Development and testing of an efficient data acquisition platform for machine learning of optical emission spectroscopy of plasmas in aqueous solution

Abstract

This study presents the development of an efficient data acquisition platform and discusses machine learning of optical emission spectroscopy (OES) of plasma in aqueous solution to demonstrate the multivariate analysis of spectra. A specially designed platform for efficient acquisition of spectra emanated from plasmas in solutions is developed, and several machine learning algorithms are tested for plasma analysis. This platform enables acquiring up to 10k spectra solutions with various pH values under constant solution conductivity, or vice versa, within 15 min. This rapid acquisition scheme provides a sufficient dataset for testing machine learning algorithms. We test the OES of plasmas ignited in solutions with designated conductivities with pH of 2.2–5.2. A total 40k spectra are collected and tested with principal component analysis (PCA) and an artificial neural network (ANN) to predict the conductivity of the solution. In PCA, the results show that most data points are overlapped in the score plot constructed using principal components 1 and 2, implying that PCA cannot discriminate the conductivity based on the spectra. In an ANN, several network structures are constructed and tested. The results show that the deep ANN significantly improves the accuracy of conductivity prediction in terms of mean squared error by three orders of magnitude compared with the method of using single emission line. Regularization techniques including dropout and early stopping show promise for mitigating overfitting in a deep ANN. Such improvements suggest that a deep ANN considers the nonlinear behaviors of plasma and can handle datasets with high complexity. In addition, the application of a deep ANN with a large number of parameters makes using this experimental platform for efficient spectra acquisition highly desirable.