Google’s DeepMind AI can now control superheated plasma in a nuclear fusion reactor
An artificial intelligence (AI) might help us tap into the limitless energy potential of nuclear fusion.
Google-owned DeepMind, the U.K.-based company building AI to take on some of the world's most complex science problems, has trained a deep reinforcement learning algorithm to control the burning plasma inside a nuclear fusion reactor, a report from MIT News reveals.
An AI built to tame the tokamak
In collaboration with the Swiss Plasma Center at EPFL, DeepMind was able to apply its machine learning knowledge to taming a tokamak — a round nuclear fusion reactor that could one day allow us to harness the energy tapped by the Sun and the stars.
The team, which outlined its findings in a paper in the journal Nature, said its new breakthrough could provide physicists with a better understanding of how fusion works. "This is one of the most challenging applications of reinforcement learning to a real-world system," said Martin Riedmiller, a researcher at DeepMind.
Nuclear fusion occurs when two atoms smash together to form a heavier nucleus, a process that releases a massive amount of energy in the form of plasma. Inside stars, this plasma is held together by gravity. Here on Earth, scientists must rely on powerful lasers and magnets, such as one being developed by MIT and the Bill Gates-backed Commonwealth Fusion Systems.
90 measurements monitored ten thousand times per second
In a tokamak reactor, controlling this plasma requires constant monitoring of the magnetic field. The DeepMind team was able to train their reinforcement-learning algorithm to control plasma in a computer simulation. After the AI successfully controlled the virtual plasma, it was then allowed to control the magnets in the Variable Configuration Tokamak (TCV), an experimental reactor in Lausanne, Switzerland. The AI was able to control the plasma for a total of two seconds, which is the total amount of time the TCV reactor can run before it overheats.
The AI closely monitored the plasma by taking in 90 different measurements ten thousand times a second. It then adjusted the voltage accordingly for the reactor's 19 magnets. The researchers said this type of AI could allow them to tightly control future tokamak experiments, meaning they will be able to experiment with a greater number of conditions. It sounds like something out of science fiction: An artificial intelligence may allow us to finally harness the same energy as the stars and the Sun.
Nuclear fusion using magnetic confinement, in particular in the tokamak configuration, is a promising path towards sustainable energy. A core challenge is to shape and maintain a high-temperature plasma within the tokamak vessel. This requires high-dimensional, high-frequency, closed-loop control using magnetic actuator coils, further complicated by the diverse requirements across a wide range of plasma configurations. In this work, we introduce a previously undescribed architecture for tokamak magnetic controller design that autonomously learns to command the full set of control coils. This architecture meets control objectives specified at a high level, at the same time satisfying physical and operational constraints. This approach has unprecedented flexibility and generality in problem specification and yields a notable reduction in design effort to produce new plasma configurations. We successfully produce and control a diverse set of plasma configurations on the Tokamak à Configuration Variable1,2, including elongated, conventional shapes, as well as advanced configurations, such as negative triangularity and ‘snowflake’ configurations. Our approach achieves accurate tracking of the location, current and shape for these configurations. We also demonstrate sustained ‘droplets’ on TCV, in which two separate plasmas are maintained simultaneously within the vessel. 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.
There are going to be plenty of crazy costumes at that mega party this Halloween, but what about some cocktails to brighten the fun?