Optimization and application of neural network models for accelerated predictive modelling of NSTX-U

A critical component of advancing modern fusion devices to cost-effective energy sources will be model-based control and scenario development. Specifically, a hierarchy of models of varying complexity, speed, and precision is needed to meet all the needs of the design process. Neural networks have proven capable of producing fast and reliable models of fusion plasmas, and their levels of complexity can be easily tuned. However, each network has many hyperparameters, the tuning of which is time-consuming and does not always result in the best possible model. An algorithm has been developed to systematically generate and optimize neural networks of varying complexities by tuning hyperparameters for the purposes of accelerated predictive models of NSTX-U. It uses a genetic algorithm to rapidly arrive at optimal model parameters that could previously be found only through exhaustive grid searches. This algorithm was used to tune and train neural networks of neutral beam deposition based on the NUBEAM code. This new approach yields more optimal neural networks and enables tuning the trade-off of model fidelity and computation time based on the requirements of any application.