Researchers have developed an algorithm to control the hot, chaotic plasma that fuels fusion energy reactors.
Why it matters: The plasma in fusion reactors needs to be controlled in order to extract energy from the reactions. Fusion is a potential source of clean fuel that has long been promised — but not yet delivered.
Background: The goal of nuclear fusion is to harness the energy released when two atomic nuclei of hydrogen are pushed together.
- Experimental reactors are notching advances in creating energy but to become useful as a commercial fuel source, reactors will have to generate more energy than is required to produce the reaction in the first place.
- One approach for creating and controlling plasma is to use powerful magnets in a donut-shaped reactor called a tokamak.
How it works: The tokamak at the Swiss Plasma Center at the EPFL is used to study the optimal conditions for confining constantly-changing plasma.
- The plasma's shape and distribution in the tokamak can be adjusted by changing the voltage in the 19 magnets that hold it in place.
- But testing new plasma configurations by changing the interconnected conditions in the tokamak requires significant engineering and design effort.
What they did: DeepMind researchers developed a deep reinforcement learning algorithm that learned to control the magnetic coils to yield — and maintain — different plasma configurations.
- The algorithm takes in 90 observations about the plasma and adjusts the coils to produce different plasma shapes.
- They first trained the algorithm in a simulator then tested it in the tokamak at the Swiss Plasma Center.
- It was able to produce a range of shapes — conventional, elongated plasmas and advanced ones, including a snowflake configuration and an arrangement of two separate plasmas at the same time, the researchers reported in the journal Nature.
The intrigue: The algorithm was able to handle a major challenge presented by the plasma — its state changes constantly. Unlike a game of Go, which a DeepMind algorithm has mastered, there aren't discrete choices to be made, Jonas Buchli of DeepMind said in a press briefing. There are "an infinite amount of possibilities."
- "AI, and specifically reinforcement learning, is particularly well-suited to the complex problems presented by controlling plasma," Martin Reidmiller of DeepMind said in the briefing.
What's next: The results suggest there may be new ways to control the plasma, Anne White, a professor of nuclear science and engineering at MIT who was not involved in the research told Axios in an email.
- Those might include "how to optimize control of the plasma shape and heat load at the divertor (the main heat “exhaust pipe”) dynamically during a tokamak pulse."
Keep in mind: The tokamak in the study is designed for research purposes. But the researchers say the machine learning approach could be used to test control of future reactors.