Identifying functional groups in IR spectra using an artificial neural network

Abstract

Artificial neural networks are capable of learning and are potentially superior to other computer programs at pattern recognition. We have used a simple two-layer, feed-forward neural network to obtain structural information from IR spectra of organic compounds. The network was taught to recognize the presence and absence of selected functional groups and bond types by simply presenting it with IR spectra of training compounds. The back-propagation algorithm was used to adjust the weights of the network. Spectra of compounds not belonging to the training set were used for testing. The trained network was able to recognize the presence and absence of the functional groups and bond types in the spectra of previously unseen compounds. Percent transmittance vs. wavenumber was the most successful input data representation. Using both bond type and functional group identification in the output layer significantly reduced the number of incorrect classifications.