J. Walton and N Fairley SIA, 2004; 36: 89-91
J. Walton and N Fairley J Electron Spectrosc. Related Phenomena, 2005; 148: 29-40
When discussing XPS imaging, the question to answer is why XPS image data does not appear more in the literature? One reason typically quoted is the lack of spatial resolution compared to other techniques such as Auger or ToF SIMS; however this in itself is not sufficient to explain the lack of published data. The problem with XPS imaging relates to the nature of XPS spectra, where the signal often sits on the back of broad energy loss features, thus a single image represents the combined influence of the transition monitored and all peaks to higher kinetic energy contributing to the background shape. Traditionally, XPS imaging of a surface consisted of choosing a binding energy representing a transition from the element of interest, acquiring an image then taking a second image at an energy representative of the background signal for the photoelectron peak previously measured. Once the peak and background images are available, interpretation of the results is based on the difference in counts per second between the peak and the background images. The rationale behind the calculation is to extract the signal representative of the peak above the background; however in spectroscopic terms, these calculations represent a very primitive mechanism for measuring a photoelectric line. Few analysts today would quantify XPS spectra using quantities analogous to those used for XPS imaging. Further, the images typically include artefacts attributable to the X-ray source, the lens system, the detection system and sample topography. Thus, visual inspection of individual background subtracted elemental images is often misleading and so explains why images are used to guide spectroscopy suitable for publication, but tends not to be seen as publishable in its own right.
All the same issues of X-ray flux, lens systems, detection systems and sample topography also exist for spectroscopy. The issues associated with XPS imaging then leads to a further question: what is it about spectroscopy that makes the results acceptable? The solution adopted for spectroscopy is to generate a normalised quantification report relative to the total composition of the surface. That is, the measured intensities are corrected for relative sensitivity and instrumental transmission, summed to provide the normalisation factor and reported as a set of intensities relative to this normalisation factor. The resulting percentage atomic concentration report provides a set of numbers by which different samples can be compared. The introduction of parallel acquisition, either in terms of energy or spatial dimensions, means that so called spectromicroscopy where images are acquired in such a way that spectra at each pixel are available for analysis, provides the image data in such a fashion that the tried and accepted spectral analysis techniques can be used to create chemical state image. What is more, these images generated from spectra at pixels can be interpreted with all the certainty of normal spectroscopy.
The key to the success of spectromicroscopy are data consisting of acceptable signal to noise. Large area analysis achieves good signal-to-noise by collecting a single spectrum from a relatively large area, whereas a set of images may be measured from the same area and rather than summing the signal, counts are assigned to spatially resolved pixels at a given energy. If a spectrum of images is acquired, it is not surprising to observe that the spectra at pixels have poor signal-to-noise compared to a single large area spectrum. Fortunately, multivariate statistics offers a means of analysing an image dataset into useful information and noise.
The following discussion is a practical example of spectromicroscopy, where chemical state images are prepared from spectra at pixel data sets. A methodology for creating peak models and applying these models to individual pixel spectra is developed.
A sample consisting of silicon dioxide islands on elemental silicon provides a good example of how energy resolved images can be treated using a combination of Principal Component Analysis (PCA), false colour image partitioning, and standard spectroscopic techniques to reveal an enhanced understanding for the nature of the surface.
The image in Figure 1 represents an example from a set of forty nine images acquired using a 0.25 eV step-size from 536 to 524 binding energy at pass energy 80 through a mirror hemi-spherical analyser onto a delay line image detector (Kratos Axis Ultra at University of Manchester). The surface was irradiated using an Al monochromatic X-ray source. The XPS image illustrates some of the issues with the technique, namely, the intensity varies from bottom left to top right and striations associated with the detection system are clearly visible.
The set of images from which the data in Figure 1 was taken also illustrates the difficulty of selecting a binding energy to represent a specific chemical state. On summing the pixels in each image and presenting these intensities as a spectrum (Figure 2), it becomes clear that the various oxide states are present but form a single peak structure deriving from overlapping component peaks. The O 1s peak in Figure 2 represents a target for peak modelling, since the sum of all the pixels necessarily include contributions from all oxygen states regardless of their spatial separation.
An advantage of spectromicroscopy is the possibility of isolating information in the spatial domain, and then generating spectra based on the spatial filter analogously to the spectrum in Figure 2. If chemically distinct spatial regions exist on the sample, the possibility of identifying and generating spectra representative of the components comprising the spectrum in Figure 2 is achievable. The current example demonstrates how these spectra from sets of similar pixel intensities can be used to create the peak model for identifying the chemical states on the surface. Indeed, the requirement that both the peak model makes chemical sense and the resulting images are similarly plausible enhances the comfort-factor when interpreting the data set.
The first task is to prepare an image from which spatial zones can be identified (Figure 3). The false colour image group pixels together based on intensity and therefore if an appropriate image is used to define the false colours, the raw spectra at each pixel can be summed to produce spectra representative of sample chemistry. Figure 4 are the set of spectra generated by summing the coloured zones from the false colour image in Figure 3. A single peak model must be prepared and applied to each of the spectra corresponding to the false colour image; then together the false colour image and these sum-of-colour spectra are used to guide the fitting of the peak model against all the spectra at pixels in the entire O 1s data set. Before fitting the peak model, the PCA based procedure is first applied to the spectra at pixels to produce reduced-noise data.
Figure 3: False colour image corresponding to spectra in Figure 4.
The spectra in Figure 4 are colour coded to match the colours in the false colour image in Figure 3. Each spectrum in Figure 4 includes a four component peak model; the number of components defined for each spectrum must be the same, however, the constraints are at liberty to be adjusted on a spectrum-by-spectrum basis. Similarly, the number of regions defined on the spectra must be the same for each spectrum; however the regions parameters, eg background type, can vary between spectra.
Figure 4: Spectra generated from the false colour image in Figure 3.
The peak models in Figure 4 contain four O 1s synthetic components. When these peak models are applied to the O 1s data set, where each spectrum-at-a-pixel is associated with one of these peak models using the false colour image, the four images in Figure 5 result. These raw O 1s component images in turn are combined with similarly created images from the Si 2p data set to produce the quantified images in Figure 6. The consequence of applying the quantification transformation is a set of images in which the artifacts due to the instrumentation are greatly reduced and the chemical state information becomes more apparent.
Figure 5: Raw O 1s component images.
The beauty of spectromicroscopy is that the results must be self consistent. That is, a peak model must satisfy the chemical requirements for the surface and, just as importantly, the images must make visual sense of the chemistry too. In this particular example, four peaks where used to analyze the O 1s peaks, where as only two oxide peaks where extracted from the Si 2p data. The O 1s peaks in Figure 4 support the need for four synthetic components and so too do the images in Figure 6. A poorly define peak model could easily result in well fitted spectra but images at odds with the spatial variation of the surface. In the current example, the images created from the Si 2p data are similar to the images synthesized from the O 1s data. The remaining question left unanswered is how these four components within the O 1s image should be interpreted.
Figure 6: O 1s and Si 2p images following quantification transformation.
The sequence of steps when performing a chemical state analysis of an image data set is as follows.
Some of these steps require further explanation.
Figure 7: Image Processing property page
An appropriate image for creating false colour zones is an image for which the intensities are uniform across the image and the intensity variation is somewhat representative of the chemical states of interest. These requirements may seem to be putting the cart before the horse, however what is required is essentially a first order approximation to the desired chemical state information, not the end result itself. For example, the image used to define the false colour scale in Figure 3 was constructed using two images from the original set of O 1s, one representative of the air-formed oxide and one from the SiO2 islands. The two images were overlaid in the active tile and the Quantify button pressed. The resulting images are I1 / (I1 + I2) and I2 / (I1 + I2) which effectively normalizes the images. Another possible image source might be to use the abstract factors from a PCA.
Once each pixel in an image is assigned a false colour, the template image for summing of pixel spectra into individual spectra is defined using the button labeled Define Image. A buffer is loaded with the false colour image information, which can be used by selecting an Experiment Frame containing the raw spectra-at-pixels data, then pressing the Sum Spectra Using Colours button. The spectra-at-pixels must be created from an image data set for which there is an identical number of rows and columns of pixels to the template image. This statement means there must be no additional VAMAS blocks from those originally created when the images where transformed into spectra at pixels.
Having prepared the spectra at pixels Experiment Frame, defined the false colour image template and established a peak model for each of the false colour spectra, the next step is to press the Convert Components to Images button. A dialog window (Figure 8) reports the number of spectral blocks and offers a tick box. The tick box allows the decision between using the false colour spectra to provide the source for the peak-model or a peak model taken from the target spectra-at-pixels. If the tick-box is ticked, then the false colour image guides the peak fit by applying the peak-model from the false colour spectrum matching the given pixel colour to each spectrum at a pixel. Once a spectrum-at-pixel is fitting with a model, the component intensities are extracted and entered into new images.