Neural Network Data Analysis in the Large Helical Device Thomson Scattering System

In Thomson scattering diagnostics systems, a combination of the lookup table and the minimum χ2 methods has been widely used to determine electron temperature. The concept of the minimum χ2 method is based on clearly defined mathematical statistics. However, the minimum χ2 method calculation requires a large amount of time because all χ2 values have to be calculated at all temperatures included in the lookup table. Thus, this method is unsuitable for the real-time data analysis required for the next generation of fusion devices, e.g., the International Thermonuclear Experimental Reactor in France. To establish real-time data analysis for Thomson scattering diagnostics, we have developed a neural network program for the large helical device (LHD) Thomson scattering (TS) system. First, we systematically studied the number of nodes and training cycles required to obtain satisfactory results, and then applied them to the LHD TS system. The calculation time was successfully reduced by approximately 1/50 - 1/100 of the χ2 method calculation time. In addition, experimental error estimation has been performed according to the concept of the neural network method used in this study