AI just cleared a major hurdle on the road to nuclear fusion energy

Futurists of the past always thought that breakthroughs in technology and science could create a utopian world fueled by limitless clean energy. Now an artificial intelligence model from Princeton researchers may have proven them right. Or at least it brought us one step closer.

Fusion – the nuclear reaction in which two or more atomic nuclei come together to form new nuclei and subatomic particles – has long been the dream as an energy source: it is non-polluting, safe and virtually unlimited, producing almost four million times as much energy en masse as burning of fossil fuels.

Unfortunately there is a problem. Fusion is actually Real difficult to do: it requires the kind of temperature and pressure found in the hearts of stars. Since we can’t really get those exact conditions in a laboratory on Earth, the relatively few examples of man-made fusion have relied on a solution: normal terrestrial pressure and temperatures more than ten times that of the core of the sun.

At those temperatures, the fuel needed for the reaction cannot exist in a solid or liquid state, and is not even in the form of a gas: it is plasma. Therein lies another problem: This state of matter is so energetic and overheated that the fuel can easily ‘rupture’ – lose stability and escape the magnetic fields that hold it in the reactor – ending any fusion within milliseconds.

It is precisely this problem that the Princeton team claims to have solved.

“Previous studies have generally focused on suppressing or mitigating the effects of these crack instabilities after they occur in the plasma,” explains first author of the new paper Jaemin Seo, now an assistant professor of physics at Chung-Ang University in South Korea. in a statement. “But our approach allows us to predict and avoid these instabilities before they ever occur.”

Their answer: an artificial intelligence (AI) trained on previous experiments at the DIII-D National Fusion Facility in San Diego.

“By learning from past experiments, rather than integrating information from physics-based models, the AI ​​could develop a definitive control policy that supported a stable, high-performance plasma regime in real time, in a real reactor,” says research leader Egemen Kolemen. , associate professor of mechanical and aerospace engineering and the Andlinger Center for Energy and the Environment and research physicist at the Princeton Plasma Physics Laboratory (PPPL).

Like any AI model, it doesn’t really understand what it’s doing at a deep level, but it doesn’t need to. The team fed the program data on real-time plasma characteristics from previous experiments and set the challenge to predict – and, crucially, avoid – crack instabilities.

“We don’t teach the reinforcement learning model all the complex physics of a fusion reaction,” explains Azarakhsh Jalalvand, a researcher in Kolemen’s lab and co-author of the paper. “We tell it what its purpose is – maintaining a strong response – what to avoid – a rupture mode instability – and what knobs it can turn to achieve those results. Over time, it learns the optimal route to achieve the goal of great power while avoiding the penalty of instability.”

After numerous simulations, which could be adjusted and refined by human observers, the team tried out the AI ​​in real life at the D-III D facility. The model proved to be able to predict crack instabilities up to 300 milliseconds in advance – not much for a human, but enough time for the AI ​​to intervene, by changing parameters such as the shape of the plasma or the strength of the beams carrying current supply. the reaction to keep the plasma stable.

Is unlimited clean energy just around the corner? Not quite. Plasma instability is far from the only problem with fusion – and rupture is just one form of possible plasma instability.

But what the paper does show, the team says, is a pretty good proof-of-concept: “We have strong evidence that the controller works quite well at DIII-D, but we need more data to show that this can work in some cases. different situations,” Seo said. “We want to work on something that is more universal.”

The article was published in the journal Nature.

Leave a Reply

Your email address will not be published. Required fields are marked *