Determination of the electron temperature and density in argon plasmas using a machine learning based collisional-radiative model

Abstract

Diagnostics of the electron temperature and density of the plasma through optical emission spectroscopy and the collision-radiation model has the advantage of performing physical characteristic analysis based on a relatively simple measurement system. The conventional method of diagnosing electron temperature and density is derived by comparing the spectral line intensity from the collisional-radiative model by the two-line ratio pair method. However, this method has the ambiguity in selecting wavelength pairs according to the experimental conditions. In addition, there is a limitation in that the spectral data of only four wavelengths is factored in the prediction rather than all spectral intensities. To solve the problem, this thesis proposes a new method to predict the electron temperature and density by applying machine learning. The synthetic data generated from the collisional-radiative model of the low-temperature argon plasma is used as the training data of machine learning. The total of 11 spectral line intensities emitted from neutron particles derived from the collisional-radiation model was trained as the input and the electron temperature and density as the output of the multilayer perceptron neural network with the prediction accuracy above 90%. In addition, uncertainty was calculated along with the prediction results by applying Bayesian machine learning. By using the deep ensemble neural network and the autoencoder, the aspect of uncertainty was mutually verified. Lastly, the machine learning method is applied to the experiments to validate the developed machine learning. From the argon capacitively-coupled plasma with a discharge power of 5-20 W and the inductively coupled plasma of 150-200 W and the 300 W class Hall thruster plasma, the values predicted by the machine learning and results measured from the Langmuir probe were compared and analyzed. As a result, the predicted value was accurately calculated compared to the previous method with an error of 10-30% based on the experimental value, and its effectiveness was verified in various plasma. Therefore, the method is expected to be widely applicable for basic and applied plasma research by measuring electron temperature and density.