Simulation of infrared spectra using artificial neural networks based on semiempirical and empirical data

Abstract

Two different methods using artificial neural networks of the backpropagation type have been applied to simulate infrared spectra of organic CHO compounds. Their performance is compared. The first method uses harmonic frequencies and intensities of a semiempirical (AM1) calculation and the second one, a list of substructures of the corresponding compound as input for the neural net. Both the networks are trained to derive the experimental spectra from the corresponding quantum chemical or structural input information. A set of 840 CHO compounds from the Specinfo database was used for training. Extensive studies to evaluate the quality of the simulated spectra were made. Both the methods are comparable in their performance. The quality of simulation is reasonable within the range of 1300–400cm −, however, the methods fail to predict the finger print region (1300–400 cm −1).