The superconducting neural network chip that can be used in fusion reactors. (Image Credit: Great Sky)
Rapidly analyzing sensor data is essential for controlling a fusion reactor’s unstable plasma. Researchers at Brookhaven National Laboratory have developed a superconducting AI chip that processes fusion signals at one million samples per second. This may eventually lead to quick plasma monitoring and control.
“This is a powerful, new platform for implementing many of the SI algorithms our team at Brookhaven Lab is developing,” said Shinjae Yoo, CDS AI department chair and the project’s principal investigator. “There are so many ways to apply this to advance the scientific missions we care about within the DOE complex."
These chips don’t use silicon transistors for handling digital information. Instead, they rely on superconducting circuitry to integrate analog neural networks. Great Sky, one of the developers behind the chip, also says it uses optical communication. Other partners include Princeton Physics Plasma Laboratory (PPPL), and the DIII-D National Fusion Facility. Although Brookhaven National Laboratory, which developed the algorithms, led the project, Great Sky manufactured the chip.
The team used three tasks, which control tokamaks, to assess the superconducting AI chip. As part of the first two demonstrations, the researchers provided the chip with measurements from the DIII-D tokamak. The algorithms then found data anomalies and predicted plasma interferences. Those fusion-plasma measurements were collected at one million samples per second by the chip, and didn’t fall behind.
Close-up image of the wafers. (Image Credit: Great Sky)
The third test involved simulated measurements produced by TORAX, a tokamak simulator that Google DeepMind’s Fusion team developed. During this demonstration, the chip analyzed signals to estimate sawtooth instabilities. Each test revealed that superconducting neural-network chips are capable of performing several types of analysis for fusion research. With those functions, it could support real-time plasma condition estimates. This can help systems respond to rapidly changing behavior.
While the team is working on a paper to detail the AI chip, DIII-D researchers will determine the best way to integrate the hardware into future experiments. In the future, this chip is expected to have a neural network 10x larger than the current one. By early 2027, the team is planning to unveil it at Brookhaven National Laboratory. The chip is being deployed at DIII-D later that year.
Apart from fusion, the team says the chip could be used for high-speed X-ray analysis, quantum-system control and error correction, cybersecurity, and video processing. They still need to run more tests to see how the technology performs with larger neural networks.
Have a story tip? Message me here at element14.