Principal component analysis and neural networks for analysis of complex spectral data from ion backscattering

The problem of ion backscattering spectral data analysis, which is to determine the physical structure of a sample from the measured spectra, was studied with neural network techniques. A new method based on principal component analysis was proposed to compress the number of nodes in the input layer so that the dimensionality of spectral data was significantly reduced. This provides a fast convergence within reasonable size of training set. The constructed neural network was applied to some computation examples, in which backscattering spectra from SiGe thin films on a silicon substrate were discussed in details. The network was trained by the resilient backpropagation algorithm with hundreds of simulated spectra of samples for which the structures were known. The trained network also was tested to analyse spectra with unknown structure of samples. The neural network prediction results were accurate within error of 5.5% and this may suggest that the approach of combining neural network and principal component analysis could be a potential tool of analysis and prediction for non-experts. The proposed approach can handle properly redundancies, which were caused by the constraint of output variables.