But a Princeton-led team made up of engineers, physicists and data scientists from the university and the U.S. Department of Energy Princeton Plasma Physicslaboratory (PPPL) harnessed the power of artificial intelligence to predict – then prevent – the formation of a specific plasma problem in real time.
In experiments at National DIII-D Fusion Facility In San Diego, the researchers demonstrated that their model, trained only on previous experimental data, could predict potential plasma instabilities, known as tear mode instabilities, up to 300 milliseconds in advance. Although this does not leave more time than is necessary for a slow blink in humans, the AI controller had plenty of time to change some operating parameters to avoid what would have turned into a tear in the plasma’s magnetic field lines, disrupting its balance and openness. the door for an escape ending the reaction.
“By learning from past experiences, rather than incorporating information from physics-based models, AI could develop a final control policy that would support a stable, high-power plasma regime in real time, in a real reactor,” said research manager Egemen Kolemen. , associate professor of mechanical and aerospace engineering and the Andlinger Center for Energy and Environmentas well as a research physicist at PPPL.
The research opens the door to more dynamic control of a fusion reaction than current approaches and provides a basis for using artificial intelligence to address a wide range of plasma instabilities, which have long been obstacles to achieve a sustained fusion reaction. The team published their discoveries In Nature on February 21.
“Previous studies have generally focused on suppressing or attenuating the effects of these tear instabilities after they appear in the plasma,” said the first author. Jaemin Seoassistant professor of physics at Chung-Ang University in South Korea, who did much of the work while he was a postdoctoral researcher in Kolemen’s group. “But our approach allows us to predict and avoid these instabilities before they appear.”
Superheated plasma swirling in a donut-shaped device
Fusion occurs when two atoms – usually light atoms like hydrogen – come together to form a heavier atom, releasing a large amount of energy. This process powers the Sun and, by extension, makes life possible on Earth.
However, getting the two atoms to fuse together is tricky, because it requires enormous amounts of pressure and energy for the two atoms to overcome their mutual repulsion.
Fortunately for the Sun, its massive gravitational pull and extremely high pressures at its core allow fusion reactions to proceed. To reproduce a similar process on Earth, scientists instead use extremely hot plasma and extremely powerful magnets.
In donut-shaped devices called tokamaks – sometimes called “stars in jars” – magnetic fields struggle to contain plasmas that reach more than 100 million degrees Celsius, hotter than the center of the Sun.
Although there are many types of plasma instabilities that can terminate the reaction, the Princeton team focused on resolving tear mode instabilities, a disturbance in which magnetic field lines at within a plasma break and create an opportunity for subsequent plasma leakage.
“Tear mode instabilities are one of the main causes of plasma disruption, and they will become even more important as we try to achieve fusion reactions at the high powers required to produce sufficient energy,” said SEO. “They represent a significant challenge for us.”
Merging artificial intelligence and plasma physics
Because tear mode instabilities can form and derail a fusion reaction in milliseconds, researchers have turned to artificial intelligence for its ability to quickly process and act in response to new data.
But the process of developing an effective AI controller hasn’t been as simple as trying a few things on a tokamak, where time is limited and the stakes are high.
Co-author Azarakhsh Jalalvandresearcher in Kolemen’s group, compared teaching an algorithm to run a fusion reaction in a tokamak to teaching someone how to fly an airplane.
“You wouldn’t teach someone by handing them a set of keys and telling them to do their best,” Jalalvand said. “Instead, you would have them train on a very complex flight simulator until they have learned enough to try the real thing.”
Much like developing a flight simulator, the Princeton team used data from previous DIII-D tokamak experiments to build a deep neural network capable of predicting the likelihood of future tearing instability on the basis of plasma characteristics in real time.
They used this neural network to train a reinforcement learning algorithm. Like a trainee pilot, the reinforcement learning algorithm could try different plasma control strategies, learning through trial and error which strategies worked and which did not in the safety of a simulated environment.
“We don’t teach the reinforcement learning model all the complex physics of a fusion reaction,” Jalalvand said. “We tell him what the goal is – maintaining a high-powered reaction – what to avoid – tearing mode instability – and the buttons he can push to achieve those results. Over time, it learns the optimal path to achieve the goal of high power while avoiding the punishment of instability.
While the model has been the subject of countless simulated fusion experiments, trying to find ways to maintain high power levels while avoiding instabilities, co-author SangKyeun Kim could observe and refine their actions.
“In the background, we can see the intentions of the model,” said Kim, a PPPL research scientist and former postdoctoral researcher in Kolemen’s group. “Some of the changes the model wants are too fast, so we are working to smooth and calm the model. As humans, we arbitrate between what the AI wants to do and what the tokamak can accomplish.
Once confident in the AI controller’s capabilities, they tested it in a real fusion experiment on the D-III D tokamak, observing the controller make real-time changes to certain tokamak parameters to prevent the appearance of instability. These parameters included changing the shape of the plasma and the strength of the beams powering the reaction.
“Being able to predict instabilities in advance can make it easier to implement these reactions compared to current approaches, which are more passive,” Kim said. “We no longer need to wait for instabilities to occur and quickly take corrective action before the plasma is disrupted. »
Projecting towards the future
Although the researchers say this work represents a promising proof of concept demonstrating how artificial intelligence can effectively control fusion reactions, it is just one of many next steps already underway in Kolemen’s group to advance the field of fusion research.
The first step is to obtain more evidence that the AI controller works on the DIII-D tokamak, and then expand the controller to work on other tokamaks.
“We have strong evidence that the controller works quite well at DIII-D, but we need more data to show that it can work in a number of different situations,” said first author Seo. “We want to work towards something more universal.”
A second line of research involves extending the algorithm to handle many different control problems simultaneously. Although the current model uses a limited number of diagnostics to avoid a specific type of instability, researchers could provide data on other types of instabilities and provide access to more buttons for the AI controller to adjust.
“You could imagine a large reward function that flips many different buttons to control multiple types of instabilities simultaneously,” the co-author said. Ricardo Shoushapostdoc at PPPL and former graduate student in Kolemen’s group that supported the experiments at DIII-D.
And by developing better AI controllers for fusion reactions, researchers could also better understand the underlying physics. By studying the decisions of the AI controller as it attempts to contain the plasma, which may be radically different from what traditional approaches might prescribe, artificial intelligence can be not only a tool for controlling fusion reactions , but also an educational resource.
“Eventually, this could be more than just a one-way interaction between scientists developing and deploying these AI models,” Kolemen said. “By studying them in more detail, they might have some things they could teach us as well.” »
The newspaper, “Avoiding fusion plasma tearing instability through deep reinforcement learning», was published on February 21 in Nature. In addition to Kolemen, Seo, Jalalvand, Kim and Shousha, co-authors include Rory Conlin, Andrew Rothstein, Joseph Abbate and Josiah Wai of Princeton University, as well as Keith Erickson of PPPL.
The work was supported by the U.S. Department of Energy’s Office of Fusion Energy Sciences, as well as the National Research Foundation of Korea (NRF). The authors also acknowledge the use of the National DIII-D Fusion Facility, a Department of Energy Office of Science User Facility.