The performance of artificial neural networks (ANNs) for modeling the Cr---Ni---Fe system in quantitative x-ray fluorescence spectroscopy was compared with the classical Rasberry-Heinrich model and a previously published method applying the linear learning machine in combination with singular value decomposition. Apart from determining if ANNs were capable of modeling the desired non-linear relationships, also the effects of using non-ideal and noisy data were studied. For this goal, more than a hundred steel samples with large variations in composition were measured at their primary and secondary K¿ and Kß lines. The optimal calibration parameters for the Rasberry-Heinrich model were found from this dataset by use of a genetic algorithm. ANNs were found to be robust and to perform generally better than the other two methods in calibrating over large ranges.