Scientists have developed an artificial intelligence program tasked with controlling superheated plasma inside nuclear fusion reactors.
Fusion reactors, sometimes referred to as artificial suns, are machines capable of fusing atoms together under extreme conditions. When two light atomic nuclei are forced together, they create a single, heavier one and release energy as a byproduct.
Theoretically, this energy can be used to generate further nuclear reactions while also being turned into electricity that humans can use. In this way, nuclear fusion is a much-researched area of physics since it offers a clean and almost unlimited source of power.
However, scientists have yet to create a nuclear fusion reactor that's capable of producing more energy than it needs to keep working, and this has been a goal for decades. But they're getting closer.
Enter tokamaks—donut-shaped devices that are designed for nuclear fusion research. They work by producing a superheated, charged gas called plasma in which fusion can take place. In order to keep this plasma contained as it flows around the machine, tokamaks use powerful magnets that keep it in the correct shape at hundreds of millions of degrees—hotter than the core of the sun.
Lengthy calculations are needed in order to find out what sort of shape the plasma should take and where in the donut-shaped machine it should be contained. To this end, scientists at Google's DeepMind AI team think they can help.
In a new study published in the journal Nature on Wednesday, DeepMind researchers in collaboration with scientists at Switzerland's Swiss Plasma Center (SPC) at the EPFL research institution outline a method known as reinforcement learning, which makes use of an artificial intelligence program that can create and maintain specific plasma configurations.
Different Simulations
They put the program through its paces by subjecting it to many different plasma simulations so it could gain some experience and automatically come up with a strategy to produce desired configurations. Eventually it was able to work with a range of different shapes and was even able to control two separate plasmas at once.
According to the study, the scientists then tested the AI on a real-life tokamak—the EPFL's TCV machine—to see if it could achieve tasks like move the plasma by small amounts, increase or decrease its current, and even operate multiple plasma bands at the same time. They found it could do so successfully.
"The collaboration with the SPC pushes us to improve our reinforcement learning algorithms, and as a result can accelerate research on fusing plasmas," Brendan Tracey, a senior research engineer at DeepMind and co-author of the study, said in an EPFL press release.
