Data-Driven Approach to Model the Influence of Magnetic Geometry in the Confinement of Fusion Devices

TL;DR

This paper uses data-driven methods to analyze how magnetic geometry affects particle confinement in stellarators, aiming to improve fusion device design by predicting omnigenity and related properties.

Contribution

It introduces a comprehensive database and machine learning models to understand the influence of stellarator geometry on omnigenity and optimize fusion device performance.

Findings

Identified key geometric parameters influencing omnigenity.

Developed models to predict solver convergence based on configuration.

Demonstrated the effectiveness of autoencoders and gradient boosting in stellarator analysis.

Abstract

The design of fusion energy devices involves a balance between competing performance metrics to achieve an energy gain. In stellarators, the geometry is very flexible and involves a large number of free parameters. These can be optimized to achieve good performance. One of the main optimization targets is omnigenity, that is, the confinement of alpha particles stemming from the fusion reactions. In this work, two classes of omnigenous stellarators are studied, namely quasisymmetric and quasi-isodynamic stellarators. The goal is to determine the influence of the geometry on omnigenity, which can lead to greater insight into the design space of stellarators. For this purpose, a database of stellarator configurations is created and analyzed for correlations, pair-wise distributions, and dimensionality reduction using a supervised autoencoder framework. Then, a classification model is trained on this database to predict the convergence of numerical solvers. Finally, two regression models, LightGBM and its probabilistic version, LightGBM LSS, as well as a feed-forward neural network, are trained to predict quasisymmetry and quasi-isodynamiticity and find the design parameters that most influence omnigenity.