(ORDO NEWS) — A collaboration between the Swiss Plasma Center and DeepMind has developed a new method for magnetically controlling plasma in a tokamak.
The deep reinforcement learning algorithm developed by DeepMind can significantly speed up the selection of tokamak settings to create pre-specified plasma configurations with high accuracy.
A tokamak, or a toroidal chamber with magnetic coils, is a toroidal device (bagel or donut) in which conditions are created for controlled thermonuclear fusion to take place – the same reactions that take place in the depths of stars.
To this end, powerful magnetic fields are generated in tokamaks and a vacuum is created to contain the high-temperature plasma and protect the walls of the facility from melting. Theoretically, the energy released in this process can be used to generate electricity.
The Swiss Plasma Center (SPC) of the Federal Polytechnic School of Lausanne (EPFL) has many years of experience in the field of plasma physics and control methods.
Not only is the SPC one of the few research centers in the world with a working tokamak, but their installation is also very difficult. Their tokamak allows for various plasma configurations given by the position of the magnetic coils, hence the name variable configuration tokamak (TCV).
The configuration of the plasma is related to its shape and position in the tokamak, and this determines the stability of the plasma and the productivity of the reactor, that is, the amount of generated energy. Before experimenting with their facility, the SPC researchers first test the control system configurations on a simulator.
“Our simulator is based on over 20 years of research and is constantly being updated,” explains Federico Felici , SPC staff member and co-author of the study. “Even so, it still takes time-consuming calculations to determine the correct value for each variable in the control system. This is where our joint research project with DeepMind comes in.”
DeepMind is a UK-based scientific discovery and AI company acquired by Google in 2014 that aims to “solve the problems of artificial intelligence to advance science and humanity.” DeepMind experts developed a deep reinforcement learning (DRL) algorithm that can create and maintain certain plasma configurations and trained it on the SPC simulator.
Initially, the algorithm tested many different plasma control strategies in the simulation to gain experience. Moreover, the training took place in two directions: first, the algorithm was given a number of settings to control the installation, according to which plasma was generated on the simulator, and the algorithm analyzed its configuration; then, the correct settings were determined by the algorithm from the plasma configuration.
After training, the DRL-based system was able to create and maintain a wide range of plasma shapes and advanced configurations in the simulator, including one in which two separate plasma fragments are simultaneously maintained in the reactor.
Finally, the research team tested their new system directly on the tokamak to see how it would perform under real-world conditions. As expected, all the configurations created by the DRL algorithm and predicted by the SPC simulator were obtained on a real installation.
Thus, the new approach to controlling the tokamak magnetic coils not only makes it possible to accelerate the creation of the necessary plasma configurations, but also provides accurate tracking of the location, current, and shape for these configurations.
Martin Riedmiller , Head of Control at DeepMind and co-author of the study, said, “Our team’s mission is to explore next-generation artificial intelligence systems—feedback controllers—that can be trained from scratch in complex dynamic environments. The control of fusion plasma in real facilities offers fantastic, albeit extremely complex, possibilities.”
Contact us: [email protected]