DeepMind A.I. helps control nuclear fusion reaction, potentially producing more energy

New research shows that artificial intelligence can be used to more precisely control a nuclear fusion reaction, potentially helping accelerate the development of nuclear fusion as a practical power source.

The A.I. was developed by computer scientists at DeepMind, the London-based A.I. research company that is part of Alphabet, and physicists from the Swiss Plasma Center at EPFL in Ecublens, Switzerland. The breakthrough research was published in the peer-reviewed scientific journal Nature on Wednesday.

The most promising path toward fusion power involves a doughnut-shaped reactor, called a tokamak, in which hydrogen is superheated into a state called plasma. This happens at temperatures of more than 100 million degrees Celsius (180 million degrees Fahrenheit). At these temperatures, the nuclei of hydrogen atoms can be fused, releasing a huge amount of energy.

But plasma is too hot to be contained by any material, so the plasma is suspended and held in place inside the tokamak by powerful magnetic fields. The heat from the fusion reaction can be used to generate steam, which in turn can power a turbine to create electricity.

The A.I. software that DeepMind is developing learns to control the magnetic fields that contain the plasma inside the tokamak. The system was able to manipulate the plasma into new configurations that can produce higher energy, but which physicists had been reluctant to attempt using previous control methods.

“This allows us to push things forward because we can take risks we would not dare take otherwise,” Ambrogio Fasoli, one of the Swiss Plasma Center scientists involved in the project, said. “Some of these [plasma] shapes that we are trying are taking us very close to the limits of the system, where the plasma might collapse and damage the system, and we would not risk that without the confidence of the A.I.”

It’s been a big two weeks for advances in fusion power. Last week, a group of European physicists working at the Joint European Torus Laboratory in England managed to create the most powerful controlled fusion power reaction in history. The experiment produced 59 megajoules of energy (the equivalent of about 11 megawatts of power) over a five-second reaction. That is twice the power of the previous record, set in 1997.

JET’s tokamak is much larger and more powerful than the TCV tokamak used by the Swiss Plasma Center. That smaller tokamak can sustain a fusion reaction only for a maximum of two seconds, members of the Swiss research team said.

But similar methods to those used for the A.I. control algorithm at the Swiss Plasma Center might also be adaptable to larger, more powerful fusion reactors. The world’s largest such system is currently under construction in southern France, with support from a consortium of governments, including members of the European Union, U.S., China, and Russia.

Experts hope that fusion power will be developed enough to start powering portions of the world’s energy grid sometime in the second half of this century. Fusion offers the prospect of almost limitless energy from simple, relatively easy-to-source elements, and produces no greenhouse gases and relatively small amounts of radioactive waste that break down within about a century. Fission reactors, which are used in all existing nuclear power plants, on the other hand, produce large amounts of highly radioactive waste, some of which remains dangerous for tens of thousands of years.

The time frame in which nuclear fusion is likely to be commercially viable, however, is not fast enough for the technology to play much of a role in the current race to decarbonize the world’s energy sources and avert catastrophic global warming.

Pushmeet Kohli, who leads DeepMind’s efforts to use A.I. to address challenges in science, said that the fusion project showed that the research company is able to make fundamental impacts in physics. In late 2020, the company showed that an A.I. system it had created, called AlphaFold, could effectively predict the three-dimensional shape of a protein from its genetic sequence, a major breakthrough in biology that is likely to have far-reaching impacts on the field, including in the area of drug discovery. Previously, the company was best known for creating an A.I. system that could beat the world’s top players at the strategy game Go.

The A.I. system that DeepMind developed to manipulate the magnetic control system of the tokamak uses a method called reinforcement learning, in which the system learns by trial and error in a simulator. A concern with using this technique, however, is whether the simulator is good enough to allow the A.I. to effectively control a real tokamak. “We thought the simulation might not be good enough,” Jonas Buchli, a DeepMind researcher who worked on the project, said.

One issue is that the simulator did not accurately capture all of the variables present in a real tokamak. But Buchli said that by using a method where these factors were represented by random numbers in the simulation, DeepMind was still able to train an A.I. that was flexible enough to transfer its knowledge to the real tokamak.

Another issue is that in order to keep the plasma controlled inside the tokamak, the control algorithm must be able to make extremely fast decisions, executing adjustments to the magnet fields in just fractions of a second. Many A.I. systems take too long to make predictions to work in such a high-speed environment.

So the DeepMind team trained the A.I. system with two components. One is a large neural network, a type of A.I. designed loosely on how parts of the human brain function, that makes longer-term predictions about how changes to the magnetic field will shape the plasma. This network is then used to help train a much smaller system that learns the best way to implement the decisions that the first network recommends. But only the smaller network interacts directly with the tokamak control system because it has to be able to make decisions in less than 50 microseconds (50 millionths of a second).