A statistical methodology to derive the scaling law for the H-mode power threshold using a large multi-machine database

In this paper, a refined set of statistical techniques is developed and then applied to the problem of deriving the scaling law for the threshold power to access the H-mode of confinement in tokamaks. This statistical methodology is applied to the 2010 version of the ITPA International Global Threshold Data Base v6b(IGDBTHv6b). To increase the engineering and operative relevance of the results, only macroscopic physical quantities, measured in the vast majority of experiments, have been considered as candidate variables in the models. Different principled methods, such as agglomerative hierarchical variables clustering, without assumption about the functional form of the scaling, and nonlinear regression, are implemented to select the best subset of candidate independent variables and to improve the regression model accuracy. Two independent model selection criteria, based on the classical (Akaike information criterion) and Bayesian formalism (Bayesian information criterion), are then used to identify the most efficient scaling law from candidate models. The results derived from the full multi-machine database confirm the results of previous analysis but emphasize the importance of shaping quantities, elongation and triangularity. On the other hand, the scaling laws for the different machines and at different currents are different from each other at the level of confidence well above 95%, suggesting caution in the use of the global scaling laws for both interpretation and extrapolation purposes.