AI Plasma Physics Discovery: The Fourth State of Matter

The AI plasma physics discovery worth explaining to a 10-year-old is not that a computer was clever. It’s what the computer found: two dust particles in a plasma push on each other unequally. A leading particle attracts the one behind it, while the trailing particle repels the leader. That breaks the everyday rule that forces come in equal and opposite pairs. An Emory University team led by experimental physicist Justin Burton and theoretical physicist Ilya Nemenman, with first author Wentao Yu, published the result in the Proceedings of the National Academy of Sciences (2025, vol. 122, issue 31), describing these non-reciprocal forces with greater than 99% accuracy. “We showed that we can use AI to discover new physics,” Burton said.

Key Takeaways

• The system was a custom physics-structured neural network, not a general-purpose model. It was built to separate particle motion into exactly three pieces: drag from the surrounding gas, environmental forces, and particle-to-particle forces.
• Data came from real laboratory experiments: tomographic imaging, where a laser sheet sweeps through a plasma chamber while cameras record 3D particle positions over time.
• Two textbook assumptions fell. Electric charge does not increase in simple proportion to particle size, and particle size changes how fast the force falls off with distance.
• The headline finding is non-reciprocity: particle A’s pull on B differs from B’s push on A, because each particle distorts the plasma around it and the wake is asymmetric.
• It took more than a year of design refinement, used limited experimental data, and required careful human interpretation. This was not a model dropped onto a dataset.

What plasma is, and what “dusty plasma” means

Plasma is the fourth state of matter: a gas hot or energized enough that electrons separate from their atoms, leaving a soup of free electrons and positive ions. It conducts electricity and responds to magnetic fields. Lightning is plasma. So is the inside of a neon sign, the surface of the sun, and the glow in a plasma ball at a science museum.

A dusty plasma adds small solid particles — micron-scale dust grains — into that soup. The grains pick up electric charge from the surrounding electrons and ions, then interact with each other. This sounds exotic and is actually everywhere: Saturn’s rings, comet tails, the material that formed planets, and the interiors of semiconductor fabrication chambers where stray particles ruin chips.

Dusty plasma is also a gift to physicists, because the grains are big enough to photograph individually. You can literally watch each particle move. That’s why it became the test bed for this experiment.

How the AI plasma physics discovery actually worked

What the data was. Not simulations. Real particles in a real chamber, tracked in three dimensions. The team used tomographic imaging: a sheet of laser light scans through the plasma while cameras capture where particles are, repeatedly, building a 3D movie of trajectories.

Why a general-purpose network wouldn’t do. A standard neural network fed positions and asked to predict future positions would learn something, but you couldn’t read a force law out of it, and with limited data it would overfit. So the team built structure into the model.

The structure they imposed. The network decomposed each particle’s motion into three separate contributions: drag from the neutral gas, forces from the environment (chamber walls, electric fields, gravity), and forces from other particles. That decomposition is the key design choice. It means the output isn’t a black-box prediction; it’s a labeled estimate of each physical term, which physicists can read, question, and compare to theory.

What it computed. Given observed accelerations, the network solved for the inter-particle force as a function of separation and particle size, while simultaneously accounting for drag and environmental forces so those didn’t contaminate the answer.

What it got right. The inferred non-reciprocal forces matched the observed motion to better than 99% accuracy. And two prior assumptions turned out to be wrong: charge on a grain does not scale in simple proportion to grain size, and grain size affects the rate at which the force decays with distance. Both of those were corrections to existing theory, not confirmations of it. Nemenman’s framing: “We’re able to correct these inaccuracies because we can see what’s occurring in exquisite detail.”

What it cost. More than a year of refining the model architecture. Limited experimental datasets. And human interpretation at every stage. That’s the honest version of “AI discovered new physics,” and it makes the achievement more impressive rather than less.

Why non-reciprocal forces are strange, and why they’re allowed

Newton’s third law says every action has an equal and opposite reaction. Push a wall, the wall pushes back equally. That law holds for direct interactions between two objects.

Here’s the loophole. In a plasma, particles don’t interact directly. Each charged grain distorts the surrounding cloud of ions, creating a wake. The second particle then feels the wake, not the first particle. And a wake has a direction. The grain in front produces a wake that reaches backward; the grain behind sits inside that wake, while the one in front does not sit in a symmetric wake of its own.

So the system as a whole still conserves momentum, because the plasma is carrying some. But if you only look at the two dust grains and ignore the medium, the forces between them look unequal. That’s non-reciprocity, and it shows up in other media too: bird flocks, active colloids, some biological systems.

For a 10-year-old the version is simpler. Two boats on a river: the one upstream sends ripples to the one downstream, but not the other way around. The water is doing the carrying.

The four states of matter, and where plasma sits

State	What the particles are doing	Everyday example	Can a kid see it?
Solid	Locked in place, vibrating	Ice cube	Yes
Liquid	Touching but sliding past each other	Water in a glass	Yes
Gas	Far apart, moving freely	Steam, air	Indirectly (steam, balloons)
Plasma	Electrons stripped from atoms; charged and conductive	Lightning, neon sign, plasma ball, the sun	Yes, in a plasma ball or a flame’s hottest part
Dusty plasma (a plasma variant)	Plasma plus micron-scale charged grains	Saturn’s rings, comet tails, chip-fab chambers	Only in a lab

Plasma is the most common state of visible matter in the universe by a wide margin, because stars are plasma. It’s the one kids learn last and see most.

How to Teach Your Kid About Plasma and AI Discovery

Ages 5–8: Four States on One Counter

Put an ice cube in a pan, melt it, boil it. Three states in ten minutes. Then, for the fourth, visit a plasma ball at a store or museum, or point at a candle’s hottest blue region, and say: hotter still, and the atoms come apart. Ask them to name all four. That’s the whole lesson at this age, and it sticks because they watched three of them happen.

Ages 9–12: The Unequal Push Experiment

Float two ping-pong balls in a bathtub. Push one gently so it moves toward the other and watch what happens to the second ball. Now try reversing which one moves. The water carries the interaction, and the effect is not symmetric in time or space. Then explain: in a plasma, the “water” is a cloud of charged particles, and the AI figured out the shape of that unequal push by watching the dust move.

Ages 13+: Fit a Force Law to Real Motion

Record a slow-motion video of two magnets sliding toward each other on a smooth table (phone slow-mo is plenty). Step through frames and log positions. Compute acceleration, and therefore force, at several separations. Plot force versus distance and try fitting a power law. Your teen just did, by hand and with two objects, what the Emory network did with hundreds of particles in three dimensions and a drag term to untangle. The PNAS paper is open enough to skim for the figure showing the inferred force curve.

The question to ask: “If the two particles push on each other unequally, where did the missing push go?”

What to actually do at home

Start with plasma as a state, not as science fiction

Kids hear “plasma” and think of TV shows. Anchoring it to lightning, neon signs, and the sun makes it ordinary. Ordinary is where curiosity actually lives, and our guide on how LED lights work pairs well because both involve electrons doing something specific.

Emphasize the structure, not the intelligence

The key design move was decomposing motion into three named terms. A model with structure produced readable physics; a model without structure would have produced a prediction nobody could interpret. That idea shows up throughout modern science, and it’s the same principle behind physics-informed machine learning.

Point out that the AI corrected theory, not just confirmed it

Two textbook assumptions about charge and force decay turned out to be wrong. Kids often assume science is finished and AI just applies it. Here the opposite happened, which is a better and more accurate picture of how science works.

Use a plasma ball if you can find one

They’re cheap and available, and watching filaments follow a finger is a genuinely good demonstration of a conductive, responsive medium. Discuss why the filaments move toward your hand: you’re changing the electrical environment.

What not to do

Don’t say the AI “figured out physics on its own.” Humans chose the experiment, built the imaging system, designed a network with physics-shaped internals, and spent over a year refining it. The AI extracted a force law from data within a structure humans provided. That’s a real and significant contribution, and describing it accurately is what lets a kid understand what these tools are for.

What to Watch For Over the Next 3 Months

• Week 4: Your kid can name all four states of matter and give one real example of plasma that isn’t the sun.
• Month 2 red flags: They describe Newton’s third law as “wrong.” It isn’t. Ask where the momentum went, and steer them to the medium.
• Month 3 self-check: Ask what the researchers had to build before the AI could do anything. If they mention the imaging setup or the three-term structure, they’ve understood that the instrument and the model design were the hard parts.

Frequently Asked Questions

Did AI really discover new physics, or is that hype?

In this case the claim is unusually well-supported, and also narrower than it sounds. The network inferred previously uncharacterized non-reciprocal force relationships in dusty plasma and corrected two specific assumptions about charge and force decay. Burton’s own statement is “we can use AI to discover new physics,” which is a claim about method. Humans designed the experiment and the model’s internal structure.

What is a non-reciprocal force?

A force where object A’s effect on B is not equal and opposite to B’s effect on A. It appears when the two objects interact through a medium that carries momentum, so accounting only for the two objects makes the books look unbalanced. Newton’s third law still holds for the complete system including the plasma.

Is plasma dangerous for kids to be around?

Consumer plasma balls are designed to be safe to touch and are sold as toys and decorations. Lightning, welding arcs, and industrial plasma equipment are not. The distinction is energy, not the state of matter itself. Normal caution applies to hot or high-voltage sources.

Why use dust particles instead of studying the plasma directly?

Because dust grains are large enough to image individually while still being fully charged participants in the plasma. That gives researchers something they almost never have: direct, particle-by-particle observation of trajectories in a many-body system. Electrons and ions move far too fast and small to track this way.

Does this have any practical use?

Semiconductor manufacturing is the nearest one: fabrication chambers use plasmas, and stray dust particles ruin chips, so understanding how grains move and cluster matters commercially. Beyond that, non-reciprocal interactions appear in astrophysics (Saturn’s rings, planet formation) and in the active-matter physics behind flocking and some biological motion.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Yu, W., Abdelaleem, E., Nemenman, I., & Burton, J. C. (2025). Physics-structured machine learning of non-reciprocal forces in dusty plasma. Proceedings of the National Academy of Sciences, 122(31). https://www.pnas.org/
2. ScienceDaily / Emory University. (2026, April 23). “AI discovers new physics in plasma.” https://www.sciencedaily.com/releases/2026/04/260422044635.htm
3. NASA Science. “Plasma: the fourth state of matter.” https://science.nasa.gov/heliophysics/focus-areas/magnetosphere-ionosphere/
4. National Institute of Standards and Technology. Plasma science and measurement programs. https://www.nist.gov/
5. Google Research. (2026, May 28). “A new era of innovation: Google Research at I/O 2026” (context on AI systems that propose experiments). https://research.google/blog/a-new-era-of-innovation-google-research-at-io-2026/
6. Shen, C., et al. (2025). Differentiable physics-plus-ML modeling. Nature Communications. Penn State summary: https://www.psu.edu/news/research/story/ai-powered-model-predicts-floods-improves-water-management-worldwide