The Fourth State

A Machine That Sees What Physicists Cannot

The research, published in Proceedings of the National Academy of Sciences in July 2025 and now attracting broad scientific attention, emerged from a collaboration between two Emory physics groups: the experimental lab of Justin Burton and the theoretical group of Ilya Nemenman. First author Wentao Yu, then a PhD student in Burton's group, and co-author Eslam Abdelaleem, who worked with Nemenman, led the work. Both are now postdoctoral fellows - Yu at Caltech, Abdelaleem at Georgia Tech.

Their subject was dusty plasma: ionized gas containing suspended micron-sized particles of electrically charged solid material. The team recorded the full three-dimensional trajectories of these dust particles as they moved through the plasma, then fed that data to a custom neural network built specifically for the task. The network's job was to learn the forces governing each particle's motion - in particular the forces between particles that standard physical models have long struggled to pin down precisely.

What the model found when it got there was surprising. Two assumptions physicists had relied on for decades turned out to be not quite right. First: that the electrical charge a dust particle carries is proportional to its radius. Larger grain, proportionally larger charge. The AI found this is not necessarily true - the relationship between size and charge is more complex than the field assumed. Second: that the rate at which inter-particle forces weaken over distance is independent of particle size. The AI found that particle size does affect how quickly those forces decay.

"What's even more interesting," said Nemenman, "is that we show that some common theoretical assumptions about these forces are not quite accurate. We're able to correct these inaccuracies because we can now see what's occurring in such exquisite detail."

What Plasma Is - and Why It Is Almost Everything

You were taught that matter comes in three states: solid, liquid, gas. This is accurate as far as it goes, but it misses the state that dominates the universe. When you heat a gas to the point where electrons are stripped free of their parent atoms, the result is an ionized mixture of positive ions and free electrons that move independently and respond to electric and magnetic fields in ways no ordinary gas can. This is plasma. The sun is plasma. Lightning is plasma. The vast tendrils of gas threading between galaxies - plasma. Estimates consistently put the fraction of all visible matter in the universe that exists in plasma form at around 99 percent. The solid world humans inhabit is a thin exception to the rule.

Dusty plasma adds a further layer. When micron-sized solid particles - dust grains - are suspended inside a plasma, they rapidly acquire electrical charge from the surrounding ions and electrons. These charged grains are not passive passengers: they interact with each other and with the plasma medium in a complex electromagnetic dance. Dusty plasma is found in the rings of Saturn, in the tails of comets, in collapsing interstellar clouds where stars are being born, and - critically for industry - inside the chambers used to manufacture semiconductor chips.

The central challenge of dusty plasma physics is modeling the forces between those charged particles. In ordinary Newtonian mechanics, forces between objects are symmetric: if A pushes B with a certain force, B pushes A back with the same force in the opposite direction. Newton's third law. In plasma, this symmetry breaks. Each charged dust grain sits embedded in an ionized medium that is itself disturbed by the grain's presence - creating a directional ion wake downstream from the particle. Because of this wake, the force that one grain exerts on another depends on their relative positions in a way that is not symmetric. The force from A to B is not the same magnitude as the force from B to A. Physicists call these non-reciprocal forces, and accurately describing them has been one of the field's standing difficulties.

Modeling non-reciprocal forces precisely requires knowing how each particle interacts with the plasma medium surrounding it - the ion wake, the local density gradients, the plasma temperature. Conventional theory handles this by making assumptions that simplify the mathematics at the cost of some accuracy. For seven decades, those assumptions have been the best available tool. The Emory team's neural network learned the true behavior directly from observation and found that the simplifications were not quite good enough.

Not a Black Box - A Physics-Informed Mind

The phrase "AI discovers physics" has become common enough to invite skepticism. Most machine learning models are black boxes: they produce predictions no one can fully explain, trained on vast datasets using computational brute force. The Emory team built something deliberately different. Their neural network was physics-tailored - designed from the outset with the known laws of the system embedded in its architecture.

Rather than feeding raw data to a general-purpose network and hoping patterns would emerge, the team engineered the model to account for what was already understood: the effects of gravity, viscous drag from the plasma medium, and the general mathematical form of inter-particle forces. The unknown parts - the precise shape of the non-reciprocal interactions - were what the network was asked to learn. Because the known physics constrained the architecture, the model did not need to rediscover the entire structure of the problem from scratch. It only needed to fill in the details the theory could not supply.

This approach had two consequences. First, it meant the model could learn from a small dataset. Where conventional deep learning requires enormous quantities of training data, the Emory network was trained on a relatively compact but information-rich set of 3D particle trajectories. Second, and more importantly, it meant the result was interpretable. The researchers could look inside the model and understand what it had learned. "Our AI method is not a black box: we understand how and why it works," said Burton. "The framework it provides is also universal. It could potentially be applied to other many-body systems to open new routes to discovery."

The distinction between a physics-tailored model and a conventional one matters for how discoveries are made. Standard deep learning finds correlations. A physics-informed model finds deviations - the gap between what the known theory predicts and what nature actually does. Those deviations are precisely where new knowledge lives. By building in what is known, the Emory team created a tool sensitive enough to detect what is not.

Chips, Fusion, Saturn - and the End of Physics as Usual

Dusty plasma is not an abstract curiosity. It appears in some of the most consequential physical environments on Earth and in the cosmos, and understanding it more accurately has practical stakes across all of them.

In semiconductor manufacturing, dusty plasma is a known hazard. The gases used to etch and deposit material on silicon wafers become ionized during processing, and the resulting plasma can allow dust particles to form, accumulate charge, and drift toward the wafer surface. Engineers call them "killer particles" - charged contaminants that can destroy the circuits being built if they land on the wafer during fabrication. A more accurate model of how these particles move and interact under different plasma conditions could improve process control and yield in ways that ripple through every device manufacturer on the planet, at a time when the world's dependence on advanced chips has never been higher.

In fusion energy research, dusty plasma is an unwanted byproduct of plasma-wall interaction inside reactor vessels. When energetic plasma ions strike the inner wall of a tokamak, material is sputtered off the wall and enters the plasma as charged dust. This dust retains tritium - the radioactive hydrogen isotope used as fusion fuel - creating a safety concern, and can destabilize the plasma in ways that degrade reactor performance. As fusion moves from scientific proof-of-concept toward engineering reality, the ability to predict and manage dust behavior becomes a design requirement, not an academic footnote.

In space, dusty plasma governs phenomena on scales ranging from the rings of Saturn to the birth of solar systems. Planetary rings are vast charged-grain systems whose structure is partly determined by the electromagnetic interactions the Emory model now describes more precisely. Cometary tails are dusty plasma. So are the dense interstellar clouds that collapse under gravity to form new stars, and the protoplanetary disks of gas and charged dust that eventually become planets. Correcting the theoretical assumptions that have underpinned models of these systems for decades could meaningfully improve our understanding of how planets - including this one - came to exist.

But the most significant implication may be the most general. Justin Burton's statement - "We showed that we can use AI to discover new physics" - is not a claim about dusty plasma specifically. It is a claim about what physics-tailored machine learning can do across any complex system where theory has had to rely on simplifying assumptions. For decades, the role of AI in science has been pattern recognition at scale: searching protein sequence space, screening drug candidates, identifying anomalies in telescope data. Powerful applications, all of them. What the Emory study demonstrates is something different in kind - a model that encodes existing physical knowledge and uses the precision of its predictions to locate where that knowledge is incomplete.

If Burton is right that the framework is universal - applicable to other many-body systems, other complex force interactions - then what happened in a plasma chamber at Emory may be the early signal of a broader transformation. The instruments of discovery are no longer only the telescope, the accelerator, and the interferometer. They include, now, the neural network trained precisely enough to see what theory cannot.