A neural network model relating extraction current characteristics with optical emission spectra for the purpose of a digital twin of miniaturized ion thrusters

Abstract

Miniaturized ion thrusters are one of the most important candidates in the task of drag-free control for space-based gravitational wave detection, whose thrust can be accurately tuned in principle by in-orbit monitoring and feedback controlling. This work investigates a neural network model (NNM) which can be used for real-time monitoring of the function relating the grid voltage and the extraction current of a miniaturized ion thruster by optical emission spectroscopy. This model is developed as a component of an ion thruster’s digital twin. A collisional-radiative model relates the plasma parameters in the discharge chamber of the thruster to the emission spectroscopy; an extraction current model relates the plasma parameters to the function relating grid voltage and extraction current. The neural network model is trained based on the dataset produced by these models, and is examined by experimental results from a miniaturized ion thruster. It is found that the difference between the thrust predicted by the NNM and the experimental value is less than 6%. Discussions are given on further improvement of the NNM for accurate thrust control in space-based gravitational wave detection in the future.

Machine learning for parameters diagnosis of spark discharge by electro-acoustic signal

“…Gidon [26] introduced a novel deep learning-assisted, non-invasive method using microwave-plasma interactions to accurately estimate electron density profiles in low-temperature plasmas, overcoming limitations of conventional methods and demonstrating promising results in plasma diagnostics through comprehensive simulations and evaluations. Chang [27] demonstrated the effective use of a convolutional neural network, specifically the InceptionTime model, for real-time classification and analysis of atmospheric-pressure plasma jet currents, achieving high accuracy in identifying working gases and discharge types, and showcasing the potential for rapid monitoring and diagnosis in plasma applications. Zhang [28] presented a neural network model for real-time monitoring and feedback control of a miniaturized ion thruster's performance, using optical emission spectroscopy to accurately relate grid voltage and extraction current, achieving less than 6% deviation from experimental values and promising improvements for precise thrust control in space-based gravitational wave detection.…”

Using transfer-learning-based algorithms as data reduction strategies for volatile organic compounds classification using plasma spectroscopy

“…Machine learning (ML) has become increasingly popular in recent years across a variety of fields [8], including the analysis of images [9,10], videos [11], and spectra [12][13][14][15][16][17][18]. The features in the data can be successfully extracted using ML algorithms [19], and the extracted features can be used for regression [6,[20][21][22][23][24], classification [25,26], and error detection.…”

Efficient data acquisition and training of collisional-radiative model artificial neural network surrogates through adaptive parameter space sampling

“…Notably, this series of studies demonstrated success in applying relatively simple ANN techniques with limited computational resources available, when compared to the present day, to streamline the use of fundamental atomic and molecular physics models into a larger plasma physics workflow. More recently, ANN(s) have been shown to be beneficial in learning surrogates for non-LTE radiation models in order to aide analysis and direct interpretation of experimental spectroscopy measurements [34,35]. Lastly, the recent work of Kluth et al [36] is a notable demonstration of the utility of employing ANN surrogates of CR models within plasma simulation codes to provide accurate physics quantities at a fraction of the computational cost of the forward-pass CR model.…”