Solves decades-old math problem with AI
- Emory University physicists reported that a physics-guided neural network inferred force laws in dusty plasma experiments and uncovered interactions that standard models missed. - The model learned one-way particle forces from 3D trajectories with R² above 0.99, then measured charge and screening lengths that diverged from common assumptions. - The paper points to AI systems that infer physical laws directly from data in many-body systems. (pnas.org)
Plasma is an ionized gas, and a dusty plasma adds tiny charged particles that push and pull on each other through the surrounding gas. Those forces are hard to write down because they can be one-way, with particle A affecting particle B differently than B affects A. (pnas.org) A team at Emory University said it used a physics-guided neural network to infer those hidden force laws from laboratory data, then tested the results against real particle motion. The paper, “Physics-tailored machine learning reveals unexpected physics in dusty plasmas,” was accepted July 1, 2025 and published in Proceedings of the National Academy of Sciences. (pnas.org) (news.emory.edu) The authors are Wentao Yu, Eslam Abdelaleem, Ilya Nemenman and Justin C. Burton. Yu was first author as an Emory PhD student and is now a postdoctoral fellow at the California Institute of Technology. (pnas.org) (news.emory.edu) Instead of treating the neural network as a black box, the group built physical constraints into the model so it respected symmetries and differences between particles. They trained it on 3D particle trajectories from dusty-plasma experiments and reported average fits above 0.99 R². (pnas.org) That let the team work backward from motion to force, like inferring the shape of a hill by watching how marbles roll across it. Using those inferred forces, the researchers said they could estimate particle masses in two independent ways and get matching answers. (pnas.org) The paper says the model then measured particle charge and a quantity called screening length, which describes how quickly electric influence fades with distance in plasma. Those measurements showed large deviations from theoretical assumptions commonly used for dusty plasmas. (pnas.org) Dusty plasmas matter outside the lab because they appear in places such as Saturn’s rings and interstellar space, and researchers also use them to study planet formation and industrial plasma processes. The paper says the same inference approach could extend to other many-body systems, including colloids and living tissues. (pnas.org) (news.emory.edu) Burton said the group showed it could use artificial intelligence “to discover new physics,” while Nemenman said the method corrected inaccuracies in standard assumptions by resolving the forces in finer detail. The National Science Foundation funded the work, with additional support from the Simons Foundation. (news.emory.edu) The result is narrower than “AI solved math” and more specific: the system inferred interaction laws in a many-particle plasma experiment that physicists could then validate in the lab. The next test is whether the same kind of physics-guided model can recover hidden laws in other messy systems where equations are still incomplete. (pnas.org) (news.emory.edu)
