Application of artificial neural networks in the quantitative analysis of gas chromatograms

This thesis demonstrates the effectiveness of artificial neural networks (ANN) in the quantitative analysis of multicomponent environmental samples. The data used in this research is from simulated soil samples contaminated with environ-mentally hazardous chemicals. Chromatographic data for three different pure analytes (Aroclors 1242, 1254, and 1260), as well as mixtures of the three, were collected by gas chromatography (GC) as training and testing data for the ANN. The networks were designed so that the concentrations of three different Aroclors in a given soil sample may be determined even when multiple Aroclors are present in one sample. This is significant since traditional linear methods tend not to han-dle overlapping contributions from multiple components. Due to the nature of a multilayered feedforward neural network, such overlapping contributions can be isolated and nonlinear classification analysis can be employed. Network param-eters and architectures were optimized to give maximum performance. This is important because the architecture affects the generalization capabilities of the network; that is, its ability to produce accurate results on patterns outside its training set. Our ANN is as effective, or more effective, in both single and mixture concentration determination as existing methods such as multiple linear regres-sion and principal component regression. Our method also has the advantage of using peak area tables as input data instead of the entire chromatogram.