Development of transport coefficient calculation algorithm for tokamak plasma impurity using neural network technique

Abstract

There are inevitably inherent impurities injected into plasma from plasma facial components such as C, W, Be, and gas impurities such as N2, Ne, Kr that are injected externally for plasma and thermal flux control within the currently under research and development. Although impurity conducts several positive effects in fusion plasma, the negative effects like plasma cooling, fuel gas dilution, plasma disruption can occur if impurity density in core plasma is too high. Analysis of transport of impurity in plasma is essential for the realization of a future fusion reactor. In plasma, impurity transport is usually described by obtaining impurity transport coefficients such as diffusion coefficients (D) and convection velocity (V), and based on diagnostic results obtained from impurity diagnostic systems, impurity transport codes such as KIM, UTC-SANCO, STRAHL, etc. The impurity transport code firstly, numerically calculates the continuity equation from the inferred impurity transport coefficient to obtain the impurity concentration in the plasma over time. Next, we compute virtually observed data from real diagnostic systems when the impurity concentration is calculated in the previous process. We compare this with the diagnostic data obtained by actual observations and repeat how to correct the transport coefficients in a way that minimizes their error and recalculates them to compare them. To obtain transportation coefficients within the margin of error compared to the diagnosis results, repeated calculations are essential, and existing methods have the disadvantage of complicated repetition procedures and take a very long time to calculate. In this thesis, we develop new artificial neural networks trained by accumulated databases from impurity transport codes, and the developed algorithm that has faster and comparable accuracy compared to conventional ones. The algorithm consists of converting diagnostic data to impurity concentrations via radiative cooling coefficients and converting time-series impurity concentrations to a peaking factor (-V/D) of impurities through the computation of artificial neural network models. We trained the artificial neural network model based on training data generated through KIM (KAIST Impurity Modeling) code and conducted learning through two kinds of training data. Training data were generated by giving variation to the transport coefficients obtained from the actual experiments, and two types of training data were generated by dividing the degree of variation into within 2% and 10% of the actual transport coefficients. The algorithm was validated based on impurity transport coefficients from the KSTAR argon injection experiment using UTC-SANCO. We confirm that the artificial neural network, which learns training data made within 10% variation of the actual transport coefficient, has a higher Pearson correlation coefficient of around 0.99. Based on the trained artificial neural network, we apply it to a KSTAR krypton injection experiment to calculate impurity transport coefficients. The calculated peaking factor has a peak value at r/a~ 0.85. Comparing the peaking factor of the argon injection experiment with the krypton injection experiment, we can know that it is consistent with the result that the degree of accumulation of impurities with high atomic numbers is much higher.