Brain-inspired chip keeps pace with fusion plasma in lab tests

Inside a fusion reactor, plasma, the superheated gas that fusion depends on, is held in place by magnetic fields, and its conditions can shift quickly. On a new superconducting chip, data from a real tokamak has just gone through at full speed: about one million samples per second, with the chip keeping up.

The bottleneck is computing. Real-time monitoring requires machines that can digest enormous volumes of data almost instantly, and conventional silicon-based systems struggle to do that fast enough. The chip was designed by researchers at Brookhaven National Laboratory, a US Department of Energy lab, with the Princeton Plasma Physics Laboratory, the DIII-D National Fusion Facility and Colorado company Great Sky. Its superconducting circuits are arranged to mimic the connections found in the human brain, which is why the team describes it as brain-inspired.

The tests were run on data from the DIII-D tokamak, a fusion research facility in California run by General Atomics. Two of them asked the chip to spot unusual plasma activity and predict disruptions. The team reports that the chip detected unusual behavior and predicted instabilities in lab tests. A third experiment used simulated data from TORAX, a tokamak core transport simulator developed by Google DeepMind's fusion team, in which the chip predicted the onset of sawtooth instabilities, rapid changes in plasma conditions. The results suggest that superconducting neural networks could eventually support real-time plasma monitoring and stabilization.

What changes is the speed of the control loop. For operators of future fusion reactors, a faster and more reliable way to watch plasma and flag trouble before it grows would be a practical gain. The timeline: a chip with a neural network 10 times larger is planned, to be presented at Brookhaven in early 2027, then tested at DIII-D during fusion experiments later that year. The same kind of chip could, according to Great Sky, find uses in quantum computing, cybersecurity and high-speed X-ray data analysis, but no commercial date has been announced.

“This is just the beginning,” said Jeff Shainline, Great Sky's founder and CEO. For now, the chip is a laboratory prototype, and the question is whether its results hold up when the larger version meets a real reactor's experiments.