Reinforcement Learning-based Density Profile Control with Active Stability-limit Enforcement in Tokamaks

S. T. Paruchuri, V. Graber, E. Schuster

31st IEEE Symposium on Fusion Energy (SOFE)

Cambridge, MA, USA, June 23-26, 2025

Achieving high plasma density is crucial for maximizing fusion power output in tokamak devices like ITER and the Fusion Pilot Plant (FPP). However, increasing the density through methods like pellet injection and gas puffing raises the risk of exceeding stability limits, leading to the onset of magnetohydrodynamic (MHD) instabilities and disruptions. Therefore, enforcing stability constraints while regulating plasma density is critical. Conventional density-control approaches rely on using pre-designed density profiles that satisfy the density limits as references for feedback control. However, real-time changes in plasma operating conditions could shift the stability limits themselves, leading to the immediate breach of the stability limits and compromising safety. This work proposes a reinforcement learning-based controller that regulates the plasma density profile while actively enforcing a pre-selected stability limit. The controller prioritizes tracking a reference density profile when the stability limit is satisfied. However, whenever the stability limit shifts, it dynamically adjusts the density profile to avoid violations, sacrificing reference tracking if required. Two controller versions, enforcing the Greenwald limit and edge stability limit, were trained in a control-oriented one-dimensional environment. Feedback simulations demonstrate the controller’s effectiveness in regulating the density profile and its robustness in satisfying the stability limit.

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