Mn. Souza et al., APPLICATION OF THE ARTIFICIAL NEURAL-NETWORK APPROACH TO THE RECOGNITION OF SPECIFIC PATTERNS IN AUGER ELECTRON-SPECTROSCOPY, Surface and interface analysis, 20(13), 1993, pp. 1047-1050
The artificial neural network (ANN) approach was applied to the identi
fication of Auger electron spectral patterns. The ANN structure employ
ed was the counter-propagation architecture with an unsupervised learn
ing algorithm. For training such a network, it is only necessary to pr
ovide a data set with samples of the patterns to be recognized, and th
e network itself will extract the relevant statistical information to
organize similar patterns into specific classes. We used a training da
ta set of five different Auger spectra (Fe, Au, Si, Sn, Cu) to which a
random fluctuation of up to 5% of the highest peak was added. To the
test set, however, the added fluctuation was up to 50% and we observed
that the network was able to identify precisely any test spectrum aft
er only a few training sessions. The ANN synapses can be interpreted a
s the average spectra of the training set for each specific class, ten
ding to zero fluctuation spectra as the number of training samples bec
omes large. The results obtained show that even by using an extremely
simple ANN structure the classification of single-element Auger spectr
a was made easy also in the case of extremely noisy spectra.