A neuromorphic superconducting chip aims to help control fusion plasma in real time

The article, published by Scientific American and written by Peter Hall, describes a new superconducting chip unveiled by Brookhaven National Laboratory. It is designed to help monitor fusion plasma in real time. The lab developed it together with other US Department of Energy facilities and AI hardware start-up Great Sky.

The backdrop is that before fusion can power anything, its machines have to control a plasma that behaves unruly. The chip's makers hope their hardware will be fast enough to track changes in the plasma and help keep them in check.

Chips of this kind are intended for tokamaks. These are large, doughnut-shaped machines that use magnets to hold plasma in place, and they sit at the heart of many fusion power plant proposals. The plasma fuel, at extreme temperatures, is highly temperamental. When it behaves unexpectedly it can trigger a disruption, a sudden collapse in which it loses its energy and which can damage the device. Michael Churchill, a researcher at the Princeton Plasma Physics Laboratory who worked on the project, says disruptions are "existential" for tokamaks. In his view, if they cannot be controlled, there will be no working tokamak.

The data problem is that trouble can arise quickly, yet analyzing plasma live requires a great deal of information, and processing it can overwhelm conventional computers. The article argues that neuromorphic AI systems could keep pace. They mimic the structure of the human brain and process huge volumes of data in parallel. Great Sky builds its version with superconducting circuits, a cousin of the hardware used in some quantum computers. The aim is to run its specialized AI directly on the chip, at the speed the task demands.

The test was carried out without connecting the chip to an operating tokamak. The team replayed data recorded at the DIII-D tokamak in San Diego (California) at its original rate of one million samples per second. The chip kept up while running algorithms designed to flag anomalies and predict disruptions. The chip's AI model is far smaller than, for example, today's large language models, and the team is still writing up the results for publication.

Churchill acknowledges that all of this is "very much in the demonstration phase." It is unclear whether the chip will make fusion more reliable and safer or remain a show of strength. The article notes that tokamaks already operate without such chips. It also points out that their control mechanisms cannot react to plasma problems as quickly as this chip is supposed to be able to detect them.

Some voices are more cautious about the application's appeal. Carlos Paz-Soldan, an associate professor of applied physics at Columbia University, says he is more excited about the use of AI in materials discovery.

Shinjae Yoo, who heads Brookhaven's AI department and led the project, likewise does not tie the chip's future solely to fusion. He wants to test it where a traditional data center won't fit. He sees an immediate benefit in edge computing, because it operates at gigahertz speeds and already has built-in memory.

The next steps are to build a successor chip with a neural network ten times larger. The team plans to unveil it at Brookhaven early next year and install it at DIII-D some time after that. Great Sky's hardware could then finally face a real fusion plasma, without a hard drive in between.

In short, this is an early demonstration whose results are still unpublished. It points to an interesting path for applying superconducting AI to a control problem that is extremely demanding in terms of speed, but its practical usefulness for fusion remains unconfirmed.