Why cracking nuclear fusion will depend on artificial intelligence


THE big joke about sustainable nuclear fusion is that it has always been 30 years away. Like any joke, it contains a kernel of truth. The dream of harnessing the reaction that powers the sun was big news in the 1950s, just around the corner in the 1980s, and the hottest bet of the past decade.

But time is running out. Our demand for energy is burning up the planet, depleting its resources and risking damaging Earth beyond repair. Wind, solar and tidal energy provide some relief, but they are limited and unpredictable. Nuclear fission comes with the dangers of reactor meltdowns and radioactive waste, while hydropower can be ecologically disruptive. Fusion, on the other hand, could provide almost limitless energy without releasing carbon dioxide or producing radioactive waste. It is the dream power source. The perennial question is: can we make it a reality?

Perhaps now, finally, we can. That isn’t just because of the myriad fusion start-ups increasingly sensing a lucrative market opportunity just around the corner and challenging the primacy of the traditional big-beast projects. Or just because of innovative approaches, materials and technologies that are fuelling an optimism that we can at last master fusion’s fiendish complexities. It is also because of the entrance of a new player, one that could change the rules of the game: artificial intelligence. In the right hands, it might make the next 30 years fly by.

Nuclear fusion is the most widespread source of energy in the universe, and one of the most efficient: just a few grams of fuel release the same energy as several tonnes of coal. These vast quantities of energy have their origins in something vanishingly small: the nucleus of an atom. Consisting of positively charged protons and neutral neutrons orbited by negatively charged electrons, the nucleus makes up the bulk of an atom’s mass.

When two or more small atomic nuclei come into contact, they can, under certain circumstances, merge to form larger nuclei, releasing huge amounts of energy in the process. On a gargantuan scale, this is what takes place within the core of stars like our sun, giving them the power they need to shine for billions of years.

Fusion’s extraordinary potential has been tantalising scientists for decades, but remains difficult to realise on Earth. It requires the creation of a “plasma” of naked atomic nuclei at huge temperatures and densities – something that is both difficult to achieve and difficult to control (see “Why fusion is so hard”).

At present, the most popular approach is to use what is called a magnetic confinement fusion device. In fusion’s early days in the 1950s, the favoured design was a kinked doughnut shape known as a stellarator. These machines created complex magnetic fields that could theoretically hold a charged plasma steady, but their twisted shape made them tricky to build.

By the 1970s, interest had turned to simpler designs: vast hollow rings called tokamaks in which trapped plasma is heated to hundreds of millions of degrees. The forces required to keep such a plasma in place can only be generated by powerful superconducting magnets cooled to close to absolute zero, creating the sharpest temperature gradients in the known universe.

These magnetic confinement devices have had a few successes over the years. In 1997, the Joint European Torus (JET) near Oxford, UK, set the world record for the amount of energy created in a fusion reaction, producing 16 megawatts of fusion energy from an input of 24 megawatts. This is the closest anyone has got to breaking even – getting as much energy out as that pumped in – but the reaction lasted for only a few hundredths of a second.

Back then, break-even seemed around the corner, but strange instabilities appeared in JET’s plasma that worked to cool down its centre and stymie the plans. Now, after years of upgrades, changes in design and materials, the reactor is back. In November 2020, JET is set to power its first fusion reaction in more than 20 years, aiming to beat its previous energy record and sustain the reaction for longer.

Meanwhile, other players have been getting in on the act. In 2018, China’s Experimental Advanced Superconducting Tokamak (EAST) sustained a plasma at temperatures of 15 million °C for 100 seconds, the longest confinement time yet.

EAST plans to start operating again in 2020, but is comparatively small fry. The heavily backed favourite in the race is the huge International Thermonuclear Experimental Reactor, or ITER. Founded in 1985 as a collaboration between 31 nations including China, the US, Russia and the European Union, ITER was originally expected to start experiments in 2016, but design challenges mean it is likely to remain under construction in France until 2025. “ITER is a first-of-a-kind facility,” says Howard Wilson at the University of York, UK. “It will take 10 years to learn how to bring it up to its full performance.”

ITER currently aims to begin fusion reactions in 2035, and it has big goals: pushing beyond break-even to produce 10 times as much power as goes in. Despite the delays, there is confidence that ITER will achieve this. “The question now is, do we have the technology to make a commercially viable power plant?” asks Simon Pinches, head of the plasma physics division at ITER.

Even if ITER achieves its goals, its journey will be far from over. The reactor isn’t set up to capture the energy it produces as electricity. Instead, the idea is that it will pave the way for demonstration power plants down the line. One is the China Fusion Engineering Test Reactor (CFETR), a follow-up tokamak to EAST three times the size, which is expected to be built in the late 2020s.

But with climate change looming ever larger, the need to find alternatives to fossil fuels has become more urgent. That has coincided with a flurry of innovation across the fusion industry, aiming to make cheap, sustainable reactors a reality within years, not decades. Most important has been the discovery of superconductors that work at higher temperatures, and so can generate strong magnetic fields with less dramatic refrigeration. Such superconductors allowed magnets to become smaller, and tokamaks to be more compact.

Other recent breakthroughs in technology, ranging from improved construction techniques to robotic systems that can inspect and maintain parts of the reactor, have made it cheaper to get into the fusion business. “It’s gone from being a purely academic activity that only government-funded research labs can fund, to something that private individuals are prepared to invest in,” says Pinches.

This has sparked a race between private companies to be the first to achieve sustainable fusion. One prominent competitor aiming to exploit the same tokamak concepts as JET, ITER and EAST is Commonwealth Fusion Systems. A spin-off from the Massachusetts Institute of Technology, it is partly funded by billionaires including Bill Gates, Jeff Bezos, Jack Ma and Richard Branson and is aiming to produce a reactor within the next 10 years. Other challengers, like Windridge’s Tokamak Energy, are also aiming to provide power to the grid by 2030.

Some are wary of promises given by private companies. “Even with those companies which have been around for a longer time, it’s still the promise of a reactor in 10 years from now,” says Tony Donné, programme manager for EUROfusion, the consortium in charge of JET. Partly, these timescales are given to keep investors happy. “I’m sceptical that they will deliver a fusion reactor much faster than we have,” says Donné. “If the community thought there was an easier way of doing it, they’d be doing that,” says Pinches.

Whether public or private, everyone designing tokamaks is facing the same problems. Chief among them is how to handle instabilities in the plasma. When hot plasma is contained within the magnetic fields of a tokamak, it behaves weirdly. Sometimes small ripples appear like on the surface of a lake, while at others huge tidal waves send the plasma shooting towards the reactor walls. It is enough for some people to seek alternatives to the magnetic confinement technique, which depends on the plasma remaining stable for a long time (see “They do it with lasers”).

Starting in the 1980s, some researchers looking for such alternatives dived into the past, dusting off the long-abandoned stellarators. Their more complex design generated magnetic field patterns capable of stabilising the plasma, says Amitava Bhattacharjee at Princeton University. What’s more, increases in computing power meant it was becoming possible to model how plasma behaved within their more complex configurations, and so potentially create more effective designs. “This produced a renaissance in stellarator research,” says Bhattacharjee. At the same time, new materials and construction methods mean building a stellarator has never been easier.

And while stellarators still lag decades behind tokamaks, they are starting to catch up. In 2015, Wendelstein 7-X, the largest stellarator in the world, was switched on at the Max Planck Institute for Plasma Physics in Greifswald, Germany, and is gearing up to maintain a plasma for 30 minutes, with this milestone expected in 2021. After that, the aim will be to start fusion.

It is still a hugely complex, time-consuming business trying to work out how best to build a fusion reactor, however. “Finding the optimum design of stellarators typically requires playing around with about 50 parameters until the best design is arrived at,” says Bhattacharjee. Plasma instabilities can plague any reactor design, and understanding the complex behaviour involved requires a lot of data and time. “A fully integrated predictive simulation for ITER could take many weeks to run at present,” says Pinches.

That is why, over the past few years, plasma physicists have been turning to a new partner to help haul a sustainable reactor design over the finishing line: machine minds. “Artificial intelligence can give us much greater speed and a much deeper exploration of the range of possibilities,” says Bhattacharjee.

TAE Technologies, a California-based fusion research company, has had a partnership with Google’s DeepMind AI set-up since 2014, while Canada’s General Fusion is working with Microsoft. Improvements are already emerging, says David Ewing at TAE, particularly with regard to modelling how the plasma reacts to different configurations of temperature, density and magnetic field. “Previous to our advancements in machine learning, optimising the performance for a particular experiment set-up could take well over a month,” says Ewing. “These can now be achieved within hours.”

“Calculations that once took a month can now be performed within hours”

Key to this speed-up is AI’s ability to recognise patterns and make predictions about future behaviour. You can’t put a thermometer inside a tokamak to understand its workings, so the temperature has to be inferred from other properties, like how much light is coming out. This can be a difficult task for a human researcher, but an AI trained on mountainous data sets can dramatically cut the time it takes – and also up the efficiency. In 2019, a team at Princeton paired the US’s fastest supercomputer with a neural network to predict plasma disruptions with an unprecedented 95 per cent accuracy.

Artificial intelligence could also be a rocket booster for ITER, too. For some tasks, like modelling the consequences of small ripples in the plasma, AI has already made the job 10 million times faster, says Pinches. Now the key is to boost the speed of the whole simulation, allowing researchers to predict problems and avoid them without needing to run the experiment.

Such innovations, and the speed at which they are now happening, is bringing a new optimism that fusion’s time could, finally, be nearing. “In the last decade, we’ve seen exponential progress in the science,” says Ewing. “That, coupled with the emergence of critical support technologies like AI, has now created the proper tool chest to bring us to the cusp of a breakthrough.” The old joke about fusion hasn’t dated, but this time its backers may have the last laugh.

newscientist.com, 10 June 2020
; https://www.newscientist.com/article/mg24632861-200-why-cracking-nuclear-fusion-will-depend-on-artificial-intelligence/?utm_source=NSNEW&utm_campaign=a4c89c20b9-nsnew_110620&utm_medium=email&utm_term=0_e65fab71ff-a4c89c20b9-378549783”>https://www.newscientist.com