| For certain problems, conventional chemical analysis does not provide adequate solutions. In some cases, these problems may be solved by applying chemometrical techniques. In general, these techniques model an empirical relationship between a number of measured variables and one or more requested variables. Recently, artificial neural networks have attracted much attention as a means of deriving empirical models from complex data. The object of this thesis is to examine the applicability of these networks as a chemometrical tool and to examine the problems connected with their application.; Chapter 2 introduces artificial neural networks and presents an overview of relevant theoretical aspects. Main emphasis is given to the backpropagation algorithm and the validation of network models. In chapter 3, practical aspects of applying backpropagation networks are discussed and illustrated with the aid of a synthetic data set. The neural network simulation package and other software tools which were developed during this project are discussed in chapter 4. In chapters 5-8, networks are applied to practical problems in chemical analysis. These are respectively the prediction of the water content of cheese from the process variables, accurate multivariate calibration in x-ray fluorescence spectroscopy, the prediction of the CO{dollar}sb2{dollar} permeability of polymer membranes from their IR-spectra, and classifying functional groups from IR-spectra. In general, accurate quantitative results could be obtained which were superior to those of conventional linear-based techniques. However, with relatively large amounts of noise present in the data, different neuron types and modifications to the learning rule were necessary to avoid excessive over-fitting. Qualitative analysis was found to improve a great deal by applying an extension to the global error function used in backpropagation learning.; In chapter 9, the results of this project and directions for further research are discussed. |
