Deepmind, perhaps best known for their Alphago AI, which defeated all the grandmasters of Go, has trained an AI to control and sculpt heated plasma in a fusion reactor. The researchers used reinforcement learning to teach the AI to adjust the magnets that hold the plasma in a Tokamak reactor.
The plasma in such a reactor reaches many millions of degrees, so it must be handled with magnets so as not to burn through the reactor's housing. The problem is that the magnets have to be adjusted very quickly and often to maintain plasma stability.
Deepminds AI can make thousands of adjustments per second. Together with researchers from the École polytechnique fédérale de Lausanne, EPFL, Deepmind has tested the AI in an experimental facility. The AI maintained the plasma for two seconds, which is as long as that particular reactor can be run before it overheats.
In addition to controlling the plasma, the AI was able to shape the plasma in different ways, increasing safety and efficiency as it is possible to adapt the plasma as needed. Just like in the case of Alphago, this AI found new ways of working that we humans have not come up with before.
"The AI created the plasma we expected, but it used the magnetic cores in a completely different way. So those who looked at the results of the experiment on how the magnetic fields changed were quite surprised", says Federico Felici, researcher at EPFL, in a comment to New Scientist.
The researchers will now do more experiments to see how the AI can be developed and what methods it can come up with.