Neural network methodology for /sup 1/H NMR spectroscopy classification

A neural network was developed for the purpose of automating the identification of skeletal structures of chemical compounds using /sup 1/H Nuclear Magnetic Resonance (NMR) spectroscopy signals. The neural net developed was a three-layer, feed forward network using 21 hidden layer neurons. Backpropagation of error was used to train the network with a database of 93 chemical compounds. The inputs to the neural network were relative peak integral and chemical shift (PPM) for the 31 largest peaks in each spectrum. Testing was performed using the same database. The trained network was able to identify the presence or lack of presence of several structural features correctly in 97% of the database. The results show great potential for further study of the application of neural networks to NMR spectroscopy classification.