Researchers are now deploying machine learning algorithms directly into the control systems of tokamak reactors to prevent sudden plasma collapses. These collapses, known as tearing modes, reconfigure magnetic field lines and break the symmetry required for stable operation. 

This development marks a transition from reactive safety measures to predictive, automated stabilization of the ultra-hot plasma necessary for nuclear fusion energy.

The tokamak is a donut-shaped device that uses magnetic fields to confine plasma. Within this environment, tearing modes form at specific locations called rational surfaces. 

“Tearing modes (TMs)—slow growing, resistive magnetohydrodynamic (MHD) plasma perturbations with resonant modal structure—are one of the primary instabilities threatening the viability of the tokamak power plant,” said the researchers in a new study.

The formation of these modes is driven by a complex balance of stabilizing and destabilizing effects. When this balance is lost, a magnetic bubble grows within the plasma. 

This growth slows the rotation of the plasma column and eventually causes the plasma to strike the reactor wall, ending the fusion process. In an unmitigated state, this magnetic bubble grows like a slug, grinding the system to a halt and dispersing the plasma.

Solution for real-time control

Traditional physics models struggle to manage these events because the underlying mechanisms are nonlinear and chaotic. Small, rapid instabilities in one part of the rotating plasma can trigger a tearing mode elsewhere. 

To address this, researchers Cristina Rea and Stuart Benjamin used machine learning to process large volumes of experimental data from previous tokamak operations to identify the precursors of these instabilities. 

These algorithms detect patterns that indicate an imminent tearing mode before it becomes visible to standard diagnostic tools.

These AI-based predictors are now being used to develop active plasma controllers. These controllers receive real-time data from the reactor and use machine learning to determine the stability of the plasma at any given moment. 

If the system detects a risk of tearing onset, the controller automatically adjusts the magnetic configuration to suppress the mode or avoid the conditions that cause it. 

This capability is essential for the International Tokamak Physics Activity, which is currently designing a trigger for the disruption mitigation system of the ITER project.

To maintain plasma stability

The demand for these AI controllers increases as fusion experiments move toward higher plasma pressures. While higher pressure is necessary for efficient energy production, it also increases the frequency and severity of tearing modes. 

By integrating machine learning into the operational core of the tokamak, engineers can maintain the stability required for continuous power generation.

“TMs remain fiendishly hard to predict with physics models, but their stochastic complexity appeals to ML-empowered scientists,” concluded Benjamin.

“That’s why we wanted to explain how recent studies have applied AI to large experimental tokamak datasets, providing new insights into the TM physics and control mechanisms we must perfect to ensure TMs don’t compromise the tokamak power plants of the future.”

These controllers essentially act as an automated steering system, keeping the plasma within the narrow parameters required for a sustained fusion reaction.