Abstract
The grid geometry is a critical performance driver in ion-thruster engine design. Research has shown that novel grid geometry can dramatically boost ISP or reduce the erosion rate. [1, 2]. At the same time, computational models of plasma physics are intensive. This paper evaluates the performance of a neural network to replicate a MATLAB-based 2D ion optics simulation and predict engine thrust performance from engine design parameters. The model employs a second-order physics approximation and the particle in cell method. The neural network is trained, via unsupervised deep learning, on a subset of previously developed model simulation solutions. Training inputs to the neural network are the grid geometry, voltages, and thrust achieved for each solution case. In testing, the neural network can predict the thrust performance for engines that it has not previously seen. This study aims to understand better the effectiveness of surface models trained on higher fidelity simulations to increase the efficiency of parametric design studies. Results are the NN predictive performance for a test case and the ion optics, as well as examples of high and low-performing designs. This method appears reasonably suitable for the rapid approximation of simulation spaces. Future work includes: techniques for bounding and reducing error, and physical testing.