INTERPRETATION OF INFRARED-SPECTRA BY ARTIFICIAL NEURAL NETWORKS

A Pascal program for the simulation of artificial neural networks, was implemented on a PC. For interpretation of infrared (IR) spectra by means of a neural network, a back propagation model with one hidden layer and a sigmoid transfer function has been proved to be the best of several network types. The full curves of less resolved IR spectra (128 measured data points) were used as input elements. 32 structural and substructural groups have been defined as output elements of the network. The hidden layer contains 40 nodes. One hundred spectra from our IR library were selected as the training set and 50 other spectra from the same library were used as the prediction set. After some training sessions (total time 16 h) the output error was sufficiently low (0.004%). In the next step it was tried to predict the structural elements of the 50 compounds of the prediction set. An iterative algorithm for the computation of the specific spectra of the predefined structural elements from the network knowledge after the training has been developed.