A data-driven approach to relate coil complexity to geometric properties of stellarator configurations

QI magnetic configurations are particularly relevant in the modern landscape of stellarator designs and still relatively scarcely investigated under a data-driven lens. An important question in progressing design is, how does the complexity of the coils depend on the geometric properties of stellarator magnetic configurations?

In this study, we try to learn an answer from data: we consider an existing, open source dataset of quasi-isodynamic (QI) stellarator configurations.

As a first step, we generate a dataset of filamentary coils which produce the magnetic field configurations in the dataset with high accuracy, while at the same time satisfying critical engineering constraints such as coil-to-coil distance, coil-to-plasma surface distance and coil curvature, as well as penalizing the length of the curves.

To investigate the dependency between the geometrical properties of the plasma boundary and the engineering features of the coils we employ statistical methods: we look at correlation metrics, we quantify the relative importance of variables in determing coil’s features with feature importance studies, and we use neural network models, including autoencoders and generative models, to predict coil features, cluster their shape according to their complexity and sample novel coil sets with desirable properties.

Our study aims at adding valuable insight when it comes to the understanding, design, and optimization of the magnetic field of such devices under the light of a feasible coil set.