Identification of chemical species using artificial intelligence to interpret optical emission spectra

The nonlinear modeling capabilities of artificial neural networks (ANN’s) are renowned in the field of artificial intelligence (Al) for capturing knowledge that can be very difficult to understand otherwise. Their ability to be trained on representative data within a particular problem domain and generalise over a set of data make them efficient predictive models. One problem domain that contains complex data that would benefit from the predictive capabilities of ANN’s is that of optical emission spectra (OES). OES is an important diagnostic for monitoring plasma species within plasma processing. Normally, OES spectral interpretation requires significant prior expertise from a spectroscopist. One way of alleviating this intensive demand in order to quickly interpret OES spectra is to interpret the data using an intelligent pattern recognition technique like ANN’s. This thesis investigates and presents MLP ANN models that can successfully classify chemical species within OES spectral patterns. The primary contribution of the thesis is the creation of deployable ANN species models that can predict OES spectral line sizes directly from six controllable input process parameters; and the implementation of a novel rule extraction procedure to relate the real multi-output values of the spectral line sizes to individual input process parameters. Not only are the trained species models excellent in their predictive capability, but they also provide the foundation for extracting comprehensible rules. A secondary contribution made by this thesis is to present an adapted fuzzy rule extraction system that attaches a quantitative measure of confidence to individual rules. The most significant contribution to the field of Al that is generated from the work presented in the thesis is the fact that the rule extraction procedure utilises predictive ANN species models that employ real continuously valued multi-output data. This is an improvement on rule extraction from trained networks that normally focus on discrete binary output