Neural network correction of PM3-predicted infrared spectra

We describe the application of neural networks to a theoretical problem: the correction of inaccuracies in infrared spectra as predicted by the PM3 semiempirical method. Twenty-eight "peak-correcting" backpropagation neural networks were trained to predict the location of a characteristic infrared peak when given a scaled topological map of one of 1116 literature spectra. The infrared spectra of 200 aliphatics were then calculated using PM3, displayed graphically in Infrared Spectrum Comparison, and submitted to the appropriate "peak-correcting" neural network(s) based on a rule set implemented via an interface to HyperCube's HyperChem software. Results show an average 8-fold decrease in prediction error between PM3-predicted and neural network-corrected peak locations.