Machine-learning assisted transport modeling. Practical cases in JT-60U

Since fusion plasma is dominated by physical phenomena in a wide range of spatiotemporal scales, an integrated model consisting of module groups dealing with physical phenomena on each spatiotemporal scale is indispensable for the performance prediction and physics elucidation of the plasma. The transport module, as one of the constituents, deals with the macroscopic time evolution of plasma. Here, the turbulent transport model was used to predict the turbulent flux, but the computational cost increased due to the refinement of the physical model, which lowered the computational speed of the entire integrated model. As a countermeasure, transport models have been developed in Japan and overseas through a data-driven scientific method that significantly reduces the computational cost while maintaining the reproducibility of physics phenomena. This paper introduces an artificial neural network (NN) model based on the data obtained by the first-principles calculation and the JT-60U experiments, and an NN surrogate model simulating the transport model that was constructed by learning an enormous amount of transport calculation data generated during the calculation of the global optimization method. It also shows the results of the transport simulation performed for the JT-60U experiments. (A.O.)