Safe Reinforcement Learning-based Controller for Disruption Free Regulation of Plasma Density in Next-generation Tokamaks

S. T. Paruchuri, H. Al Khawaldeh, V. Graber, E. Schuster, A. Pajares, J-W. Juhn

33rd Symposium on Fusion Technology (SOFT)

Dublin, Ireland, September 22-27, 2024

Achieving high particle density is desirable in fusion reactors because of its direct correlation to the fusion power. However, operational limits constrain the maximum achievable particle density. Several density control algorithms have been designed to inject particles using gas puffing and pellet injection to track carefully selected safe targets. However, a controller unaware of these operational limits may modulate the particle densities beyond the safe limits during transients caused by changing operating conditions. Designing plasma control algorithms that recognize these operational limits and ensure that the particle density remains within the safe operational space is desirable. This work focuses on the safe regulation of deuterium, tritium, alpha-particle and impurity densities. Under the assumption of quasi-neutrality, these particle densities can be related to the electron density, which is constrained by the well-known Greenwald's limit. To regulate the particle densities within the safe operational space defined by Greenwald's limit, a novel safe reinforcement learning-based controller is developed in this work. Such control-design approach can be critical in designing learning-based safe plasma control algorithms for next-generation tokamaks. The performance of the controller is evaluated through one-dimensional nonlinear simulations in COTSIM. The results demonstrate the controller's effectiveness under various operating conditions and in the presence of uncertainties.

*Supported by the US DOE (DE-SC0010661, DE-SC0010537).