Comparison of the training of neural networks for quantitative x-ray fluorescence spectrometry by a genetic algorithm and backward error propagation

Neural networks are shown to be useful as emplncal mathematical models m the calculation of quantltatlve analytical results, gvmg sufflclent accuracy to compete successfully with various common cahbratlon procedures. The performance of these neural-network models for cahbratlon data from x-ray fluorescence spectrometry (XRF) was evaluated for two trammg methods, 1 e , backward error propagation (BEP) and a genetic algonthm (GA) For a small triumng set (13 members) of data from Fe/Nl/Cr samples taken from the hterature, the BEP-trained models compared favourably with other hterature methods. The GA-trained models performed poorly for these samples. The two models performed equally well when tramed on a larger data set (30 members) conslstmg of XRF data for thm Fe/N1 layers on a substrate, for whch both the composltlon and the thckness were determmed The predictive power of both models for samples outslde the range of the traming set was unsatisfactory