Predicting plasma kinetic profiles with Gaussian process and a neural network in KSTAR based on magnetic and heating information

Abstract

I expected that inferring plasma status with magnetic signals would be especially important when only a few diagnostics such as magnetic field probes would be available at nuclear fusion power plants. I conducted preparatory research predicting kinetic profiles in the KSTAR with magnetic and heating information using an artificial neural network. I implemented Gaussian process regression (GPR) for kinetic profile reconstruction, and results - where GPR kernel function's hyper-parameters are predicted with the Maximum a posteriori (MAP) estimator and marginalized with the No-U-turn sampler (NUTS) - are compared. To get the spatially continuous neural network prediction, I conducted the principal component analysis (PCA) on the kinetic profiles for the dimensionality reduction. In addition, I adopted support vector machine regression (SVMR) to handle data outliers. I report the KSTAR ion temperature prediction results utilizing all the methods mentioned above.