Modification of a machine learning‐based semi‐empirical turbulent transport model for its versatility

Abstract

AbstractA machine learning‐based semi‐empirical turbulent transport model DeKANIS has been modified to apply it independently of the device. DeKANIS predicts particle and heat fluxes, distinguishing diffusive and non‐diffusive transport processes. DeKANIS consists of a neural network (NN) model, which computes coefficients of the non‐diffusive terms and the ratio of the fluxes based on the gyrokinetic calculations, and a scaling formula, which estimates the turbulent saturation level founded on empirical fluxes. The datasets used for NN training have been prepared based on JT‐60U plasmas so far, but by exploiting JET plasmas, the datasets have been expanded and the parameter ranges covered by the NN models have become wider. The scaling formula has been rebuilt considering the decrease in the residual zonal flow level due to collisions. The new DeKANIS has demonstrated a reasonable profile prediction of an ITER plasma in the pre‐fusion power operation 1 phase with an integrated model GOTRESS+. In validating the prediction results with the gyrokinetic calculations, transport processes causing the fluxes have been exhibited.