Tokamaks is a machine that helps to retain and utilize the power of the sun. These fusion machines use powerful magnets to contain plasmas hotter than the solar core, and force the plasma atoms to fuse and release energy. If Tokamaks can operate safely and efficiently, the machine can one day provide clean and endless fusion energy.
Today, there are many experimental tokamaks in operation around the world, and more is underway. Most are small research machines built to investigate how devices spin up plasma and utilize that energy. One of the challenges facing Tokamaks is how to safely and reliably turn off plasma currents circulating at temperatures above 100 million degrees Celsius at speeds of up to 100 km per second.
This “ramp down” is necessary when the plasma becomes unstable. The operator runs down the plasma current to prevent the plasma from destroying further inside the device and potentially damaging it. However, sometimes the ramp down itself can make the plasma unstable. On some machines, ramp down causes scuffs and scars inside the tokamac. This requires a considerable amount of time and resources to repair.
Now, MIT scientists have developed a method to predict how tokamac plasma will behave during ramp down. The team combined machine learning tools with a physics-based model of plasma dynamics to simulate plasma behavior and instability that can occur as the plasma increases and turns off. Researchers trained and tested a new model of plasma data for experimental tokamacs in Switzerland. They discovered that this method quickly learned how plasma evolves when it was tuned in a variety of ways. Furthermore, this method achieved a high level of accuracy using relatively small amounts of data. This training efficiency is promising given that it is expensive for each experimental run of Tokamak and, as a result, limited high-quality data.
The new model the team highlights in open access this week Natural Communication Paper may improve the safety and reliability of future fusion power plants.
“Fusion must be reliable to be a useful source of energy,” says Allen Wang, a graduate student in aerospace and asthma studies and a lead author who is a member of the destruction group at MIT's Plasma Science and Fusion Center (PSFC). “To be reliable, you need to be good at managing your plasma.”
MIT co-authors of this study include Christina Lea, a PSFC Principal Research Scientist and Confusion Group Leader, and members of Professor Oswin (LID) of Information and Decision Systems (LIDS), along with Commonwealth Fusion Systems and Collaborators Mark (Dan) Boyer at Swiss Plasma Centre in Switzerland.
“Subtle balance”
The Tokamaks is the first experimental fusion device built in the Soviet Union in the 1950s. This device gets its name from the Russian acronym, translated into “Toroidal Chamber with Magnetic Coils.” As the name explains, tokamacs are toroidal or doughnut-shaped, using powerful magnets to contain and spin up the gas, allowing the resulting plasma atoms to fuse and release energy.
Today, Tokamak's experiments are relatively low energy, and few people are approaching the size and power needed to produce safe, reliable, usable energy. In general, experimental, low-energy tokamak confusion is not a problem. However, as the fusion machines expand to grid-scale dimensions, controlling the much higher energy plasma at all stages is paramount to maintaining the safe and efficient operation of the machine.
“Uncontrolled plasma terminations can produce intense heat fluxes that damage the inner walls, even during ramp down,” Wang points out. “In particular with high-performance plasmas, ramp down actually allows the plasma to approach some unstable limitations. So it's a delicate balance. And there's a lot of focus on how to take these plasmas regularly and reliably and manage instability in the way they safely power down.
Lower the pulse
Wang and his colleagues have developed a model to predict how plasma will work during the ramp down of the tokamak. Although machine learning tools such as neural networks could be simply applied to learn indications of instability in plasma data, “we'll need a profane amount of data,” Wang says, to identify very subtle and short-lived changes in very hot, high-energy plasmas.
Instead, researchers combined existing models with neural networks that simulate plasma dynamics according to basic rules of physics. This combination of machine learning and physics-based plasma simulations led the team to discover that only hundreds of pulses at low performance and a small number of high performance pulses are sufficient to train and validate the new model.
The data they used in their new research come from TCV, a Swiss “variable configuration Tokamak” run by the Swiss Plasma Centre at the EPFL (Lausanne Swiss Federal Institute of Technology). TCVs are small experimental fusion experimental devices that are often used for research purposes as testbeds in next-generation device solutions. Wang used data from hundreds of TCV plasma pulses, including plasma characteristics such as temperature and energy during ramp up, run and ramp down for each pulse. He trained a new model of this data, tested it, and found that it could accurately predict the evolution of the plasma, taking into account the initial conditions of a particular tokamac performance.
The researchers also developed algorithms that translate model predictions into practical “trajectories” or plasma management instructions. For example, a plasma management instruction that a tokamac controller can be automatically executed to maintain the stability of the plasma to regulate the magnet or temperature. They have implemented algorithms in some TCV runs and found that, in some cases, generate trajectories that safely increase plasma pulses without confusion, compared to running without new methods.
“At some point, the plasma always disappears, but when the plasma disappears at high energy, we call it a mess. Here, we did nothing of energy,” Wang points out. “We did that many times, and we all made things better, and so we had the statistical confidence that we made things better.”
This work was supported in part by Commonwealth Fusion Systems (CFS), the MIT spin-out aimed at building the world's first compact, grid-scale fusion power plant. The company is developing DeMotocamac SPARC, designed to generate net energy plasmas. This means it should generate more energy than it takes to heat the plasma. Wang and his colleagues are working with CFS on how new predictive models and tools can better predict plasma behavior and prevent costly confusion to enable safe and reliable fusion forces.
“We're trying to tackle science questions and make fusion useful on a daily basis,” Wang says. “What we did here is still the beginning of a long journey. But I think we've made great progress.”
Additional support for research was funded by the Swiss State Secretariat for education, research and innovation from the framework of the Eurofusion Consortium through the Euroatom Research Training Program.
