Paper AS-WeM2
Data Scaling for Quantitative Imaging XPS

Wednesday, October 22, 2008, 8:20 am, Room 207

Multivariate analytical techniques are seeing increasing use in surface analysis due to the ability of current instrumentation to acquire multispectral data sets. Their use can provide a significant improvement in signal/noise, and simplify analysis of the large amount of data present by reducing its dimensionality. In the case of XPS this enables quantification by measurement of photoelectron peak areas, and chemical state determination using curve fitting to the spectrum at every pixel in an image. Selection of the most appropriate technique is dependent upon the characteristics of the data, and for XPS, where there are relatively few components compared with the number of objects in the data set, Non-linear Iterative Partial Least Squares (NIPALS) affords a significant saving in computational requirements, as the procedure may be terminated after the appropriate number of components has been calculated. A key aspect in the use of these techniques is their ability to order the data so that the chemical information is easily separated from the noise. For data acquired by pulse counting, which is governed by Poisssonian statistics and where the variance scales as the data, ordering the data by variance may result in noise from high intensity photoelectron peaks dominating low intensity chemical information. Effective separation of noise and chemical information therefore necessitates prescaling the data, so that the noise is evenly distributed. Individual objects, whether images or spectra, may be scaled by the square root of the variance. This provides good separation of the chemical information from the noise, but leads to a signal level below which information cannot be extracted. This is not important where the inelastic background is greater than this base level, but leads to significant errors in quantification where the background is equal to, or below this level, and data is lost from photoelectron peaks. This is particularly important for photoelectron peaks occurring at low binding energy. It will be shown that scaling to the square root of the mean variance in both image and spectral domains, known as optimal scaling, avoids this limitation and allows use of the NIPALS procedure for fully quantitative imaging XPS.