MPC-Guided Policy Learning for Real-Time Tokamak Density Control

H. Al Khawaldeh, S. T. Paruchuri, T. Rafiq, E. Schuster

68th Division of Plasma Physics (DPP) Annual Meeting of the American Physical Society (APS)

Chicago, IL, USA, November 2-6, 2026

Density-control frameworks for reactor-grade tokamaks must maintain reliable performance across a broad range of plasma conditions. Meeting this requirement calls for predictive models that capture variations in the density response throughout the operating space. However, incorporating comprehensive transport models directly into real-time control frameworks can impose a substantial computational burden. To balance physical fidelity with control execution speed, synthetic data generated by the physics-based Control-Oriented Transport SIMulator (COTSIM) is used to identify a data-driven response model. The model is then integrated into a Model Predictive Control (MPC) framework to compute optimal fueling commands. While this MPC framework delivers strong closed-loop performance, solving the associated optimization problem online may limit its real-time applicability in future multi-objective integrated-control architectures, where additional objectives, such as detachment control, increase the optimization dimension and, consequently, the computation time. To proactively ensure real-time feasibility, a control policy is trained through interactive imitation learning using measurement-action pairs generated by the MPC during simulations. Across the evaluated operating scenarios, the learned policy matches MPC performance while significantly reducing execution time. These results demonstrate a scalable pathway toward real-time, multi-scenario plasma control.

*Supported by the U.S. DOE under Award DE-SC0010661.