Machine-learning assisted steady-state profile predictions using global optimization techniques

Abstract

Predicting plasma profiles with a stiff turbulent transport model is important for experimental analysis and development of operation scenarios. Due to the sensitivity of turbulent fluxes to profile gradients, robust predictions are still arduous with a stiff model incorporated in a conventional transport code. With global optimization techniques employed, the new steady-state transport code, global optimization version of the transport equation stable solver, has been developed to overcome these difficulties. It enables us to attain smooth profiles of diffusivity and temperature even though jagged profiles thereof are inclined to emerge in simulations with a stiff model. A neural-network-based surrogate model of a transport model is developed to compensate slow computation inherent to global optimization. Hyperparameter optimization realizes the surrogate model with very good accuracy.