Machine learning mitigates tearing modes in plasmas tori using real-time feedback

Tokamaks rely on high-powered magnets to confine plasmas into the shape of a torus. However, modern tokamaks looking to increase fusion gain risks increasing the probability of destabilizing the plasma. To prevent disruptions and damage to reactors, researchers seek to predict such instabilities far enough in advance to control and mitigate them.

Fu et al. developed a real-time feedback control based on machine learning algorithms that successfully avoided tearing modes and disruptions in DIII-D plasmas. To maximize performance in the simulations without any human pre-programmed design, their algorithms used ensemble learning methods to process the data in the real-time Plasma Control System coming from several thousand DIII-D discharges.

“For the first time, machine learning control was implemented on a fusion reactor,” said author Egemen Kolemen. “We believe that the approach outlined in this work will form the basis for a much more significant part of the fusion reactor control going forward.”

The approach produces “tearability” and “disruptivity” metrics that quantify the likelihood of 2/1 tearing modes in a given time window and the likelihood of plasma disruptions. The algorithms continuously feed the results back to the system control and help it maximize normalized plasma beta while avoiding tearing modes and high-current disruptions in ramp down.

The group looks to expand their work for describing a full plasma evolution model only from data, including such features as ion and electron temperature and density, and plasma rotation. Using extensive computation resources at the National Energy Research Scientific Computing Center, they hope to achieve the combination of high performance and high stability in fusion reactors.

Source: “Machine learning control for disruption and tearing mode avoidance,” by Yichen Fu, David Eldon, Keith Erickson, Kornee Kleijwegt, Leonard Lupin-Jimenez, Mark D. Boyer, Nick Eidietis, Nathaniel Barbour, Olivier Izacard, and Egemen Kolemen, Physics of Plasmas (2019). The article can be accessed at https://doi.org/10.1063/1.5125581 .