Neural Pattern Recognition was used for extracting chemical state info
rmation from electron energy-loss (EEL) spectra. The purpose was to ob
tain a quantitative composition profile from sets of low-loss and core
-loss EEL spectra measured along a line across an amorphous inclusion
at a grain boundary in a silicon bicrystal. The spectra were presented
serially to the artificial neural network to obtain the number and sh
ape of the spectra, whose linear combinations reproduce each single sp
ectrum. The results indicate the existence of a different chemical env
ironment at the interfaces between inclusion and crystal. The data ana
lysis proved to be fast, robust, relatively immune to noise or artifac
ts and capable of extracting relevant information from subtle spectral
features.