ZAF, PAP, Φ(ρz), α-FACTOR..: A Comparison of the Relative Accuracies of The Alphabet Soup of Correction Factors for “Non-Optimized” Samples
The ultimate practical test of the utility of a correction procedure for quantitative x-ray microanalysis is how well the analyses of standards of interest conform to their known compositions. Many papers have been published testing various correction algorithms by processing data from sets of quantitative analyses through the corrections and then comparing the resulting error histograms {e.g., Fig. 1). The best correction procedure is usually considered to be the method that results in error histograms with the minimum mean relative error and the smallest standard deviation for the distribution of relative errors. Such evaluations can be misleading for a number of reasons: (1) the correction procedure may have been “adjusted” by adding empirical factors to produce superior results for a particular type of specimen. If the test data either include these data or include samples of similar composition to those employed for the refinement, the results may appear artificially good and may not work nearly as well for other types of specimens. (2) The analytical data used for testing may itself be flawed, either because the samples were not actually the compositions they were thought to be (or the sample surface analyzed was not the same composition as that published for the bulk material); or because the surface or the sample was contaminated, rough or charging; or because the analytical conditions were not well controlled. (Many of the published k-factors used in evaluating correction procedures [e.g., ref. 1] were obtained in the early days of microbeam analysis, using instruments having poor control over high voltage and beam current stability with low spectrometer take-off angles.) (3) The analytical data may contain specimens analyzed under unusual conditions {e.g,, very high or veiy low accelerating potentials) that may have very large corrections dominating the data set that may never be encountered in normal analysis.