Fusion reactors promise cheap, abundant and relatively clean energy – if we
can get them to work. Now, thanks to artificial intelligence firm DeepMind,
fusion researchers are one step closer to extracting power from plasma
hotter than the surface of the sun.
DeepMind worked with scientists at the Swiss Federal Institute of Technology
in Lausanne (EPFL), Switzerland, to create a neural network capable of
controlling the magnetic fields within EPFL’s Variable Configuration Tokamak
(TCV) fusion reactor.
These magnetic fields are essential for keeping the plasma generated by the
reactor safely contained. If the plasma touches the walls of the reactor it
rapidly cools, stifling the reaction and potentially causing significant
damage.
Researchers at the TCV previously used 19 magnetic coils, each controlled by
a separate algorithm that monitored the interior of the reactor thousands of
times a second with a host of sensors. DeepMind instead created a single
neural network to control all the coils at once, automatically learning
which voltages needed to be supplied to them to best contain the plasma.
The team trained the AI on a precise digital simulation of the reactor
before conducting experiments on the real machine. Ultimately, it was able
to successfully contain the plasma for around 2 seconds, which is
approaching the limits of the reactor – TCV can only sustain the plasma in a
single experiment for up to 3 seconds, after which it needs 15 minutes to
cool down. The record for fusion reactors is only 5 seconds, set recently by
the Joint European Torus in the UK.
As well as controlling the plasma, the AI was able to shape it and move it
around within the reactor. New plasma shapes may bring efficiency or
stability improvements in new fusion reactors such as ITER, which is
currently being built in France and will be the world’s largest tokamak when
complete in 2025. The AI even demonstrated the ability to control two
separate beams of plasma at once.
Federico Felici at EPFL says that although there are many theoretical
approaches that could be used to contain the plasma with a magnetic coil,
scientists have tried-and-tested strategies. But the AI surprised the team
with its novel approach to forming those same plasma shapes with the coils.
“This AI algorithm, the reinforcement learning, chose to use the TCV coils
in a completely different way, which still more or less generates the same
magnetic field,” says Felici. “So it was still creating the same plasma as
we had expected, but it just used the magnetic cores in a completely
different way because it had complete freedom to explore the whole
operational space. So people were looking at these experimental results
about how the coil currents evolve and they were pretty surprised.”
Gianluca Sarri at Queen’s University Belfast, UK, says that AI is key to the
future of control systems for fusion reactors, which have yet to sustain a
reaction that produces more power than is consumed.
“Once this is done, this is not the end of the story. Then you have to make
it a power plant,” he says. “And this AI is, in my opinion, the only way
forward. There are so many variables, and a small change in one of them can
cause a big change in the final output. If you try to do it manually, it’s a
very lengthy process.”
To make fusion reactors efficient, practical power sources, physicists need
to increase the ratio between the pressure of the plasma and the power of
the magnetic fields containing it, a value called beta, says Howard Wilson
at the University of York, UK.
“The plasma writhes and wiggles and tries to escape the clutches of the
magnetic fields, and as one is pushing up that beta parameter, one is having
to work harder and harder to get the control that one needs to just hold the
plasma there,” he says. “The further you push the plasma, the more chance
you just suddenly lose it.”
Wilson believes these AI experiments show promise for containing plasma in
“extreme geometries”, which paves the way for experiments with different
plasma shapes that might yield improvements in stability or efficiency. “It
makes the risky parameter space less risky to operate in, but also opens up
new parameter space that we can go into and explore,” he says.
Reference:
Degrave, J., Felici, F., Buchli, J. et al. Magnetic control of tokamak plasmas
through deep reinforcement learning. Nature 602, 414–419 (2022).
DOI: 10.1038/s41586-021-04301-9
Tags:
Physics