Accelerated predictive modeling of the current profile evolution on NSTX-U using neural networks

Fast, real-time modeling of data will be vital for designing experiments and simulations for present-day and future fusion devices like ITER. The modeling presented here focuses on the rapid evaluation of terms needed to evolve the magnetic diffusion equation for current profile prediction. A neural network has been developed to model plasma conductivity, bootstrap current, and flux surface averaged geometric quantities. The model drew from a database of 2016 NSTX-U TRANSP runs and used dimensionality reduction and an optimization algorithm to best select inputs, outputs, and hidden layer sizes. A fully-connected neural network topology was used, and multiple models were trained to maximize profile prediction. Comparison of the models to test data shows that they can closely reproduce calculated profiles and scalar quantities relevant to the evolution of the magnetic diffusion equation. Combined with the recently developed NubeamNet model for beam current drive, these models demonstrate progress towards real-time simulation of NSTX-U current profile evolution that can account for changes in plasma shaping.