AI Can Predict and Prevent Fusion Reaction Instabilities

Ever since the concept of nuclear fusion was understood close to a century ago, from engineers to scientists, everyone has been in search of ways to create and harness it. After all, once we achieve nuclear fusion at an industrial scale, it could offer affordable, safe, non-polluting, and virtually limitless energy to meet rising demand.

What is nuclear fusion? It's a process where two or more light atomic nuclei come together to form a single, different, and heavier atomic nucleus, releasing massive amounts of energy. Interestingly, fusion reactions power the sun and the stars, making life on Earth possible. 

However, fusing two atoms is rather difficult, as it requires a massive amount of pressure and energy for them to get over their mutual repulsion. The sun does so through its massive gravitational pull and very high pressures at its core. 

Now, researchers are attempting to build fusion reactors of their own. These fusion reactions take place in plasma, a state of matter. Plasma, a hot, charged gas consisting of free negative electrons and positive ions, allows electric current to flow through it.

To replicate this process, scientists use extremely hot plasma and super strong magnets. However, the superheated, highly complex, and disorderly plasma can lose its stability really fast and escape the strong magnetic fields that enclose it within the donut-shaped fusion reactor, usually signaling the termination of the reaction. These donut-shaped devices are known as tokamaks.

This, of course, poses a critical challenge in developing fusion as a clean and limitless energy source. However, a team of data scientists, engineers, and physicists from Princeton University and the U.S. Department of Energy's Princeton Plasma Physics Laboratory (PPPL) is utilizing artificial intelligence (AI) to forecast the formation of a specific plasma problem in real time so as to avoid it.

At San Diego's DIII-D National Fusion Facility, which is an Office of Science operated by General Atomics for the U.S. Department of Energy (DOE), the researchers ran their experiments and demonstrated their AI model, which was shown to predict potential plasma instabilities. Dubbed tearing mode instabilities, these instabilities can be predicted by AI well in advance by as much as 300 milliseconds. For context, we blink on an average of 100–150 milliseconds, as per UCL researchers.

While an extremely short duration, this provides plenty of time for the AI controller to change specific operating parameters in order to prevent the tearing within the magnetic field lines of the plasma that would upset the equilibrium and potentially end the reaction.

The AI model was trained on data from previous trials instead of utilizing information from physics-based models. This way, “the AI could develop a final control policy that supported a stable, high-powered plasma regime in real-time, at a real reactor,” said the study lead Egemen Kolemen, who's a staff research physicist at PPPL along with being an associate professor of mechanical and aerospace engineering at the Andlinger Center for Energy and the Environment.

Unlike current approaches, this allows for more dynamic control of a fusion reaction and further provides a foundation for leveraging AI to find a solution to a wide set of plasma instabilities. This is critical because it has been presenting a challenge in accomplishing a constant fusion reaction for a long time.

Click here to learn all about investing in artificial intelligence (AI).

AI's Growing Usage in Fusion Research

Published in Nature, this study concentrates on predicting and avoiding tearing instabilities before they even appear in plasma. This makes it different from previous approaches, which, according to Jaemin Seo, the study author, who's an assistant professor of physics at South Korea's Chung-Ang University, have normally worked on mitigating or subduing the effects of these instabilities after they occur. 

The Princeton team tackled tearing mode instabilities, a type of disturbance where magnetic field lines fail to contain plasmas exceeding 100 million degrees Celsius, leading to plasma disruption. This is hotter than the center of the Sun. According to Seo:

“Tearing mode instabilities, one of the major causes of plasma disruption, will only become even more prominent as we try to run fusion reactions at the high powers required to produce enough energy, presenting a major challenge that needs to be solved.”

Given that this type of instability can form and then thwart the fusion reaction in a matter of milliseconds, researchers have turned to AI for its ability to quickly process and act in response to new data.

This wasn't the first time AI was leveraged in fusion research, though. In fact, there has been a growing interest in machine learning and AI to evaluate enormous amounts of data produced in these experiments. 

On evaluating training data, the AI then recognizes patterns and derives principles from them. For instance, Wendelstein 7-X deals with live detection of the state of plasma equilibrium in stellarators, a plasma device that depends on external magnets to confine plasma.

As Riccardo Betti, chief scientist at LLE, and Robert L. McCrory, Professor in the Department of Physics and Astronomy and the Department of Mechanical Engineering, said:

“We now have a wealth of experimental data that we can harness with machine learning to systematically correct the simulations and guide real-time adjustments to experiments.”

Meanwhile, in a study in 2021, Diogo Ferreira, a professor of information systems at the University of Lisbon's Instituto Superior Técnico in Portugal, detailed three different uses for AI, ML, and DL models for fusion research. One of his models predicts disruptions in a super-hot plasma, the second one detects anomalies in the plasma, and the third one is about visual representations of plasma radiation patterns.

Now, this latest study, which is supported by the DOE's Office of Fusion Energy Sciences and the National Research Foundation of Korea, is working on preventing fusion plasma tearing instability with the help of machine learning (ML) subfield deep reinforcement learning (RL). DOE has also been providing funding support to MIT Plasma Science and Fusion Center to improve access to fusion data.

To build a successful artificial intelligence controller, the Princeton team has to use data from tests run in the past at the DIII-D tokamak. A deep neural network was then created with the ability to predict the possibility of upcoming tearing instability depending on the characteristics of plasma in real time.

This, according to Azarakhsh Jalalvand, the study's co-author and a research scholar in Kolemen's group was like teaching someone how to fly a plane, which you wouldn't be “handing them a set of keys and telling them to try their best,” rather, “you'd have them practice on a very intricate flight simulator until they've learned enough to try out the real thing.”

The team then trained a reinforcement learning (RL) algorithm on their neural network, which can try out different strategies for controlling plasma. The algorithm learns through trial and error within the safety of a simulated environment.

Instead of teaching it all the complex physics of a fusion reaction, the team only told the reinforcement learning model their goal, which was to maintain a high-powered reaction along with what to avoid, i.e., a tearing mode instability and the knobs it can use to achieve those outcomes. Then, over time, the model learns the optimal pathway to achieve high power levels while avoiding the punishment of instability.

As the model ran through countless simulated fusion experiments, the team observed and refined its actions due to some changes being too rapid. 

“As humans, we arbitrate between what the AI wants to do and what the tokamak can accommodate.” 

– SangKyeun Kim, Co-author of the study & staff research scientist at PPPL

The team only tested the AI controller during an actual fusion experiment once they were confident in its abilities. The AI controller then made real-time changes to certain D-III D tokamak parameters, including changing the plasma's shape and the strength of the beams that put power into the reaction to avoid the onset of instability.

A Universal Solution

This latest research, whose findings were published last week in Nature, presents an active approach as opposed to the “more passive” current approaches to predicting instabilities ahead of time, which can make it easier to run these reactions. Kim said:

“We no longer have to wait for the instabilities to occur and then take quick corrective action before the plasma becomes disrupted.”

This study is certainly a promising proof-of-concept showcasing just how AI can effectively control fusion reactions. However, researchers noted that the Kolemen's group is already working on several next steps to advance the field of fusion research. 

So, while there is “strong evidence” that the AI controller works “quite well” at DIII-D tokamak, the focus is now to get more evidence of the controller in action, have more data that it also works in different situations, and then expand it to function at other tokamaks. 

But this is not all. The current AI model uses only a limited number of diagnostics that allow it to avoid just one specific type of instability. Hence, the researchers want to provide data on other types of instabilities as well as provide access to more knobs for the AI controller to make adjustments. So, here the team aims to expand its algorithm to handle several types of instabilities by controlling many different knobs at the same time.

On this road to creating more efficient AI controllers, the team hopes the AI will have enhanced clarity and comprehension of fusion reactions and physics and teach us more about it all, too.

Click here to learn why 2023 was the breakout year for AI.

Companies Working in the Field of AI & Fusion Reaction

Now, let's take a look at some prominent names in the sector:

#1. Google 

The tech giant has its hands in every pie, and that includes AI research and fusion energy. A couple of years ago, Google-backed DeepMind announced that it had trained an AI system to control plasma inside a nuclear fusion reactor. The AI Lab created an RL AI system to control magnets and change voltage thousands of times every second.

The $1.75 trillion market company has its shares trading at $140.38, up 3% YTD. Google posted revenue (TTM) of 307.39 bln and has an EPS (TTM) of 5.80, P/E (TTM) of 24.17, and ROE (TTM) of 27.36%.

#2. TAE Technologies

Formerly known as Tri Alpha Energy, the California-based company is focused on developing fusion energy tech. TAE Technologies is currently upgrading its fusion platform, Norman, to a sixth-generation machine called Copernicus. If all goes smoothly, the company expects to build its first prototype power plant that could connect to the grid in the early 2030s, scaling which to develop “robust and reliable” commercial power would continue through the decade. Fusion, according to its CEO Michl Binderbauer, would take us into a “paradigm of abundance.”

In 2022, Google and Chevron invested in TAE Technologies as part of the company's $250 million funding raise. Google has actually been partnered with TAE for a decade now and provides the company with AI and computational power. 

#3. ITER

International Thermonuclear Experimental Reactor is an international nuclear fusion research and engineering megaproject that has been advancing its fusion research and development with ML and AI.

“Big Science projects like ITER offer a wealth of data which is ideal for AI. They provide us with a one-of-a-kind opportunity to learn, train, extrapolate, and apply these skills in other fields of manufacturing.” 

– María Ortiz de Zúñiga, Senior Technical Officer at Fusion for Energy

Conclusion

As we talked about throughout this piece, AI has the potential to transform various aspects of the fusion journey. However, recent advances in technology have raised hopes of overcoming the long-standing challenges faced by this industry. By leveraging artificial intelligence's computational power and pattern recognition capabilities, we can accelerate the development of fusion power and finally make a sustainable energy future a reality.