The design of systems for spectral data interpretation requires clustering
of chemical compounds based on their spectral characteristics. Kohonen netw
orks have been shown to be efficient tools to achieve this clustering. Thes
e auto-organising systems perform a mapping between a high-dimensional vari
able space and a two-dimensional one. An application to infrared spectra of
organic compounds is presented here. The non-supervised learning algorithm
used allows classification of compounds by spectral characteristics withou
t a priori knowledge. An analysis of the distribution of spectra on the res
ulting maps is used to build models for predicting the presence or absence
of specific structural features. The performance of the models in recognisi
ng structural features is discussed and compared with the prediction of a m
ultilayered feed forward network (MLFFN). Localisation of compounds wrongly
classified by the MLFFN on the Kohonen maps allows to establish a link bet
ween the supervised and the unsupervised approaches.