Machine learning surrogate model for ideal peeling-ballooning pedestal MHD stability

Magnetohydrodynamic (MHD) stability simulations are central to predicting the performance of pedestals in high-confinement mode plasmas. A machine learning surrogate model, called KARHU, for the ideal MHD stability code MISHKA has been developed using a feed-forward convolutional neural network trained on a database of equilibrium simulations spanning a subset of the JET-ILW parameter space. A dataset of about 16 000 equilibria was created and MISHKA was used to assess the stability of these equilibria for eight toroidal mode numbers ranging between 3 and 50. KARHU was then trained to predict the maximum growth rate out of these toroidal mode numbers. The surrogate model was integrated into the Europed workflow. The Europed predictions using the surrogate model were compared to respective predictions using Europed with MISHKA, in order to demonstrate the improvement in simulation time and the accuracy of the predictions. A Europed run for an example scan was accelerated by 72%, where the MHD stability evaluation part of the model took less than 1% of the runtime. The accuracy was not compromised significantly. While the equilibria in this proof-of-principle work assume the standard Europed ballooning critical profile constraint to reduce the dimensionality of the dataset, the surrogate model was also tested on equilibria outside this constraint. Even for these equilibria that are strictly speaking outside the training domain, the model retains relatively good prediction performance within an average error of 22% for these pressure profiles.