DeepMind develops 'AI that can control nuclear fusion'


Eye Steel Film

With the construction of the world's largest fusion reactor ' ITER ' progressing, fusion is highly expected as a future energy source. The tokamak fusion reactor represented by ITER has a mechanism to confine ultra-high temperature plasma with strong magnetism, and it is necessary to adjust the magnetism of the fusion reactor flexibly. DeepMind , a sister company of Google that develops artificial intelligence (AI), announced that it has succeeded in magnetically controlling a tokamak fusion reactor with trained AI.

Magnetic control of tokamak plasmas through deep reinforcement learning | Nature

Latest success from Google's AI group: Controlling a fusion reactor | Ars Technica

The bottleneck of the tokamak fusion reactor is the development of its control software. The control software must be designed for each furnace, and it is also required to 'read the ever-changing ultra-high temperature plasma with a sensor, model the state, control and maintain it'. Therefore, the versatility of the control software is very low, and if you want to change the shape of the plasma in the fusion reactor significantly, you may need to make major modifications.


Steve Jurvetson

Therefore, the application of AI technology was expected for the development of more versatile control software. By letting AI learn how to control fusion reactors, researchers can study with different plasma configurations without having to modify the control software every time. So DeepMind's team embarked on the development of fusion control AI at the Swiss Plasma Center at the Swiss Federal Institute of Technology Lausanne.

However, it is very dangerous to suddenly train in a real fusion reactor, so first use a simulator that reproduces the fusion reactor of the Swiss Plasma Center, and use the simulator to generate plasma without producing inaccurate results on the simulator. Trained AI to control. In order to learn the control method efficiently, an algorithm called 'Critic' that receives a reward when changing the magnetism of the fusion reactor and outputting the expected result has been incorporated into AI.

After the training was over, we first brought the actual fusion reactor to a high energy state with normal control software and entrusted AI to control the reactor. Then, AI was able to work almost as expected. In one experiment, after raising the energy to stabilize the plasma, we succeeded in changing the shape of the plasma and lowering the energy. In addition, an experiment was conducted to maintain two plasmas with different structures at the same time in one furnace.

by DJ Quietstorm

The research team proposes a software-first research scheme that 'realizes a fusion reactor that can build a desired plasma by AI' instead of developing software based on the fusion reactor as in the past. I am. The research team also says that it can be expected to optimize the performance of fusion reactors that have already been developed.

in Science, Posted by log1i_yk