Under enormous heat and gravity, the nuclei of hydrogen in the sun's core collide with each other, coalescing into heavier helium atoms, and releases a lot of energy in the process. For decades, scientists and engineers have explored the use of a doughnut-shaped tokamak device to constrain plasma for controlled nuclear fusion. How to effectively control the plasma is the key to nuclear fusion.
"We need to heat these substances and keep them long enough to get energy from them." Ambrogio Fasoli, director of the Plasma Center at the Swiss Federal Institute of Technology in Lausanne, said.
To achieve nuclear fusion, three conditions must be met: extremely high temperatures, sufficient plasma particle density, and sufficient time limiting. This is where artificial intelligence comes in. On Feb. 16, a paper titled "Magnetic control of tokamak plasmas through deep reinforcement learning through deep reinforcement learning" was featured in Nature, a collaboration between Deepmind, an artificial intelligence company owned by Google, and physicists at the Plasma Center at the Swiss Federal Institute of Technology in Lausanne.
In an environment of more than 100 million °C, hydrogen superheats into a plasma state. No material can control a plasma at such a temperature, but in a tokamak device, a strong magnetic field suspends and fixes the plasma inside the tokamak, forcing it to maintain its shape and preventing it from touching the reactor wall (touching the reactor wall will cool the plasma and damage the reactor).
Under the previous control methods, it was very risky to produce higher energy, and physicists did not dare to try it easily. What Deepmind does is train the AI to learn to precisely control the magnetic field containing the plasma inside the tokamak.
"It allows us to move things forward because we can take risks that we wouldn't have dared to take." Ambrogio Fasoli, one of the scientists at the Swiss Plasma Center involved in the project, said, "Some of the plasma shapes we are trying are bringing us very close to the limits of the system. In such a situation, the plasma may collapse and damage the system. Without confidence in AI, we wouldn't have taken that risk. ”
Precise plasma control requires constant monitoring and manipulation of magnetic fields. "The more complex and performance the tokamak, the greater the reliability and accuracy to control." Dmitri Orlov, a scientist at the San Diego Energy Research Center, said in an interview with Wired.
To achieve this, the world's first AI to achieve autonomous control of plasma in a nuclear fusion device (tokamak) needs to solve two problems: accurately capture all variables present in a real tokamak device, and make a decision in less than 50 microseconds (50 millionths of a second).
Shaping plasma by magnetic field changes (00:04)
The research team used a large neural network to train the shapes and positions of 90 plasmas ten thousand times per second, continuously making long-term predictions about how magnetic field changes shape the plasma, and adjusting the voltages of the 19 magnets accordingly. This neural network is then used to train a small system to learn to perform the best decisions recommended by the first network. This makes it both accurate and fast.
"This is one of the most challenging applications of reinforcement learning in the real world to date," said Martin Riedmiller, a deepMind research scientist, "but to be clear, that doesn't mean we've solved the fusion problem." What it stands for... this is an important step in our understanding of how to design a new flexible tokamak controller. ”
After creating AlphaGo, which beat the World Champion of Go, DeepMind entered the public eye. Since then, it has also launched AlphaFold, which predicts the three-dimensional structure of proteins through genetic sequences.
"Today's big problems in science rarely reduce to a small set of elegant or compact formulas, solved by one person or a small team," deepMind research scientist Jonas Buchli once said, "We believe that artificial intelligence is a multiplier of human creativity, it opens up new areas of exploration that allow us to reach our full potential." Today, AI systems are powerful enough to be applied to many real-world problems, including scientific discoveries themselves. ”
This isn't the first time AI has been used to control nuclear fusion. Since 2014, Google has been working with fusion company TAE Technologies to apply machine learning to different types of fusion reactors — accelerating the analysis of experimental data. Research from the UK's Joint EUROPEAN Torus fusion project has used AI to try to predict the behavior of plasma. The concept even appeared in 2004's Spider-Man 2, where the villain Dr. Oak created an AI-powered, brain-controlled exoskeleton to control his experimental fusion reactor.
Recently, JET generated a total of 59 megajoules of energy in a nuclear fusion experiment lasting 5 seconds, significantly breaking the record it set in the 1997 experiment. This achievement will play an important role in promoting the International Thermonuclear Fusion Experimental Reactor (ITER), which is still under construction.