AI harnessed to improve fusion plasma control

Image: EPFL

Google’s DeepMind has developed a deep learning system to advance the development of nuclear fusion in a tokamak.

The initiative, conducted in association with the Swiss Plasma Centre at the Swiss Federal Institute of Technology Lausanne (EPFL), was aimed to control the powerful magnetic coils that surround a tokamak and that in turn control the shape of the hot plasma therein where the fusion takes place.

Such control of the plasma is necessary to keep it away from the walls of the tokamak vessel, which would otherwise result in heat loss and deterioration of the plasma and the switching off of the fusion process.

A tokamak is a doughnut shaped vessel in which the plasma is contained with the control system required to constantly adjust the voltages on the coils to maintain its shape.

Have you read?
UK’s JET facility achieves fusion energy record
Italian nuclear fusion challenge is on

The variable configuration tokamak at the Swiss Plasma Centre is unique in allowing for a variety of plasma configurations and hence enabling the investigation of new approaches for confining and controlling plasma shapes and positions.

DeepMind and the Swiss Plasma Centre have now demonstrated that a deep reinforcement learning algorithm that was ‘trained’ in a simulator environment has produced controllers that can both keep the plasma steady and be used to accurately sculpt it into different shapes.

This ‘plasma sculpting’ shows that the reinforcement learning – a form of sequentially-based outcome machine learning – has successfully controlled the superheated matter and allows investigation of how the plasma reacts under different conditions, according to the researchers, whose work has been published in the journal Nature.

Plasma sculpting

Existing plasma control systems are complex, with separate controllers required for each of the variable configuration tokamak‘s nineteen magnetic coils. In contrast, the deep learning architecture uses a single neural network to control all of the coils at once, automatically learning which voltages are the best to achieve a plasma configuration directly from sensors.

With these, a range of plasma shapes have been created, from droplets of two plasmas and snowflakes to triangles and ovals, which are now being studied by plasma physicists for their usefulness in generating energy.

For example, the snowflake shape with many ‘legs’ could help reduce the cost of cooling by spreading the exhaust energy to different contact points on the vessel walls.

The DeepMind researchers consider that the capability of autonomously creating controllers also could be used to design new kinds of tokamaks while simultaneously designing their controllers.

No posts to display