Comparing different pulse processors

For accurate and efficient EDS analysis the performance of the pulse processor is as important as the detector.

EDS detector specifications typically reveal the best possible performance that the detector can achieve. Any pulse processor will perform its best at very low count rates, when the voltage steps on the ramp are widely spaced and easy to measure. Therefore detector specifications are often quoted at 1000cps. However, this count rate is well below what is required for efficient analysis and only the best designs of pulse processor will maintain good and stable detector performance as the input count rate is varied.

How does performance change when the count rate is increased? At more commonly used input rates between 2000 and 10000cps, processors may not be able to maintain the best resolution if resolution degrades with count rate, or process time is shortened to achieve a useful acquisition rate. Therefore one useful measure of how an EDS hardware system will perform is the resolution achieved when the input rate is at least 2500cps, and the acquisition rate is below the maximum for the process time chosen.

Modern EDS software is designed to give reliable automatic identification of X-ray peaks and accurate standardless analysis. This is a relatively straightforward task when peaks are well separated, but for overlapped peaks where there is no clear valley between the peaks, accurate energy calibration is vital. Some systems may appear to have stable performance with count rate because peaks do not move more than 5eV and X-ray line markers are always in the correct channel at 10eV/channel. However, when quantifying peaks about 35eV apart (e.g. SiKα and WMα) only a 4eV shift in energy calibration can introduce a 10 weight% error (Statham 2002). If the resolution also varies and the width of peaks is wrongly predicted by the software then larger errors may occur.

Analog pulse shaping designs have difficulty maintaining a stable baseline so the energy calibration may vary as count rate increases. Time variant shapers or digital pulse shapers should show much less variation, and provide more reliable results, without the need to keep input rate constant. Moreover, processors which are able to constantly monitor the zero level will be able to measure shifts in the baseline and, in addition, some systems can also monitor resolution changes for correction by software.

The best method to test how energy calibrations and resolutions change over a useful operating count rate range, and how well these changes are compensated, is to analyze real samples at a useful range of input rates, for example 1000, 2500, 5000 and 10000cps. By doing this the effect of any variation on the ability of a system to perform reliable analysis can be tested. The ability to resolve a severe overlap can be tested by analyzing a pure material for elements known not to be present. For example, by collecting spectra from a pure element such as Si, at different count rates.

If the spectra are quantified assuming Si, and Ta and W are present (TaM and WM are very close to SiK), a system that will provide accurate analysis in this count rate range should find levels of Ta and W below statistical significance at all count rates. By overlaying spectra collected at the different rates it may also be possible to see peak shifts or resolution changes with count rate. For example Fig. 20 shows four spectra collected from pure silicon at different input rates using a digital pulse processor where count rate stability is good. Quantitative analysis of these spectra, assuming Si, Ta and W are present (Table 1) shows that within statistical significance only Si is present, whatever count rate is used to collect the data.

	1000cps	2500cps	15000cps	10,000cps
Si	99.68 ± 1.78	99.43 ± 1.41	99.90 ± 0.97	99.23 ± 0.90
Ta	-0.82 ± 1.56	0.31 ± 1.22	-0.99 ± 0.85	0.04 ± 0.78
W	1.14 ± 0.89	0.26 ± 0.70	1.09 ± 0.49	0.74 ± 0.46

Table 1 Quantitative analysis of pure silicon calculated from the spectra shown in Fig 20 assuming Si, Ta and W are present. At all input rates Ta and W are below statistical significance (3 sigma). Note that negative numbers should be reported, because negative values greater than statistical significance also indicate incorrect spectrum deconvolution.