Neural Network-Enhanced COTSIM Predictive Capabilities for Fast DIII-D Simulations

K. Shabbir, T. Rafiq, E. Schuster

31st IEEE Symposium on Fusion Energy (SOFE)

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

Sustaining fusion reactions in tokamaks requires heating the plasma to thermonuclear temperatures while maintaining confinement and stability. Neutral beam injection (NBI) is used for heating, current drive, plasma rotation, and fueling, while radio frequency waves, such as electron cyclotron range of frequency (ECRF), are employed for heating and current drive. Together, these methods shape the plasma current, temperature, and density profiles. COTSIM, a control-oriented predictive code, has been enhanced with neural-network surrogate models for transport and sources. MMMnet, a neural network version of the updated Multi-Mode Model (MMM 9.0.10), predicts turbulent transport with high accuracy and significantly faster computation times compared to the traditional MMM [1]. Neoclassical transport is calculated using the Chang-Hinton model, while NubeamNet [2], a neural network surrogate for the Monte Carlo NBI module, efficiently predicts NBI effects with reduced computational overhead. To compute electron cyclotron (EC) heating and current drive, a simplified Gaussian beam model has been implemented in COTSIM. This model incorporates the Gaussian beam structure and the cold plasma dispersion relation to calculate the beam trajectory, accounting for weakly relativistic effects. EC power deposition is localized at resonance layers, where the wave frequency matches the electron cyclotron frequency, and is computed using Westerhof's model. Plasma resistivity is modeled using a simplified Spitzer approach, while the non-inductive bootstrap current, essential for steady-state tokamak operation, is calculated using the Sauter model. The equilibrium is solved with a fixed-boundary solver, and the PEDESTAL module determines the H-mode edge pedestal height and width using an empirical model for the L-H transition, based on the power crossing the separatrix threshold. This integrated modeling approach, enhanced by neural-network-based surrogate models, significantly improves COTSIM’s predictive capabilities, enabling fast and accurate predictions of temperature and safety-factor profiles in DIII-D discharges. Comparisons with experimental data from DIII-D demonstrate the model’s accuracy and reliability. These advancements in predictive modeling are essential for advancing tokamak plasma research and optimizing fusion reactor design and operation.

[1] T. Rafiq et al., Phys. Plasmas 20, 032506 (2013).
[2] S.M. Morosohk, et al., Fusion Engineering and Design 163, 112125 (2021).

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