DeepMind Develops AI to Control Nuclear Fusion Plasma
Courtesy of pxhere
DeepMind, a British artificial intelligence company, announced on February 16 that it had successfully taught an AI to control nuclear fusion in a joint research project with the Swiss Plasma Center. The company aims to use this research to bolster nuclear fusion’s candidacy as a global source of clean energy.
“We predict reinforcement learning will be a transformative technology for industrial and scientific control applications in the years to come, with applications ranging from energy efficiency to personalised medicine,” the company wrote in a blog post.
Today in @nature, with @EPFL, the first deep reinforcement learning system that can keep nuclear fusion plasma stable inside its tokamaks, opening new avenues to advance nuclear fusion research.
— DeepMind (@DeepMind) February 16, 2022
Paper: https://t.co/UdIIfNgaap pic.twitter.com/gH2nNc71nK
DeepMind’s AI can manipulate plasma inside a tokamak, a machine that contains plasma within a doughnut-shaped vacuum surrounded by magnetic coils. By sculpting the plasma into different shapes, the AI ensures that the plasma does not touch the inner walls of the tokamak, a process which involves coordinating the machine’s magnetic coils and adjusting their voltage thousands of times per second.
“This approach has unprecedented flexibility and generality in problem specification and yields a notable reduction in design effort to produce new plasma configurations,” the researchers’ paper, published in Nature, reads.
A blogpost with a more readable overview of the fusion work. And pretty videos of what shaping plasma actually looks like! On the left, you can see a camera looking into the tokamak, while a single NN controller goes through the different phases of a shot.https://t.co/ymK3jrxVQA pic.twitter.com/jvaJoPoAPB
— 317070 (@317070) February 17, 2022
These various plasma configurations enable the optimization of the nuclear reaction’s stability, confinement, and energy exhaust. The AI was capable of producing elongated shapes, a negative triangularity, and a snowflake shape. It was even able to form sustained plasma droplets, in which two separate plasmas exist simultaneously within the machine.
“This represents a notable advance for tokamak feedback control, showing the potential of reinforcement learning to accelerate research in the fusion domain, and is one of the most challenging real-world systems to which reinforcement learning has been applied,” the paper reads.
This tokamak reactor diagram (Courtesy of Wikimedia) shows the internal chamber where the plasma is controlled as well as the coils that encase it.
“Deep reinforcement learning is pretty great at working with sci-fi things where human intuitions tend to break down,” said Jonas Degrave, one of the DeepMind researchers.
