(ORDO NEWS) — It turned out to be quite difficult to keep the plasma in a stable state in the reactor. Many factors lead to the instability of the plasma column and its extinction.
Modern systems for maintaining plasma stability do not have time to respond to everything, and this work was entrusted to AI.
Many unsolved problems lie in the way of commercial fusion reactors, although the first reactor with a positive energy yield promises to be operational in three years (project ITER).
How artificial intelligence was taught to control plasma
Researchers at DeepMind (bought by Alphabet in 2014), along with scientists at the Swiss Federal Institute of Technology in Lausanne (EPFL), were able to teach artificial intelligence how to control plasma inside a real fusion reactor.
Interesting fact! In the past, DeepMind has made impressive strides in building learning-capable platforms by teaching them how to code, play chess, Go, and StarCraft II, as well as solving a half-century-old problem in biology by teaching AI to envision the spatial shapes of proteins.
A new task for AI DeepMind was the task of controlling the shape of the plasma in a tokamak-type fusion reactor. And he solved it.
How it all goes
In modern tokamaks and on the experimental Swiss tokamak TCV at the center of the EPFL (variable configuration tokamak), the parameters of the magnetic field around the working chamber of the reactor are set by several programmable controllers.
The controllers control electromagnets, the field of which keeps the plasma cord with a temperature of tens to hundreds and more millions of degrees Celsius from touching the inner walls of the working chamber. Thus, the walls of the reactor are protected from destruction, and the plasma becomes more stable.
In the Swiss tokamak, separate controllers set the plasma flow, its profile, and the vertical and horizontal positions of the flow. The work of DeepMind and Swiss scientists came down to the development of a single controller capable of learning and controlled by a neural network.
First, neural networks showed the reactions of plasma to a series of combinations of operating parameters of electromagnets, then they taught how to control plasma on a software simulator.
After that, the neural network was connected to the reactor through a single controller. Artificial intelligence, as practice has shown, was able to independently hold the plasma bundle in the given configurations.
– The developers claim that further AI can independently and much faster than a person find stable parameters for holding plasma, reacting to changing conditions faster than a live operator.
– Experiments with a tokamak in controlling a neural network can accelerate the emergence of commercial solutions in the field of generating clean and almost infinite fusion energy.
Contact us: [email protected]