Yes, true, fusion reactors don't work quite yet, but, er, maybe AI can help us stop our experiments from imploding

No neural networks here, though, only decision tree algos


Machine-learning algorithms may help scientists develop a warning system capable of preventing fusion reactor meltdowns in the future.

Scientists have attempted to harness the power of nuclear fusion for decades in the hopes of providing clean energy. Clouds of plasma made up of positively and negatively charged ions are placed in a donut-shaped chamber and controlled with superconducting electromagnets in tokamak constructions. As the strength of the magnetic field is cranked up, the ions whiz around the reaction chamber coming into close contact with one another to fuse. Thermal energy is released in the process.

Tokamaks, however, are temperamental. A lack of control can lead to a sudden release of the heat generated. Disruptions like these can melt the surface of the reaction chamber. Now, researchers led by Princeton Plasma Physics Laboratory (PPPL), a fusion research lab working under the US Department of Energy (DoE), are attempting to build a control system that monitors and predicts the likelihood of disruption events unfolding during the operation of tokomaks using machine-learning software.

Specifically, a paper published in the American Institute of Physics's Physics of Plasma journal describes using decision-tree algorithms. Details from various fusion reactors, such as their plasma current, plasma energy, radiation power, strength of radial magnetic field, and more are taken into account and used as inputs into the algorithm.

"Using 10 signals at different time moments, we made use of a broader range of signals and calculated the mean, trend, and variance within a specific time window," the paper said.

If the value of each signal is above a certain threshold, it increases the chance of a meltdown. All these different scenarios for each input are run through the decision-tree algorithm until it reaches the last two branches that predict whether there will be a disruption event or not within the next 250 milliseconds.

That's an incredibly short amount of time, so the results from the algorithm will probably have to be fed into a wider diagnostic system that automatically controls the tokamak rather than relying on human intervention. "In the real-time experiments we did, the algorithms successfully predicted the disruption," Yichen Fu, a graduate student, and Egemen Kolemen, an assistant professor at Princeton University, who both work at PPPL and are co-authors of the study, told The Register. "We used the prediction of our algorithm as a control input to dynamically adjust neutral beam power so that we could reach the best performance while avoiding instabilities."

The researchers took data from thousands of experiments run on the DoE's DIII-D National Fusion Facility operated by General Atomics, an energy and defense biz in San Diego, to train their algorithms. After they tested their methods, they found it was generally accurate approximately 80 per cent of the time.

Although the algorithms are promising, the study is more of a proof of concept and won't be used for real tokomaks anytime soon. "This work represents significant progress in the use of machine learning to develop a disruption prediction and avoidance method in fusion devices," said Raffi Nazikian, who heads the ITER and Tokamaks Department at PPPL. ®


Biting the hand that feeds IT © 1998–2020