Machine learning for emission spectroscopy of Hall thrusters
ORAL · Invited
Abstract
Optical emission spectroscopy is one of the most commonly applied methods for non-invasive characterization and diagnostics of plasmas. In order to extract from the recorded spectra the plasma parameters such as electron density and temperature, collisional-radiative models (CRM) are employed. With the advancement of machine-learning techniques, new ways of analysing the emission spectra are becoming available [1,2]. These methods allow to train models that act as an inverse of a CRM. This provides fast estimation of the plasma parameters from the emission intensities and opens the possibility for the development of fast-response systems that can act as feed-back controllers for improved control of the plasma system [3].
In this work, we present results from the application of machine learning methods to spectra collected from Hall effect thrusters operating in Xe and Kr. Unsupervised learning methods allow characterization of the ability of a CRM to generate the experimental spectra and serve as a way for estimating the accuracy of the CRM. Synthetic spectra are used for the training of supervised learning algorithms. Both analytical distribution functions and plasma parameters from particle in cell simulations are used to generate the training spectra. The supervised learning methods then provide predictions for the plasma parameters in the thruster. In a combination with experimental data, the machine learning methods allow also the correlation of the emission spectra and the thruster control parameters such as voltage and mass flow rate. In the talk, the machine learning methods used will be briefly described and their results will be discussed.
[1] Arellano, F. J., etal, J. Vac. Sci. Technol. A, 42, 053001 (2024).
[2] Ben Slimane, T., "Investigation of the Optical Emission of Hall Effect Thrusters using Collisional Radiative Model, Particle-In-Cell Simulations, and Machine Learning", PhD thesis (2023, IP Paris).
[3] Ben Slimane, T., etal, J. Appl. Phys., 136, 153302 (2024).
In this work, we present results from the application of machine learning methods to spectra collected from Hall effect thrusters operating in Xe and Kr. Unsupervised learning methods allow characterization of the ability of a CRM to generate the experimental spectra and serve as a way for estimating the accuracy of the CRM. Synthetic spectra are used for the training of supervised learning algorithms. Both analytical distribution functions and plasma parameters from particle in cell simulations are used to generate the training spectra. The supervised learning methods then provide predictions for the plasma parameters in the thruster. In a combination with experimental data, the machine learning methods allow also the correlation of the emission spectra and the thruster control parameters such as voltage and mass flow rate. In the talk, the machine learning methods used will be briefly described and their results will be discussed.
[1] Arellano, F. J., etal, J. Vac. Sci. Technol. A, 42, 053001 (2024).
[2] Ben Slimane, T., "Investigation of the Optical Emission of Hall Effect Thrusters using Collisional Radiative Model, Particle-In-Cell Simulations, and Machine Learning", PhD thesis (2023, IP Paris).
[3] Ben Slimane, T., etal, J. Appl. Phys., 136, 153302 (2024).
–
Presenters
-
Tsanko V Tsankov
LPP, CNRS, Ecole Polytechnique, Sorbonne Université, Laboratoire de Physique des Plasmas, CNRS, Ecole Polytechnique, Laboratory for Plasma Physics, CNRS, Ecole Polytechnique
Authors
-
Tsanko V Tsankov
LPP, CNRS, Ecole Polytechnique, Sorbonne Université, Laboratoire de Physique des Plasmas, CNRS, Ecole Polytechnique, Laboratory for Plasma Physics, CNRS, Ecole Polytechnique
-
Tarek Ben Slimane
Instituto de Plasmas e Fusão Nuclear, Instituto Superior Técnico, Universidade de Lisboa
-
Clarence Deltel
Safran Spacecraft Propulsion, Vernon, France
-
Federico Petronio
Laboratoire de Physique des Plasmas