English Abstract The usefulness of artificial neural networks (ANN) and
partial least squares regression (PLS) for computerized interpretation of
infrared (IR) spectra has been studied. Experiments have been carried out
to establish the capabilities of these methods to recognize characteristic
band shapes and patterns as used for the interpretation by experts. Spectra
have been classified by (i) the complete spectral profile (ii) the band
pattern in a limited preselected region and (iii) individual band shapes.
The results are compared with classifications using computer generated
frequency/intensity-structure correlations and as performed by experienced
spectroscopists. Classification by skilled interpretators is found to be
superior in all cases but a significant improvement of the results from ANN
and PLS is established compared with predictions obtained from
frequency/intensity-structure correlations. Differences in scores between
ANN and PLS were small when full spectra or limited spectral regions are
considered. Networks scored better in recognizing individual bands. Both
the absorption frequency and the band width play an important role in the
recognition process.
