Prediction of the hydrocarbon gas composition from the optical emission spectra of argon-based plasma using machine learning with Shannon entropy

Diagnostic methods for gas composition are essential in plasma processes, but conventional techniques such as gas chromatography are time-consuming and not suitable for real-time diagnostics. In this study, a machine learning approach was applied to predict the composition of hydrocarbon gases—methane, ethane, and propane—based on optical emission spectroscopy (OES) of argon-based plasma. Despite the similar spectral features of these hydrocarbons, particularly between ethane and propane, the constructed artificial neural network model achieved accurate composition prediction even under mixed-gas conditions.

To enhance the model performance, Shannon entropy—calculated from the spectral intensity distribution—was introduced as an additional input feature. The Shannon entropy reflects the dispersion of the data in a spectrum and serves as a feature that reduces overfitting to local peak intensities and emphasizes the overall shape and characteristics of the spectrum. This can help rebalance the spectral data in the model. By comparing models with and without the entropy input, whether the Shannon entropy can improve compositional prediction and/or generalize the model was evaluated. The results show that the entropy model made more accurate predictions for unseen OES data, indicating Shannon entropy helps improve the reliability and robustness of the model [1].