Development of optical emission spectroscopy method with neural network model: Case study of determining the electron density in a xenon microwave discharge

Optical emission spectroscopy (OES) is an important technique for plasma diagnostics. Random deviation is inevitable during the measurement of plasma emission spectra due to the imperfection of instruments and other interferences. On the other hand, inaccuracies in the collision cross-section data can lead to distortion of the collisional-radiative (CR) model. The coupling of theoretical and experimental error factors can pose difficulties for accurate diagnostics of plasma. This work presents the development of the OES method for xenon plasma that employs a neural network model to integrate prior information on the characteristics of instrument noise and model distortions, thereby improving the accuracy of OES diagnostics. The neural network model takes emission line ratios as input and normalized electron density as output and is trained using a dataset that is generated with a CR model and an instrument disturbance model. The neural network-based OES method is implemented to determine the electron density in a microwave discharge chamber and compared with a traditional OES method with a multi-variant fitting technique. A significant improvement on relative deviation of diagnostic results is observed, which promises a good prospect for further development.