OPTIMIZATION OF FUNCTIONAL-GROUP PREDICTION FROM INFRARED-SPECTRA USING NEURAL NETWORKS

In a large-scale effort, numerous parameters influencing the neural ne twork interpretation of gas phase infrared spectra have been investiga ted. Predictions of the presence or absence of 26 different substructu ral entities were-optimized by systematically observing the impact on functional group prediction accuracy for the following parameters: tra ining duration, learning rate, momentum, sigmoidal discrimination and bias, spectral data reduction with four different methods, number of h idden nodes, individual instead of multioutput networks, size of the t raining set, noise level, and 12 different spectral preprocessing func tions. The most promising approaches included constant monitoring of t raining progress with a 500 spectra cross-validation set, increasing t he number of spectral examples in the training set from 511 to 2588, e mploying variance scaling, and using specialized instead of multioutpu t networks. An overall recognition accuracy of 93.8% for the presence and 95.7% for the absence of functionalities was achieved, while perfe ct prediction was reached for several present functional groups.