Avoiding fusion plasma tearing instability with deep reinforcement learning.

Area of Science:

• Fusion energy research
• Plasma physics
• Control systems engineering

Background:

• Stable tokamak operation requires active control to prevent plasma disruptions.
• Tearing instability is a primary cause of disruptions, difficult to predict and avoid.
• Previous work developed a dynamic model for predicting tearing instability likelihood.

Purpose of the Study:

• To develop an AI-driven control system for preventing disruptive tearing instabilities in tokamaks.
• To leverage a multimodal dynamic model within a reinforcement learning framework for automated control.
• To demonstrate the AI controller's effectiveness in maintaining stable plasma operations.

Main Methods:

• Developed a multimodal dynamic model to estimate future tearing instability.
• Utilized this model as a training environment for reinforcement learning (AI).
• Implemented and tested the AI controller on the DIII-D tokamak.

Main Results:

• The AI controller successfully reduced the likelihood of disruptive tearing instabilities.
• It maintained tearing instability below a threshold under challenging conditions (low safety factor, low torque).
• The controller enabled stable plasma tracking and H-mode performance, surpassing traditional methods.

Conclusions:

• AI control, trained on a predictive dynamic model, effectively prevents tokamak plasma disruptions.
• This approach facilitates stable, high-performance plasma scenarios for future fusion reactors like ITER.
• Automated instability prevention is crucial for advancing fusion energy production.