How does magnetic geometry affect plasma turbulence? A machine learning approach

Magnetic geometry has a significant effect on the level of turbulent transport, so there is hope of reducing turbulence through the design of the geometry. To do so, it is desirable to better understand which specific geometric features affect the turbulence. Here we take the approach of attempting to gain this understanding based on data. The dataset is generated using a new procedure to sample random stellarator shapes from a plausible distribution. Several thousand nonlinear gyrokinetic simulations are performed in these random geometries. At fixed aspect ratio, gradients, and other parameters, the turbulent heat flux varies between geometries by over an order of magnitude. Patterns are apparent among the configurations with particularly high or particularly low heat flux. Regression techniques from machine learning are then applied to extract patterns in the dataset. Several feature engineering approaches are compared. A noteworthy aspect of these regressions is that the prediction should be invariant to translations of the raw features in the parallel coordinate, similar to computer vision applications.