COTSIM: An AI-Powered Tokamak Digital Twin for Predictive Simulation, Scenario Optimization, Plasma Estimation, and Integrated Control

S. T. Rafiq, B. Leard, H. Al Khawaldeh, K. Shabbir, F. Galfrascoli, Y. Tao, V. Graber, Z. Wang, E. Schuster

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

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

Built in MATLAB/Simulink, COTSIM is a modular, AI-accelerated, 1.5D control-oriented integrated tokamak simulation framework for predictive studies, scenario optimization, observer design, and feedback control. It self-consistently couples 1D profile transport (temperature, density, momentum, and current) with fixed- and free-boundary Grad-Shafranov MHD equilibrium evolution over full plasma discharges from ramp-up through ramp-down. Beyond empirical transport models, COTSIM incorporates a neural-network surrogate of the physics-based Multi-Mode Model (MMMnet), together with modular heating, fueling, pedestal, and ELM models. To bridge high-fidelity physics and real-time control, COTSIM embeds additional neural-network surrogates, including NUBEAMnet, TORAYnet, and PEDESTALnet, along with a parameterized SOLPS model, while ongoing work incorporates neural-network surrogates of SOLPS for core-edge integration. This AI-powered architecture achieves orders-of-magnitude speedups, enabling faster-than-real-time simulation while maintaining predictive accuracy. Validated against high-fidelity codes and experimental discharges, COTSIM provides a unified platform for AI-driven pulse optimization, reinforcement learning control training, real-time state estimation from limited and noisy diagnostics, and closed-loop synthesis and testing of integrated actuator-sharing control strategies in next-generation fusion devices.

*Supported by the U.S. DOE under Awards DE-SC0010661 and DE-SC0021385.