Ctg sit23 hub banner.jpg

Google's DeepMind trained an AI to control nuclear fusion

DeepMind trained an AI to control the plasma inside a tokamak reactor

Image:

DeepMind trained an AI to control the plasma inside a tokamak reactor

Fusing hydrogen atoms is safer and more efficient than nuclear fission, but also much more complex

Google-backed DeepMind has trained a machine learning algorithm to control the hot plasma inside a tokamak nuclear fusion reactor.

It might sound like the start of an '80s blockbuster, but the system could give scientists a better understand of how fusion works, and how the matter inside a reactor interacts under various conditions.

Researchers believe the breakthrough could speed the development of an endless supply of renewable energy.

For scientists, nuclear fusion remains a promising potential energy source.

Fusion energy is far cleaner and safer than fossil fuels or traditional nuclear power, which is generated by fission - the splitting of nuclei.

But in nuclear fusion, the hydrogen atoms' atomic nuclei are driven together to produce heavier elements like helium, rather than being split apart. The process generates a large amount of energy, making fusion a particularly efficient power source.

However, controlling nuclear fusion is hard. Because atomic nuclei repel each other, smashing them together inside a reactor requires extremely high temperatures, often reaching hundreds of millions of degrees. At these temperatures matter transforms into plasma, which is a swirling, superheated soup of particles. At this point, it is a struggle to keep a reactor's plasma together long enough to harvest energy from it. Researchers employ a number of techniques to do so, including magnets and lasers.

In a magnet-based reactor, called tokamak, the plasma is contained inside an electromagnetic cage, forcing it to maintain its form and preventing it from reaching the reactor walls.

DeepMind has been working on controlling plasma inside a tokamak, for which it worked with the Swiss Plasma Center (SPC) at École Polytechnique Fédérale de Lausanne (EPFL).

The SPC's tokamak is known as a variable-condition tokamak (TCV). It differs from other tokamaks in allowing for a wide range of plasma configurations. However, trying a new configuration - in the hunt for more power or cleaner plasma - is expensive and time-consuming, requiring a huge amount of engineering and design work.

"Our simulator is based on more than 20 years of research and is updated continuously," said Federico Felici, an SPC scientist.

"But even so, lengthy calculations are still needed to determine the right value for each variable in the control system. That's where our joint research project with DeepMind comes in."

DeepMind created an AI algorithm that was trained on the SPC's simulator by putting it through a variety of control scenarios.

From the expertise it gained in the simulations, the system was eventually able to calculate control strategies for producing desired plasma configurations, by changing the settings on each of the 19 coils that control the shape of the plasma. These configurations included a D-shaped cross section - similar to what will be used inside ITER, the large-scale experimental tokamak being built in France - and a snowflake shape that could be better at dissipating heat.

The programme was also tested on a real-world tokamak, where it was able to create and control a variety of plasma shapes.

"While there is still much work to be done … we are pleased that the results indicate the power of AI to accelerate and assist fusion science, most likely augmenting human expertise in the field and serving as a tool to discover new and creative approaches for [fusion reactor control] and beyond," said Martin Riedmiller, a research scientist at DeepMind.

"[The work] also suggests that there might be potential for wider adoption of deep reinforcement learning on physical systems for complex scientific and industrial machines, from simple motor control to complex robots," he added.

Fasoli said SPC is "always open to innovative win-win collaborations" that enable the organisation to explore new perspectives, thereby "speeding the pace of technological development".

The research was published in the journal Nature.

Google's DeepMind trained an AI to control nuclear fusion

Fusing hydrogen atoms is safer and more efficient than nuclear fission, but also much more complex

Public Sector

National Grid is turning analogue to digital - Ctrl Alt Lead podcast

Green

AI to blame for Google's rocketing greenhouse gas emissions

Artificial Intelligence

IT Essentials: The AIrony