Proof of concept of a fast surrogate model of the VMEC code via neural networks in Wendelstein 7-X scenarios

TL;DR

This paper develops a neural network surrogate model for the VMEC code to rapidly compute magnetic equilibria in Wendelstein 7-X stellarator scenarios, significantly reducing computation time for large-scale applications.

Contribution

The paper introduces a convolutional neural network-based surrogate for VMEC, enabling fast equilibrium calculations with high accuracy, covering extensive operational space of W7-X.

Findings

Normalized root-mean-squared error from 1% to 20%

Inference time per equilibrium is milliseconds

3D CNNs outperform fully-connected networks

Abstract

In magnetic confinement fusion research, the achievement of high plasma pressure is key to reaching the goal of net energy production. The magnetohydrodynamic (MHD) model is used to self-consistently calculate the effects the plasma pressure induces on the magnetic field used to confine the plasma. Such MHD calculations serve as input for the assessment of a number of important physics questions. The VMEC code is the most widely used to evaluate 3D ideal-MHD equilibria, as prominently present in stellarators. However, considering the computational cost, it is rarely used in large-scale or online applications. Access to fast MHD equilbria is a challenging problem in fusion research, one which machine learning could effectively address. In this paper, we present artificial neural network (NN) models able to quickly compute the equilibrium magnetic field of W7-X. Magnetic configurations that extensively cover the device operational space, and plasma profiles with volume averaged normalized plasma pressure $\langle \beta \rangle$ ($\beta$ = $\frac{2 \mu_0 p}{B^2}$) up to 5% and non-zero net toroidal current are included in the data set. By using convolutional layers, the spectral representation of the magnetic flux surfaces can be efficiently computed with a single network. To discover better models, a Bayesian hyper-parameter search is carried out, and 3D convolutional neural networks are found to outperform feed-forward fully-connected neural networks. The achieved normalized root-mean-squared error ranges from 1% to 20% across the different scenarios. The model inference time for a single equilibrium is on the order of milliseconds. Finally, this work shows the feasibility of a fast NN drop-in surrogate model for VMEC, and it opens up new operational scenarios where target applications could make use of magnetic equilibria at unprecedented scales.