Real-time capable neural network approximation of NUBEAM for use in the NSTX-U control system

摘要

Present-day and next step tokamaks will require precise control of plasma conditions, including the spatial distribution of rotation and current profiles, in order to optimize performance and avoid physics and operational constraints. The coupled nonlinear dynamics of equilibrium profiles and the complex effects of actuators on the equilibrium evolution motivates embedding physics-based and data-driven models within real-time control algorithms. Due to the important role of beam heating, current drive, and torque in establishing scenario performance and stability, a high-fidelity beam model suitable for use in real-time applications is desired. Motivated by the successful application of neural networks for rapidly calculating transport and pedestal pressure [1], this work describes a neural network that has been developed to enable rapid evaluation of the beam heating, torque, and current drive profiles based on measured equilibrium profiles. The training and testing database has been generated from the NUBEAM calculations output from interpretive TRANSP analysis of shots from the 2016 NSTX-U campaign [2, 3], augmented with scans of Z_eff, fast ion diffusivity, beam voltages, and beam modulation patterns. Neural network predictions made for the testing data demonstrate the ability of the model to generalize and accurately reproduce NUBEAM calculated profiles and scalar quantities. Results of processor-in-the-loop simulations of the model within the NSTX-U plasma control system demonstrate the suitability of the approach for real-time use and accelerated offline analysis.