AI Discovers New Physics: When Action Doesn't Equal Reaction in Dusty Plasmas
> Paper: Yu, W., Abdelaleem, E., Burton, J.C., & Nemenman, I. (2025). *Physics-tailored machine learning reveals unexpected physics in dusty plasmas*. PNAS.
Key points
- Researchers at Emory University trained a physics-tailored neural network on 3D trajectories of 10–20 charged melamine-formaldehyde microspheres (8–12.8 μm diameter) suspended in a low-pressure argon plasma, reconstructing interparticle forces with R² > 0.99.
- The AI found that three long-standing theoretical assumptions are wrong: (1) particle charge does not scale with particle size—charge depends mainly on vertical position in the plasma sheath; (2) the effective screening length is much shorter than the electron Debye length (1–2 mm) and size-dependent, because the classical two-parameter formula cannot capture wake-mediated interactions; (3) vertical forces are strongly non-reciprocal—upper particles attract lower ones, while lower particles repel upper ones.
- The non-reciprocity arises from ion wakes: supersonic ion flow past a negatively charged particle leaves a region of depleted ion density downstream, acting as a positive virtual charge. This violates Newton's third law locally—momentum is transferred to the flowing ions, and is recovered only when the plasma environment is included in the system boundary.
- Dusty plasmas are ubiquitous: Saturn's rings (including the mysterious B-ring "spokes"), lunar dust levitated by electric fields, wildfire smoke interfering with radios, and contaminating particles in semiconductor plasma processing.
- The authors emphasize the network is not a black box: physical symmetries (translation and rotation invariance, pairwise interactions, handling of unequal particle sizes) are built into its architecture, making the learned force laws interpretable. Reliability was checked via cross-validation—two independent methods for inferring particle mass agreed closely.
- The framework is general: co-senior author Ilya Nemenman suggests it can be applied to any many-body interacting system, including active matter—chemotactic bacteria, bird flocks, cell collectives—where non-reciprocal interactions are common.
- The work exemplifies a shifting paradigm of scientific discovery: AI as discoverer, humans as interpreters. Data are collected, the machine learns the governing force law directly from trajectories, and physicists then translate the result into physical understanding.
- Institution: Emory University (Department of Physics, Department of Biology)
- Funding: U.S. National Science Foundation (NSF), Simons Foundation
- Author whereabouts: Wentao Yu → postdoc at Caltech; Eslam Abdelaleem → postdoc at Georgia Tech
Why this matters
Newton's third law has held for nearly 340 years as a cornerstone of classical mechanics. This study shows that in nonequilibrium systems coupled to a momentum reservoir (here, supersonic ion flow), the law can be locally broken without physics being wrong—it simply requires enlarging the system boundary. That such a hidden violation lurked in a system studied for decades suggests that interpretable, physics-constrained machine learning can uncover overlooked regularities even in "classical" domains.