The Effect of Different Machine Learning Surrogate Models for Bayesian Optimization on the Thrust-to-Power Ratio of Physical Electrohydrodynamic Ion Thrusters

Abstract

Humanity continues to be constrained by terrestrial resource requirements, necessitating decarbonized propulsion for space and atmospheric applications. Ion thrusters are known to achieve 10 times the efficiency of traditional rocket engines; moreover, atmospheric electrohydrodynamic (EHD) ion thrusters offer a thrust-to-power ratio 55 times that of turbofan aircraft engines. Despite this potential, EHD thrusters are insufficient for extended atmospheric flight, especially as current methods for traditional simulation-based thruster optimization remain extremely time intensive. This study investigates Bayesian Optimization as an engineering tool to accelerate ion thruster optimization, specifically on surrogate model selection. An EHD thruster was constructed and continuously calibrated to generate an empirical dataset of 7,222 datapoints, then preprocessed through methods of standardization and outlier removal to reduce selection bias. An XGBRegressor (R-Squared = 97%, MAE = 0.04 N/kW) was then trained to serve as the objective function, and the surrogate models Gaussian Process (GP), Random Forest (RF), and Extra Trees (ET) were tested by utilizing a weighted sum that combines optimization time and thrust-to-power ratio into a single score, with both metrics given equal weights. ETR had the most balance between thrust-to-power ratio and optimization time, with 0.27 N/kW and 8.98 seconds respectively. This resulted in a weighted sum of 0.777 (p<0.001, F(2, 297) = 1737), significantly outperforming GP and RF. From this study, ET’s proficiency as a surrogate model reduces the reliance on exhaustive methods of manual parameter sweeping and facilitates testing of broader variables with fewer iterations, significantly accelerating development of high-performing ion thrusters.