research is 10.1073/pnas.2505725122.