NEURAL-NETWORK ASSISTED RAPID SCREENING OF LARGE INFRARED SPECTRAL DATABASES

A new method of prefiltering applicable to spectral database searches has been developed. It employs a backpropagation neural network to classify 609 matrix isolation FT-IR spectra with respect to the presence or absence of 35 functional groups, sewing as sortable bit string keys to the spectral library. These bit strings or integers are used to construct a binary search tree for 100% successful fast spectral retrieval. Compared with a sequential library search, this method yields a more than 25-fold increase in search speed and could easily be adapted to handle other types of spectral information. For larger databases, this advantage is expected to increase by a factor of n(0.5) relative to the number of spectra. Additionally, the backpropagation neural network training has been modified with an extended version of the flashcard algorithm for 100% successful training of 2651 matrix isolation FT-IR spectra.