Machine learning assisted optical emission spectroscopy to determine electron density and electron temperature in a cascaded arc plasma

• A machine learning model was developed to predict n e / T e in cascaded arc plasma. • The n e / T e from LTS and line intensity ratio from OES were used to train the model. • The model achieves high accuracy with or without discharge conditions as input. • The prediction accuracy is robust to the combination of spectral lines selected. High-density cascaded arc plasma has been widely applied in linear plasma devices (LPDs), in which the laser Thomson scattering (LTS) and optical emission spectroscopy (OES) are two popular diagnostic methods for the fundamental parameters, electron density ( n e ) and electron temperature ( T e ). However, the complicated LTS setup lacks spatial flexibility, while the accuracy of simple OES is limited. To address this, this study develops a machine learning model based on Support Vector Machine (SVM) with a grid search optimization. This model combines the high accuracy of LTS with the spatial flexibility of OES to predict n e and T e in cascaded arc plasma in DUT-PSI. The model utilizes four pairs of “double-peak” spectral lines, bypassing the complicated calibration for plasma emission spectrum. The results show that when discharged conditions are included as input (Case 1), the model achieves R 2 values around 0.97 for n e and about 0.92 for T e . When excluding discharge conditions and using only line intensity ratios (LIRs) as input (Case 2), the R 2 values for n e and T e remain approximately 0.90 and 0.80, respectively. The other index, root mean square error (RMSE), follows a similar tendency to R 2 . These findings demonstrate that the predicted n e and T e in both cases are highly consistent with LTS measurements. Meanwhile, sensitivity analysis reveals that the model’s prediction accuracy is robust to the specific combination of spectral lines selected in both cases. Thus, by integrating the strengths of OES and LTS, this model features calibration-free for plasma spectroscopy and flexible spectral line selection, enabling comprehensive diagnosis of cascaded arc plasma and showing potential for application in other similar LPDs.