The sub-divertor neutral gas pressure in relation to seven plasma parameters is predicted, based on the data of the Wendelstein 7-X OP1.2b experimental campaign, via two machine learning (ML) approaches, the extra trees (ET) and symbolic regression (SR) for the standard, high iota and high mirror magnetic configurations. To capture the changing behavior of moving from attached to detached conditions, the dataset in each configuration is divided into three clusters. Via a formal feature importance analysis, in all cases considered, the line integrated electron density is ranked as the most dominant parameter, while the heating power, followed by the radiated power, are always included in the key parameters. In the standard configuration, the coil control current has also been ranked very high. On the contrary, the center and edge temperatures and the toroidal current are of much less importance. The ET model yields nearly perfect predictions but it lacks interpretability and may struggle to generalize to unseen data. The two- and three-parameter expressions obtained by the SR model explicitly demonstrate the dependency of the neutral gas pressure on the plasma parameters and have greater potential for extrapolation to unseen estimates. Although SR predictions lag behind those of ET, they still remain very accurate. The present investigation provides valuable insight into the interconnection between the sub-divertor pressure and plasma parameters and may support the design and optimization of the particle and impurity exhaust system of W7-X and other fusion reactors, broadening the scope of ML applications in fusion.
Team et al. (Thu,) studied this question.