Fusion reactors hold the promise of abundant energy which is relatively clean, but so far it has been easier said than done. The tokamak - a doughnut-shaped vessel designed to contain nuclear fusion rection-contains a special kind of chaos. Hydrogen atoms which are highly unstable, are smashed together at extremely high temperatures to create plasma that is hotter than the sun's core.

Nuclear fusions are the powerhouses of the stars. However, the sheer gravitational forces of these starts are enough to pull hydrogen atoms together and keep them together for the process. On Earth, scientists use powerful magnetic confinement with the help of magnetic coils to confine the fusion.

However, researchers at the Alphabet-owned British firm DeepMind, along with the scientists at the Swiss Federal Institute of Technology in Lausanne (EPFL), have made it possible theoretically to extract power from plasma hotter than the surface of the sun. Together with the Swiss Plasma Center, they have developed an Artificial Intelligence (AI) for controlling a nuclear fusion reaction. They've created a neural network that has the ability to contain the magnetic fields within the Tokamak fusion reactor.

Before this invention, the researchers in Switzerland used different algorithms for all 19 magnetic coils to monitor each second of the interiors of the reactors. However, DeepMind's invention replaces these with one single neural network which gradually learns to supply voltage to different coils to help them contain plasma.

A paper published in the journal Nature describes how researchers from the two groups taught a deep reinforcement learning system to control the 19 magnetic coils inside TCV, the variable-configuration tokamak at the Swiss Plasma Center.   

The algorithm was first trained on highly accurate digital stimulation of the reactor. It observed the impact of the coils on the shape of the plasma. Additionally, it was given different shapes which are used inside ITER (formerly the International Thermonuclear Experimental Reactor) and a snowflake configuration that helps distribute the fusion heat evenly around the Tokamak. Then the algorithm experimented on a real tokamak. In the experiment, it was able to contain the plasma for around 2 seconds. 

The AI was further able to shape the plasma and move it around within the reactor. Interestingly, the AI even demonstrated that it could control two separate beams of plasma at once.

“This AI algorithm, the reinforcement learning, chose to use the TCV coils in a completely different way, which still more or less generates the same magnetic field,” says Federico Felici at EPFL. “So it was still creating the same plasma as we had expected, but it just used the magnetic cores in a completely different way because it had complete freedom to explore the whole operational space. So people were looking at these experimental results about how the coil currents evolve and they were pretty surprised.”

“Sometimes algorithms which are good at these discrete problems struggle with such continuous problems,” says Jonas Buchli, a research scientist at DeepMind. “This was a really big step forward for our algorithm because we could show that this is doable. And we think this is definitely a very, very complex problem to be solved. It is a different kind of complexity than what you have in games.”