AI reveals one-sided forces in plasma that appear to violate Newton’s third law

Physics has a rule that feels almost sacred. Every action produces an equal and opposite reaction. It’s the kind of principle you learn early and never question again. Except someone should have been questioning it, because out in the dusty plasma clouds that stretch across most of the universe, that rule simply stops working.

Particles out there push each other without getting pushed back the same way. The interaction is one-sided, lopsided, and by every traditional measure, illegal. Physicists have known about this problem for a while. What they didn’t have was a way to actually measure it properly. That just changed.

A team at Emory University, led by Wentao Yu, trained a machine learning model to read the hidden rules inside this chaos. The findings came out in PNAS, and they raise an uncomfortable possibility: some of what our textbooks have been saying about how planets form might simply be wrong.

The weird physics of floating dust

To understand what the Emory team did, you first need a picture of what dusty plasma actually looks like up close.

Inside a vacuum chamber filled with argon gas, microscopic dust grains levitate inside an electric field. They hang there, suspended, but they’re not still. Around them, a stream of positive ions rushes past at over two kilometers per second. That ion flow leaves a wake behind each particle, the same basic idea as the trail behind a speedboat cutting through water.

Those wakes are where the trouble starts. They create small pockets of attraction in a system that should, by all standard physics, only be repulsive. And because the wakes all point the same direction toward an electrode, the symmetry that physics depends on breaks down. One particle feels its neighbor’s wake. The neighbor might feel nothing at all in return.

That’s not how the universe is supposed to work. But there it is.

Teaching a machine to see the rules, not just the data

Most machine learning tools are pattern machines. Feed them enough data and they’ll tell you what happens next, but they won’t tell you why. Yu’s team built something with more backbone than that.

They constructed physics-constrained neural networks, meaning the AI wasn’t left to figure everything out from scratch. It was given a skeleton based on Newton’s second law and then split into three separate networks, one handling grain-to-grain interactions, one accounting for the surrounding environment, and one dealing with gas drag on the particles. Each had a specific job.

Using scanning laser sheet tomography, the team captured detailed 3D motion data from the particles. The AI worked through that data and inferred the mass and charge of each grain in real time, measurements that are genuinely difficult to nail down in a live experiment. The model’s accuracy score came back above 0.99. For context, 1.0 is perfect.

That near-perfect fit wasn’t just a technical win. It confirmed that these strange, one-sided forces are real and measurable.

The number that changes everything

Here’s where the findings get genuinely disruptive.

Standard plasma physics says the screening length, basically the distance at which a particle’s electric field fades out, should depend only on the plasma itself, not on the particles floating inside it. The Emory data showed the particle size actually shifts that value. That alone was unexpected.

But the charge-to-mass ratio finding was harder to brush off. For decades, a framework called Orbital-Motion-Limited theory has predicted that a particle’s charge scales with its mass to the power of one-third. Clean, consistent, settled science. The AI found that in real experimental conditions, that power ranges anywhere from 0.3 to 0.8, and it shifts depending on gas pressure.

If that holds up, and the accuracy of the model suggests it does, then the simulations we’ve been running about how matter clumped together in the early solar system are working with the wrong numbers. The rate at which dust accumulates into larger bodies, eventually into planets, could be meaningfully different from what we’ve assumed.

Why specks of dust in a lab deserve your attention

Dusty plasma isn’t some exotic laboratory curiosity. It’s in Saturn’s rings. It fills interstellar nebulae. It shows up in the semiconductor manufacturing process that produces the chip in your phone. Some researchers think it may have played a role in the chemistry that eventually led to life.

The bigger prize here, though, is the method itself. The Emory team showed that you can drop particles into an unknown environment, watch how they move, and let an AI reverse-engineer the physics governing them. No intrusive instruments. No assumptions baked in beforehand. Just motion, math, and inference.

That approach isn’t locked to plasma research. It could work on colloids, on biological systems, on anything where many particles interact in ways that are too complex to untangle by hand.

For a long time, the universe’s messiest systems have been out of reach, not because we lacked curiosity, but because our tools weren’t sharp enough. That gap is closing faster than most people realize.