> Paper: Yu, W., Abdelaleem, E., Burton, J.C., & Nemenman, I. (2025). *Physics-tailored machine learning reveals unexpected physics in dusty plasmas*. PNAS.
Overview
In 2025, a team of physicists at Emory University published a striking study in *PNAS*. Using a "physics-tailored" neural network trained on the motions of charged microparticles in a laboratory dusty plasma, they reconstructed interparticle forces with over 99% accuracy—and found that several long-standing theoretical assumptions about dusty plasmas are wrong.
Most dramatically, the AI confirmed and quantified a force that violates Newton's third law: action does not always equal reaction in this system.
Background: Dusty Plasmas
A dusty plasma is an ionized gas (plasma) containing tiny solid particles. These micron-scale grains acquire negative charge (roughly 10⁴ elementary charges) by collecting electrons, then interact electromagnetically with each other and with the surrounding plasma.
Dusty plasmas are ubiquitous:
- Saturn's rings — the mysterious "spokes" observed by Voyager 2 in 1980 are collective motions of charged dust.
- The lunar surface — electrostatically levitated dust coated Apollo astronauts' suits.
- On Earth — wildfire smoke can interfere with radio communications; in semiconductor fabrication, charged dust contaminates wafers during plasma processing.
- A higher particle attracts a lower particle.
- A lower particle repels (or weakly attracts) a higher one.
- Non-reciprocity is essentially absent in the horizontal plane.
- Yu, W., Abdelaleem, E., Burton, J.C., & Nemenman, I. (2025). *Physics-tailored machine learning reveals unexpected physics in dusty plasmas*. PNAS.
- Institution: Emory University (Departments of Physics and Biology)
- Funding: NSF, Simons Foundation
- Author outcomes: Wentao Yu → Caltech postdoc; Eslam Abdelaleem → Georgia Tech postdoc
Why Dusty Plasmas Are Hard to Study
With 10–20+ charged particles interacting through shielded Coulomb forces, plasma sheath fields, ion drag, and thermal noise, exact analytical solutions are impossible. Physicists have long relied on approximations:
1. Charge proportional to size — larger particles capture more electrons. 2. Yukawa (screened Coulomb) forces: \(F \propto \frac{q_1 q_2}{r^2} e^{-r/\lambda}\), with the screening length λ taken as the electron Debye length (~1–2 mm). 3. Pairwise additivity of interactions. 4. Reciprocity — particle A's force on B equals B's force on A in magnitude.
The Experiment
In Justin Burton's lab, a vacuum chamber filled with low-pressure argon is ionized by RF fields. 10–20 melamine-formaldehyde microspheres (diameters ~12.8, 9.5, and 8.0 µm) levitate at the edge of the plasma sheath. A laser-sheet scanning system with high-speed cameras reconstructs full 3D trajectories over a ~10 mm × 10 mm × 1 mm volume for several minutes.
The Physics-Tailored Neural Network
Rather than a black box, the network's architecture is shaped by physical constraints:
1. Translation invariance — forces depend only on relative positions. 2. Rotation invariance — forces rotate with the frame. 3. Pairwise interactions at lowest order. 4. Handling of non-identical particles of three different sizes.
Crucially, the network is allowed to learn non-reciprocal forces. Trained directly on 3D trajectories, it achieved \(R^2 > 0.99\).
The Physics: Ion Wakes and Non-Reciprocity
Supersonic ions flow downward toward the electrode. Each negatively charged particle repels ions, creating an "ion wake"—a region of reduced ion density downstream that acts like a positive virtual charge.
For two vertically aligned particles: the upper particle's wake extends downward and attracts the lower particle, but the lower particle's wake flows away from the upper one. Result: the upper particle strongly attracts the lower one, while the lower one barely attracts (or even repels) the upper one. Momentum is transferred to the flowing ions, so the two-particle subsystem alone does not conserve momentum—Newton's third law breaks because the system is coupled to a non-equilibrium reservoir.
Three Surprises from the AI
1. Charge is independent of particle size
Contrary to theory, inferred charge did not grow with size. Charge appears to be set mainly by vertical position in the sheath—the local electron density and ion flux—rather than by surface area.
2. The screening length is much shorter than expected
The AI inferred effective screening lengths significantly smaller than the electron Debye length, and size-dependent. This reflects the wake's positive charge partially canceling the particle's negative charge at long range: the traditional two-parameter formula (charge + screening length) is simply insufficient to describe wake-mediated interactions.
3. The precise form of non-reciprocal forces
Broader Implications
Co-senior author Ilya Nemenman notes the framework is general: any complex system of interacting components—colloids, cells, active matter—could be analyzed this way. Non-reciprocal interactions already appear in chemotactic bacteria, bird flocks, and even opinion dynamics.
The study also raises a philosophical question: the AI was the first "discoverer," with human physicists acting as interpreters. Reliability was addressed via cross-validation—two independent methods of inferring particle mass gave consistent results.
Conclusion
Newton's third law (1687) holds for isolated systems, but breaks when coupled to non-equilibrium environments. Even in a classical field studied for centuries, secrets remained—and this time, a physics-constrained neural network found them.
References