Predicting the antenna properties of helicon plasma thrusters using machine learning techniques

Abstract

When designing helicon plasma thrusters, one important characteristic is the impedance of the radio-frequency antenna that is used to deposit power into the plasma. This impedance can be characterized both experimentally and numerically. Recently, a numerical tool capable of predicting the antenna impedance, called Adamant, has been developed. However, Adamant takes a long time to run and has high computer resource demands. Therefore, this work has been done to evaluate whether machine learning models, trained on Adamant-generated data, can be used instead of Adamant for small design change evaluations and similar works. Six different machine learning models were implemented in MATLAB: decision trees, ensembles, support vector machines, Gaussian process regressions, generalized additive models and artificial neural networks. These were trained and evaluated using nested k-fold cross-validation with the hyperparameters selected using Bayesian optimization. The performance target was to have less than 5% error on a point-to-point basis. The artificial neural network performed the best when taking into account both maximum error magnitudes and generalization ability, with a maximum error of 3.98% on the test set and with considerably better performance than the other models when tested on some practical examples. Future work should look into different solver algorithms for the artificial neural network to see if the results could be improved even further. To expand the model’s usefulness it might also be worth looking into implementing different antenna types that are of interest for helicon plasma thrusters.