In Simple Terms
Researchers have built a computer chip designed to spot warning signs inside fusion machines faster. It kept up with previously recorded measurements, but it has not yet been tested inside a working reactor. The chip is still experimental, and it is too soon to know whether it can make fusion reactors safer.
Why Fusion Reactors Need Faster Warnings
In a fusion reactor, a small problem in the plasma can suddenly grow into a disruption that threatens stable operation. Simply collecting measurements is not enough: they must be analyzed as quickly as conditions inside the machine change. That challenge inspired a new computer chip developed by U.S. national laboratories in collaboration with a startup, in an effort to bring artificial intelligence closer to the source of the data and help it detect warning signs sooner.
The research is focused on machines called tokamaks. These large reactors have ring-shaped chambers, like doughnuts, and use magnets to confine plasma—the fuel that fusion facilities aim to use. But confining the plasma does not make its behavior stable or easy to predict. It keeps changing, and it can quickly drift away from the desired state.
If a disruption occurs, the plasma can suddenly collapse and lose its energy, potentially damaging the machine. Disruptions are therefore more than brief glitches that can be ignored; they are among the challenges that will determine whether tokamaks can operate practically. Michael Churchill, a research scientist at the Princeton Plasma Physics Laboratory who worked on the project, has said that the ability to manage these events is essential if such reactors are to keep running.
Keeping track of the plasma means collecting a large, rapidly arriving stream of measurements and analyzing them while conditions are still changing. That is where conventional computing can struggle: the data may arrive so quickly and in such large volumes that analysis lags behind, even if a computer can eventually process everything. Timing matters because spotting a warning sign after events have already advanced is not as useful as catching it early.
Computing Closer to the Measurement
The new chip is based on an approach called neuromorphic computing. Its processing is designed to take advantage of many operations happening in parallel, drawing inspiration from some aspects of how the brain works. The aim is not to build an artificial mind that imitates every detail of a human brain. Instead, the approach uses a computing design suited to rapidly handling large volumes of signals and looking for patterns that could signal trouble.
The chip uses superconducting circuits, a technology also associated with some types of quantum-computing hardware. The team’s idea is to run specialized artificial-intelligence algorithms directly on the chip, rather than first sending the data to a distant computer or a conventional computing system. Processing the measurements closer to where they originate is intended to reduce the time between a signal arriving and its analysis.
That matters because the chip is meant to analyze information as it comes in, rather than wait until an experiment’s recording is complete. Xinghai Yu, head of artificial intelligence at Brookhaven National Laboratory and the project’s leader, says that having memory built into the chip, along with a speed reaching the gigahertz range, could make it useful for edge computing. That means processing information at or near its source instead of relying on a large data center.
Testing Recorded Data from the DIII-D Tokamak
The chip has not yet been connected to an operating tokamak, so it has not faced live plasma inside a reactor. Instead, the team used measurements previously recorded at the DIII-D tokamak in San Diego, California. The researchers replayed the data through the chip at the original collection rate: one million samples per second.
During the replay, the chip was able to keep pace with the stream of measurements. The team ran algorithms on it designed to detect unusual conditions and predict disruptions. The important result at this stage is that the hardware could process data at the rate at which it had been recorded—not that it stopped an actual disruption or improved a reactor’s performance while it was operating.
This distinction is essential to understanding what the team has and has not achieved. Replaying an earlier recording tests whether the chip can receive and analyze data quickly, but it does not test everything that might happen when the system is connected to a live facility. In that setting, handling real-time measurements and responding to them would also be part of the experiment. And showing that an algorithm can issue a warning or make a prediction does not, by itself, prove that a control system can turn that information into the right action in time.
Churchill describes the work as still being at the demonstration stage. The team’s results had also not yet been published; researchers were preparing them for publication. The project should therefore be viewed as an early test of a computing platform’s potential, not as a technology ready for use in fusion reactors.
Detecting a Threat Is Not the Same as Controlling It
There is a difference between a chip quickly detecting signs of a possible disruption and preventing that disruption from damaging the machine. The first involves sensing, analysis, and prediction. The second requires an appropriate response that can be carried out quickly enough to affect the plasma’s behavior. The test described for this project examined how the chip processed replayed data; it has not yet demonstrated that the full control process works.
According to the team, tokamaks already operate without this kind of chip. But existing control systems may not be able to respond to a problem as quickly as the new chip is expected to detect it. That leaves an open question: could faster detection narrow the gap between a threat appearing and a response being made, or would it improve analysis without changing the outcome of reactor operations? So far, there is no conclusive experimental answer.
The step from a model working on stored recordings to a test connected to a reactor is scientifically important. In a live operation, the team would need to verify that the chip handles the incoming measurements as intended and that its outputs are useful in an actual control setting. The current test does not establish that disruptions will become less frequent or less damaging, or that the chip alone will make fusion operation safer.
A Project with Uses Beyond Fusion Reactors
Yu does not see the technology’s future as limited to fusion. Fast computation combined with built-in memory could be useful in situations where a conventional data center is not nearby or is not suitable. That is the idea behind edge computing: devices close to the source of information handle some of the analysis, rather than sending everything to centralized computing infrastructure and waiting for it to be processed there.
That does not mean the chip has proved effective in all of those possible applications. These are areas Yu hopes to explore, while fusion remains the field in which the team tested algorithms for detecting unusual conditions and predicting disruptions using DIII-D data. It is important to distinguish potential future uses from what the experiment has actually demonstrated.
Even for fusion, the artificial-intelligence model running on the chip is much smaller than today’s familiar large language models. That is to be expected because the tasks are different: this system needs to analyze specific signals, not generate general-purpose text or answer open-ended questions. The model’s small size alone says nothing about whether the project is strong or weak. What matters is whether it can do its particular job within the required time, and whether it proves useful when integrated into a real operating environment.
Next Step: A Neural Network Ten Times Larger
The team plans to build a next-generation chip with a neural network ten times larger than the current model. Under the announced plan, the chip was expected to be presented at Brookhaven early the following year and then installed at DIII-D at a later date. The project description did not give a more precise calendar date, so this timing reflects the team’s plan, not a guaranteed completion date.
If those steps go ahead, installing the next-generation chip at the facility would be a different test from replaying recorded data. The technology could then move from showing that it can keep pace with an earlier recording to handling a stream arriving from an actual reactor, without the data first passing through a storage drive. Even reaching that stage, however, would not by itself prove that the chip can protect a tokamak. The team would still need to measure how accurate and practically useful its warnings are, and how well they connect to the reactor’s control system.
A larger network could be useful if it enables the chip to analyze more complex patterns, but that is a possibility, not an established result. The bigger model will need to be evaluated, and increasing the network’s size alone does not guarantee faster responses or better predictions. Future experiments will have to show whether combining superconducting hardware with specialized algorithms offers a tangible benefit beyond successfully processing recorded data.
The Bottom Line
The chip developed by Brookhaven and Great Sky is an attempt to address a central challenge for tokamaks: detecting changes in the plasma quickly enough to respond before a disruption worsens. In tests using recorded DIII-D data at one million samples per second, the chip kept up with a fast stream of measurements and ran algorithms designed to detect unusual conditions and predict risks. But it has not yet been tested on live plasma, and it has not been shown to prevent disruptions or make reactors safer. A planned next-generation chip with a network ten times larger could bring the technology closer to a real-world test. Whether it matters in practice will depend on whether faster analysis can be turned into useful control.