• Highly enhanced plasma confinement, reaching the highest fusion G in the Edge Localized Mode-free (ELM-free) scenario of the two machines, with G increased by up to 90%;

• First fully automated 3D field optimization using ML-based 3D field simulator;

• Simultaneous establishment of burst suppression from the beginning of plasma operation , achieving almost complete edge-free burst operation close to ITER-related levels. This achievement represents a crucial step for future devices such as the International Thermonuclear Experimental Reactor (ITER), where reliance on empirical RMP optimization is no longer a feasible or acceptable approach.

"There are instabilities in the plasma that can cause serious damage to the fusion device. We cannot use these substances in commercial fusion vessels. Our work advances the field and And shows that artificial intelligence can play an important role in managing fusion reactions to avoid instability while allowing the plasma to generate as much fusion energy as possible," said corresponding author Egemen Kolemen, associate professor in the Department of Mechanical and Aerospace Engineering at PPPL.

Fully Automated ML-Based 3D Field Optimization

In this experiment, a series of discharges are used to find optimized 3D waveforms for safe ELM suppression.

In this context, the study introduces ML techniques to develop novel paths for automated 3D coil optimization and demonstrates the concept for the first time.

Researchers developed a surrogate model of the GPEC code (ML-3D) to leverage physics-based models in real time. The model uses ML algorithms to accelerate computation time to the ms level and is integrated into the adaptive RMP optimizer in KSTAR.

ML-3D consists of a fully connected multilayer perceptron (MLP) driven by nine inputs. To train the model, 8490 KSTAR balanced GPEC simulations were utilized.

The algorithm utilizes the ELM status monitor (Dα) signal to adjust the IRMP in real time, which can maintain sufficient edge 3D fields to access and maintain ELM suppression. At the same time, the 3D field optimizer uses the output of ML-3D to adjust the current distribution on the 3D coil, thus ensuring a safe 3D field to avoid interruptions.

In KSTAR experiments, the ML-integrated adaptive RMP optimizer triggered in 4.5 seconds and achieved safe ELM suppression in 6.2 seconds.

Research also demonstrates 3D-ML as a viable solution for automating ELM-free access. ML-3D is based on physical models and does not require experimental data, making it directly scalable to ITER and future fusion reactors. This strong applicability to future devices highlights the advantages of ML's integrated 3D field optimization approach. Furthermore, better field optimization and higher fusion performance are expected to be achieved in future devices with higher 3D coil current limitations.

Research successfully optimized controlled ELM-free states in KSTAR and DIII-D devices with highly enhanced fusion performance, covering low-n RMP related to future reactors to nRMP related to ITER = 3 RMP, and achieved the highest level in various ELM-free scenarios in two machines.

In addition, the innovative integration of ML algorithms with RMP control enables fully automatic 3D field optimization and ELM-free operation for the first time, and with the support of adaptive optimization processes, performance is significantly enhanced . This adaptive approach demonstrates compatibility between RMP ELM suppression and high limits.

Additionally, it provides a robust strategy to achieve stable ELM suppression in long pulse scenarios (lasting more than 45 seconds) by minimizing the loss of limiting and uninductive current fractions.

Notably, significant performance (G) improvement was observed in DIII-D with nRMP = 3 RMP, showing an improvement of more than 90% from the initial standard ELM suppression state. This enhancement is attributed not only to the adaptive RMP control but also to the self-consistent evolution of the plasma rotation. This response enables ELM suppression at very low RMP amplitudes, thereby enhancing the base. This feature is a good example of a system transitioning to an optimal state through a self-organized response to adaptive modulation.

In addition, the adaptive scheme is combined with the early RMP-ramp method to achieve ITER-related ELM-free scenarios with almost completely ELM-free operation. These results confirm that integrated adaptive RMP control is a very promising approach to optimize ELM suppression states, with the potential to address one of the most difficult challenges in achieving practical and economically viable fusion energy.

Reference content: https://phys.org/news/2024-05-ai-intensive-aspects-plasma-physics.html

【 Recommended reading]

• Log into Nature and predict the "tearing" of plasma in fusion 300 milliseconds in advance. The Princeton team develops an AI controller
  

• Quickly filter massive data and make informed decisions instantly, MIT, Princeton & Carnegie Mellon University team uses LLM for fusion research

• #Machine learning to find the right combination of hydrogen isotopes for future nuclear fusion power plants

• The simulation accuracy is increased by 65%, the training time is reduced by 3 times or more, and DeepMind has made progress in controlling nuclear fusion with reinforcement learning