The Science
Commercial fusion energy production with a tokamak will require stable and efficient fusion reactions. However, the conditions required for high fusion performance in current devices and devices now under construction, such as ITER, can result in damaging energy bursts. These bursts, called edge-localized modes (ELMs), occur when the plasma escapes confinement by the magnetic fields of the tokamak. ELMs can disrupt the reaction or even damage the machine. One of the main ways to stabilize ELMs is to make small adjustments to the magnetic field. However, this approach has been limited to manual adjustments that rely on preprogrammed responses. In this project, scientists integrated machine learning with adaptive control to achieve real-time adjustment. The controller responded to the constantly changing conditions of fusion plasmas in the DIII-D National Fusion Facility tokamak and the Korea Institute of Fusion Energy’s KSTAR tokamak. The controller achieved comparable performance on these two devices, demonstrating its broad usability.
The Impact
The machine learning-integrated adaptive controller developed in this study is a major technological advance. It produces responsive, real-time control of disruptive ELMs in the plasma edge, addressing a critical challenge in fusion. The controller achieved consistent high fusion performance on two separate tokamaks. This suggests that this approach will be reliable for fusion devices in the years ahead, from ITER to future commercial devices. Its ability to establish stable plasmas without ELMs may also prolong the life of fusion device components, making fusion energy more economically viable.
Summary
In the race to commercialize fusion energy generation, the ability to produce stable plasmas without energy bursts capable of disrupting fusion reactions and damaging machines remains a critical challenge. In this study, researchers combined machine learning, adaptive control, and multi-machine applicability to develop a real-time controller capable of achieving nearly fully edge burst-free plasmas while boosting fusion performance in two separate tokamaks.
This controller functioned by adjusting resonant magnetic field perturbation (RMP), a small change in the magnetic field that alters the plasma edge, in response to real-time conditions in a tokamak. Specifically, the controller finds the optimal RMP amplitude and spectrum for robust ELM suppression without harmful effects, supported by a machine learning algorithm, the surrogate version of the plasma response model. Ultimately, this approach enabled consistent achievement of the highest fusion performance on the tested devices with minimized confinement loss.
The controller also drove a novel 3D field waveform in time, which led to the discovery of a coupling of plasma flow and edge transport that can further contribute to plasma confinement. This result suggests a broader applicability for artificial intelligence-integrated control; it can enable scientists to explore new states that have not been accessed with traditional manual approaches, which may lead to the development of new optimal states for commercial fusion. Thus, this controller and future iterations can pave the path to developing burst-free high-performance plasma scenarios for future fusion devices.
Funding
This material is based on work supported by the Department of Energy (DOE) Office of Science, Office of Fusion Energy Sciences, using the DIII-D National Fusion Facility, a DOE Office of Science user facility. This research was also supported by the KSTAR Experimental Collaboration and Fusion Plasma Research (EN2401-15) R&D Program of the Korea Institute of Fusion Energy.
