2025 年 12 巻 1 号 p. 249-266
The magnetic fields confining plasma in fusion reactors are analyzed using Poincaré plots, which show intersections of magnetic field lines on a poloidal crosssection. Traditional methods for designing vacuum vessels and related structures involve slicing the reactor vertically and analyzing these plots, but this approach is inefficient for understanding plasma shapes globally. Key challenges include the computational cost of magnetic field line tracing and the difficulty in constructing continuous surfaces for the divertor legs due to the limited number of magnetic field lines reaching this area. To address these issues, a new and automatic method has been proposed to increase the number of magnetic field lines constituting the divertor legs by predicting optimal starting points for tracing. The proposed method involves placing starting points on orthogonal lines through the magnetic axis, resulting in a better representation of divertor legs. This new algorithm enhances the efficiency of generating Poincaré plots that depict divertor leg regions more clearly than previous methods. Neural Networks predict voxel data representing the shape of magnetic field lines. Considering the Larmor radius, we calculate an envelope surface that encompasses the region where plasma exists and create 3D modeling data.
