Artificial neural network approach to the evaluation of the coordination geometry in organotin(IV) compounds

Artificial neural networks (ANNs) are a simple and rapid system for pattern recognition. In this study they were used to classify Mossbauer spectra of penta-coordinated and octahedral Sn(IV) complexes. Mossbauer spectra recog nition is a lengthy procedure requiring a great deal of experience. The app lication of a system such as artificial neural networks provides a rapid an d accurate method for the correct classification of Mossbauer spectra. As t he two categories of spectra are not linearly separable, conventional techn iques like principal component analysis (PCA) or perceptron can not be used . A more complex ANN was therefore used to solve this problem. The network was built as a standard three-layer back-propagation network with 256 input neurons, 2 hidden neurons and 1 output neuron and a sigmoidal activation f unction. The network's performance was tested with test sets of 10, 20 and 50% of the total number of spectra. The mean square error (MSE) of the diff erent test sets show significant differences. This type of network was able to classify correctly the spectra with an MSE of less than 0.030. Moreover , the network was even able to classify in the appropriate class a spectrum that had been deliberately inverted, demonstrating the ability of ANN to r ecognize objects affected by noise or distortion.