Tokamaks are machines whose task is to maintain and harness the power of the sun. These fusion machines use powerful magnets to trap plasma hotter than the Sun's core and push the plasma atoms to melt and release energy. If tokamaks can operate safely and efficiently, the machines will one day be able to deliver clean and unlimited fusion energy.
There are currently many experimental tokamaks in operation around the world, with more being developed. Most of them are small research machines built to investigate how devices can spin plasma and harness its energy. One of the challenges facing tokamaks is to safely and reliably turn off the plasma current circulating at speeds of up to 100 kilometers per second and temperatures exceeding 100 million degrees Celsius.
Such “slowdowns” are necessary when the plasma becomes unstable. To prevent the plasma from further interfering with and potentially damaging the interior of the device, operators reduce the plasma current. Sometimes, however, the slowdown itself can destabilize the plasma. In some machines, the drops have caused scratches and scars on the interior of the tokamak – minor damage that, however, requires significant time and resources to repair.
Now, MIT scientists have developed a method to predict the behavior of plasma in a tokamak during braking. The team combined machine learning tools with a physics-based plasma dynamics model to simulate the behavior of the plasma and any instabilities that might arise from shrinking the plasma on and off. Scientists trained and tested the new model on plasma data from an experimental tokamak in Switzerland. They found that the method quickly learned the evolution of the plasma as it was adjusted in various ways. Moreover, the method achieved a high level of accuracy using a relatively small amount of data. This training efficiency is promising, considering that each experimental tokamak launch is expensive and, as a result, data quality is limited.
The new model the team is spotlighting this week in publicly available Nature communication papercould improve the safety and reliability of future fusion power plants.
“For fusion to be a useful energy source, it must be reliable,” says lead author Allen Wang, an aeronautics and astronautics graduate student and member of Disruption Group at MIT's Plasma Science and Fusion Center (PSFC). “To be reliable, we need to learn to manage our plasma well.”
MIT study co-authors include PSFC principal investigator and Disruptions Group leader Cristina Rea and Laboratory Information and Decision Systems (LIDS) members Oswin So, Charles Dawson and Professor Chuchu Fan, as well as Mark (Dan) Boyer of Commonwealth Fusion Systems and colleagues at the Swiss Plasma Center in Switzerland.
“A Delicate Balance”
Tokamaks are experimental thermonuclear devices first built in the Soviet Union in the 1950s. The device takes its name from a Russian acronym that translates as “toroidal chamber with magnetic coils.” As its name suggests, a tokamak is toroidal, or doughnut-shaped, and uses powerful magnets to store and accelerate gas to a temperature and energy high enough for the atoms of the resulting plasma to fuse and release the energy.
Currently, tokamak experiments are conducted on a relatively small scale, with few achieving the size and power necessary to produce safe, reliable and useful energy. Interference in experimental, low-energy tokamaks is usually not a problem. However, as fusion machines scale up to grid size, controlling the much higher energy plasma in all phases will be critical to maintaining the safe and efficient operation of the machine.
“Uncontrolled plasma terminations, even during braking, can generate intense heat fluxes damaging the internal walls,” notes Wang. “Quite often, especially with high-performance plasmas, dips can actually push the plasma closer to certain instability limits. So it's a delicate balance. There's a lot of emphasis right now on how to manage the instabilities so that we can routinely and reliably pick up these plasmas and turn them off safely. And relatively little research has been done on how to do this well.”
Lowering the pulse
Wang and his colleagues developed a model to predict the behavior of the plasma during the descent of the tokamak. While they could simply apply machine learning tools like a neural network to detect signs of instability in the plasma data, “it would take an ungodly amount of data” for such tools to spot the very subtle and ephemeral changes in the extremely high-temperature, high-energy plasma, Wang says.
Instead, the researchers combined the neural network with an existing model that simulates plasma dynamics according to basic principles of physics. Using a combination of machine learning and physics-based plasma simulation, the team found that just a few hundred pulses at low efficiency and a small handful of pulses at high efficiency were enough to train and validate the new model.
The data used in the new study came from TCV, a Swiss “variable configuration tokamak” operated by the Swiss Plasma Center at EPFL (Swiss Federal Institute of Technology Lausanne). A TCV is a small experimental fusion device used for research purposes, often as a test bed for next-generation device solutions. Wang used data from several hundred TCV plasma pulses, which included plasma properties such as its temperature and energies during the rise, run and fall of each pulse. Using this data, he trained a new model, then tested it and discovered that he could accurately predict the evolution of the plasma given the initial conditions of a specific tokamak pass.
Researchers also developed an algorithm to translate the model's predictions into practical “trajectories,” or plasma management instructions, that the tokamak controller can automatically execute, for example to adjust magnets or temperature to maintain plasma stability. They implemented the algorithm on several TCV runs and found that it produced trajectories that safely reduced the plasma pulse, in some cases faster and without disruption compared to runs without the new method.
“At some point, the plasma will always decay, but we call it a disruption when the plasma decays with high energy. In this case, we reduced the energy to zero,” Wang notes. “We did this several times. And overall, we did much better. So we had statistical certainty that we had improved the situation.”
The work was supported in part by Commonwealth Fusion Systems (CFS), an MIT spinoff that intends to build the world's first grid-scale compact fusion power plant. The company is developing the SPARC demonstration tokamak, designed to produce net-energy plasma, meaning it should generate more energy than is needed to heat the plasma. Wang and his colleagues are working with CFS on ways the new predictive model and similar tools can better predict plasma behavior and prevent costly disruptions to ensure safe and reliable fusion power.
“We're trying to solve scientific problems to make thermonuclear fusion routinely useful,” Wang says. “What we have accomplished here is the beginning of what is still a long journey. But I think we have made a lot of progress.”
Additional research support came from the EUROfusion consortium, through the Euratom Research and Training Program, and was financed by the Swiss State Secretariat for Education, Research and Innovation.
