how artificial intelligence is helping to crack fusion

If you’re familiar with the quest to bring nuclear fusion to the masses, you may already know the joke that fusion is a technology that’s always just 20 years away. Big promises and a lack of visible progress has prompted many a commentator to condemn the research as a total waste of time and effort, and opened the door for charlatans claiming they found some sort of quick and easy shortcut to creating a miniature star. But lately, there’s been some promising news and movement, with new reactor designs coming online, and duration and generation records being achieved by test reactors. And this is before ITER, the massive international experiment, gets things really fired up at an unprecedented scale.

So, does this mean we’re on the verge of commercially viable fusion? Not yet, and not by a long shot. But the reason why we’re starting to see a trickle of positive news is because despite all of the naysaying and shoestring budgets, scientists have been able to demonstrate that it really is possible to ignite and sustain fusion reactions outside of nuclear weapons. Unlike 70 years ago, we have a firm grasp on the fundamentals, and the kinds of lasers and magnets that make the ignition, the critical first part of the process, almost easy. Whereas before, the only way to start fusion was to set off a fission bomb inside another bomb, forcing deuterium and lithium to fuse into helium, we can now do it with fuel pellets and lasers.

Early fusion boosters seemed to think that as soon as we can reliably start the ignition process, we could work out what to do next with enough magnets. Since the ignition would create very hot clouds of plasma, and plasma can be contained and channeled with magnets, we’d just need enough good magnets, they assumed. Unfortunately, it turns out that containing all that plasma is the hard part. Fusion reactions require the plasma to reach at least 100 million C, and ideally get closer to 150 million C, nearly ten times hotter than the core of the Sun. Meanwhile, the magnets meant to manage this plasma are cooled to near absolute zero, creating up some of the biggest temperature gradients in the known universe within a single device.

why taming unruly plasma is the key to fusion

Why do we need such insane temperature differences? Well, because the sun cheats when it fuses hydrogen into helium, using its gravity to generate core pressures over 6 trillion psi. At that extreme, fusion is the only option because the ions are literally crushed together. Unless we figure out how to add at least 30,000 Earth masses without imploding our planet into the smallest possible star capable of fusion, the only route we have is to heat atoms to absolutely insane temperatures and use magnetic fields generated by superconductors, kept within just a few degrees of absolute zero for optimal function, to squeeze those atoms in an energetic coil where they lose their electrons and get shoved together until they fuse into helium.

The longer we can keep the plasma coil circulating, the more fusion reactions we’ll create, the more energy we’ll produce. Unfortunately, that’s easier said than done because at some point, that 150 million degree plasma will overheat the magnets, break its confinement, and disperse within the chamber, terminating fusion. Figuring out the right shapes and mechanics for those magnets is what’s been holding fusion back from viability, and having to nail this down using very limited, complex modeling followed by nine figure trial-and-error experiments has taken a lot of time. But a Google backed venture called DeepMind, famous for the psychedelic imagery its artificial neural networks generate, had some helpful ideas.

Partnering with the Swiss Plasma Center, its engineers created a program which can detect the minute behaviors of the plasma coil and adjust the 19 independent magnets in the test reactor accordingly to extend the lifetime of the reaction cycle. The goal was to figure out exactly what configurations would lead to optimal confinement and whether an AI would be capable of swift and careful adjustments if the plasma’s behavior changes. So far, the test appears to yield very promising results, showing that an AI system controlling a fusion reactor would be both useful and informative since it could also be unleased to design new magnet configurations based on real world data. And that should hopefully lead us to a commercially viable reactor.

all hail our fusion-powering machine overlords

By itself, trying to use AI to crack the plasma confinement problem isn’t new, and models have been trained to understand the fundamentals of plasma flows in a fusion reactor. But this is the first time a network has been so elaborate and had such ambitious goals, guiding humans to the optimal configurations and energy levels, and creating room to add more data to detect and avoid various flares and undesirable feedback loops seen in test reactors. Partly, this is a function of more researchers feeling comfortable with artificial intelligence trying to tackle very complex problems since they now have a track record of doing that in the real world, and partly it’s thanks to the standardization and ubiquity of AI tools.

In other words, more ambitious AI models are starting to be applied to making fusion a viable energy source because it’s now a far more mature technology and researchers understand it much better than before. Right now, we can keep fusion going for just a few minutes. When we’re routinely maintaining it for a few hours at a time, we can start planning how to build a new power plant and the exact feedstock for these new reactors, since our experimental mixes wouldn’t be the best fuel, generating too many neutrons for long term use. Instead, we’d more likely use helium-3 found in lunar regolith or a mix of hydrogen and boron, if it works in future experiments as a cleaner, more accessible alternative.

If we do reach that point, it seems almost certain that AI models will get us there and start to push us further, helping miniaturize these reactors for use in microgrids, spacecraft, and bases on other worlds. Ideally, this would be the exact future of artificial intelligence in our lives: as an army of silent, helpful algorithms focused on helping us solve complex, dynamic problems that require intense number crunching to solve, and empower engineers to build machines that make our lives better, more comfortable, and less environmentally impactful. At this point, to keep wasting AI on cramming useless trinkets and monetizable outrage down our throats when fusion and space travel call, seems absolutely insane.

See: Degrave, J., et al. (2022) Magnetic control of tokamak plasmas through deep reinforcement learning, Nature 602, 414–419, DOI: 10.1038/s41586-021-04301-9