Chapter 3.37. Nonlinearities in solid-state fluorescence detectors

For the purpose of XAS, a fluorescence detector converts fluorescence photons into charge clouds, a preamplifier turns the charge clouds into a voltage signal, and a digital pulse processor turns the voltage signal into a photon energy. Dead time is a nonlinear characteristic of the fluorescence detector, primarily the pulse processor, which limits the number of events that it can process per unit time. While the detector electronics are not the subject of this section, a general description of detector signal processing is provided here since the electronics are the key to understanding the origin of detector dead time.

The following will describe dead time in the context of a reset preamplifier and a trapezoidal filter digital pulse processor, which is the most common combination at many XAS beamlines. This section assumes that the reader is familiar with the trapezoidal filter approach to digital pulse processing; excellent implementation-specific descriptions can be found in Redus et al. (2008), Redus (2009) and XIA (2008). A key element of a digital pulse processor is the analog-to-digital converter (ADC), which digitizes the voltage signal from the preamplifier. The ADC typically samples at a given frequency, returning one measurement per cycle. Registering a photon event requires multiple ADC samples, sufficient to determine an increase in voltage above a threshold, and thus takes a finite amount of time. The more precisely the magnitude of this increase needs to be measured, the longer it takes. Precise determination of the energy of a photon necessitates measuring the voltage increase due to the incident photon in isolation. The arrival of a second, or more, photons during the period that the first photon is being measured will superimpose a second voltage increase on the first, compromising the measurement, and it must be discarded to avoid spurious peaks in the energy spectrum. During this period, where the measurements of the pulse processor have been compromised, the processor will be effectively `dead' for the purpose of detecting photons, thus coining the term dead time.

Trapezoidal filter parameters can typically be modified by the operator to suit specific needs. The main parameter accessible to the user is the `peaking time', which represents the number of ADC samples collected on either side of the preamplifier voltage step. `Faster' filters permit a higher count-rate throughput with lower dead time at the expense of baseline and signal sampling precision, which in turn introduces uncertainty in signal amplitudes and therefore reduces the energy resolution. In the extreme case, small preamplifier signals (low-energy photons) may not be resolved from the baseline noise if an insufficient number of ADC samples are averaged. Another parameter of the trapezoidal filter, known as the `rise time' or `gap time', is characteristic of the detector and preamplifier, encompassing electron mobility, charge-collection time at the anode and the time constants of the pre­amplifiers used. The specifics depend on the type and the technology of the digital signal processor used. Since optimal digital filter settings also depend on the type of solid-state detector, the preamplifier used, the desired energy resolution and the operating count rate, the operator needs to make choices depending on the experimental environment. These choices will hence influence the dead-time behaviour. In this section, we thus only provide a general practical approach for dead-time correction.

The most prominent models for describing dead time are the `paralyzable' and `non-paralyzable' models (Jenkins et al., 1995; Knoll, 2010). In the `non-paralyzable' case, the pulse processor will be busy (`dead') for a period of time τ and will simply be impervious to further events arriving. After the dead-time interval τ has elapsed, the pulse process is then active. In the `paralyzable' (or `extendable') model, the arrival of an event during τ will extend the dead time. The `paralyzable' model typically applies to pulse processors employed with fluorescence detectors and will be used in the discussion in this section. This model leads to a nonlinear relationship between the true rate of incoming events, R_true, and the event rate actually reported by the pulse processor, R_out (see, for example, Reed, 1972; Leo, 1994; Knoll, 2010):

It is useful to analyze this equation further to understand the detector response to photon exposure at different count rates. In the limit of low count rates and/or very fast pulse processors (τR_true → 0), the pulse processor can handle all events received (R_out → R_true). For very high count rates (τR_true → ∞) the pulse processor will be completely saturated and unable to register any events (R_out → 0). In between the extremes, the function R_out shows a maximum at R_true,0 = 1/τ. Up to this point, increasing the fluorescence intensity will result in an increased number of events processed. Any further increase of R_true only leads to a reduction of R_out as the damping of the exponential function takes over. This `loss' of photons increases over-proportionally, i.e. nonlinearly, with R_true.

The concept of dead time applies to any detector event-processing chain. Trapezoidal filter digital pulse processors provide a `fast' and a `slow' channel. The slow channel processes the event and determines the energy of the fluorescence photon received. The fast channel operates on the same principle as the slow channel, but is designed to run as fast as possible, with the purpose of determining R_true with as little dead time as possible. It is used to identify and reject compromised events, i.e. the arrival of another photon during processing of the previous event (XIA, 2008; Redus, 2009, 2017). Rejection of compromised events is commonly referred to as pile-up rejection. In addition, the fast channel also provides a convenient method of correcting for dead time of the slow channel (see below).

The dead time of the fast channel is not negligible (Cramer et al., 1988; Nomura, 1998). For example, using equation (1), a fast channel with τ = 0.25 µs dead time will still exhibit count-rate loss of 0.25% even at a low incoming true count rate of R_true = 10 000 counts s⁻¹ (see Table 1). Therefore, correcting the dead time of both the fast and slow channels is important to optimize the linearity of signals for XAS.

1.2. Concluding notes on dead time

The paralyzable dead-time model tends to be a good descriptor for detector dead time even at very high count rates. For low to moderate rates, it is possible to approximate the paralyzable model (equation 1) as R_out ≃ R_true(1 − τR_true), which is indeed a Taylor approximation to the linear term. This gives an error of <0.1% for small values of τR_true < 0.044. For example, for fast and slow detector channels with τ_fast = 0.25 µs and τ_slow = 1.5 µs, the approximation limits the count rate to R_true,fast < 176 000 counts s⁻¹ and R_true,slow < 30 000 counts s⁻¹, respectively (Fig. 1). In this case, τ_slow would be the rate-limiting factor as the paralyzable dead-time model would predict a dead-time-related decrease from R_true,fast = 176 000 counts s⁻¹ to only 135 000 counts s⁻¹.

Figure 1 Example of dead-time curves for a fast (τfast = 0.25 µs) and slow (τslow = 1.5 µs) channel. The solid line is the expected true count rate. Note that for high rates the fast channel also responds nonlinearly, thus requiring correction.

Synchrotron sources are pulsed, and the storage-ring bunch-fill pattern can influence the magnitude of detector dead time. Using silicon drift detectors (SDDs), Walko et al. (2011) demonstrate how, in a low-bunch mode, the resulting noncontinuous illumination of the detector leads to an approximate twofold increase in the dead time. The results emphasize the importance of correcting nonlinearities in both the fast and slow channels. The count-rate response still follows the paralyzable dead-time model, and the practical approaches to dealing with dead time still apply.

As discussed in the preceding section, pulse-processing electronics require a period of time to process an incoming event. When two photons arrive at the detector within the resolving time of the digital pulse processor, the two events appear in the fluorescence spectrum as a single event with an energy that is the sum of the component photons, referred to as a `sum peak', `coincidence peak' or `pile-up peak' (Statham, 1977, 2006; Mott, 2010). Fig. 2 illustrates how pile-up manifests in a fluorescence spectrum. The data were acquired using a <sup>55</sup>Fe radioactive source and an SDD. The spectrum shows the characteristic Mn Kα and Mn Kβ lines and their respective first-order and second-order summed peaks. Note that any event recorded in the spectrum has a probability to pile up with any other event. Hence, the pile-up signals observed are not merely a `copy' of the Mn Kα/β duplet but their convolution. For the second-order sum peaks, the combinations are manifold but are not resolved in this spectrum.

Figure 2 A Mn Kα/β fluorescence spectrum (from a 55Fe source) showing various pile-up peaks at double and triple the primary photon energies of the Mn Kα/β duplet. The data were acquired using a single-element Vortex SDD at a count rate of approximately 70 000 counts s−1 (total sampling time 3128 s).

For the purpose of further discussion in this section, we assume that the fast channels of digital pulse processors at XAS beamlines provide fairly short `pulse pair resolving times' (t_p) for pile-up to occur of typically around a few 0.1 µs or below (Mott, 2010; Ritchie et al., 2012). The precise time depends on the type of detector and its preamplifier(s)¹. Following Poisson statistics, the probability for any, i.e. one or more, events to arrive within the resolving time is ρ′ = 1 − exp(−R_truet_p), where R_true is the rate of incoming events. For pulse pile-up we need to consider the probability for two or more events to arrive, which is expressed as the above ρ′ minus the probability for only one event to arrive (Scott et al., 1995; Walko et al., 2008),

In Fig. 3, ρ(R_true) is shown for selected values of t_p, illustrating that the pile-up probability is nonlinear. However, the increase of ρ(R_true) is small in the limit of low count rates and short resolving times, and its deviation from a linear function is also small in this limit. As an example, the yellow graph in the figure shows that ρ(t_p = 500 ns) < 0.2% for R_true < 130 000 counts s⁻¹, and a linear fit to this curve gives a small error of <0.05%. At very high count rates, >1 × 10⁶ counts s⁻¹, ρ(R_true) readily increases to a few percent and the linear approximation will remain within 0.1% only in the case of fast electronics (for example t_p = 50 ns).

Figure 3 Probability ρ(Rtrue, tp) for at least one pile-up event to occur during tp. The bottom panel shows the same data over a wider range to the very high count rates relevant to more recent detector electronics developments.

In the linear regime, pulse pile-up turns into a linear scaling factor of the XAS spectrum intensity recorded. Increasing the count rate by scanning through an absorption edge means that we `travel' up on the pile-up curve and scale the spectrum accordingly, and the pile-up effect can then be expected to normalize out satisfactorily. This stands in contrast to dead time, which results in significant nonlinear non-normalizing terms.

A fluorescence detector measures the energy of a photon by measuring the total charge in a charge cloud generated following the absorption of that photon. Hence, any physical loss of charges from the cloud on its way from the absorption site to the anode will result in an event of apparently reduced photon energy. If charge-loss mechanisms were solely proportional to the energy of the photon detected, those losses would merely result in a linear scaling factor of the fluorescence spectrum recorded. In reality, charge-loss mechanisms are more complex, not necessarily linear, and result in a continuum of incomplete charge clouds collected at the anode. Correspondingly, these losses contribute to the background continuum observed in fluorescence spectra.

An important charge-loss avenue arises from the fact that a detector has a finite size. Absorption of a photon near the edges of the active volume of the detector can lead to some charges leaving that volume. This becomes especially important for monolithic multi-element detectors where charge clouds can be shared between two neighbouring detector elements. Multi-element detectors play a significant role in fluorescence XAS due to their high count-rate capabilities, and as such charge sharing is an important aspect of the detector technology supporting XAS.

A quality marker that is often used to describe the effects of charge losses and sharing is the `peak-to-background ratio' (P/B). It is commonly defined as the peak maximum of a fluorescence line divided by the average background in the low-energy part of the spectrum (Iwanczyk et al., 1988; Eggert et al., 2003; Lépy et al., 2003; De Geronimo et al., 2008, 2010); see also Fig. 4).

Figure 4 The peak-to-background ratio P/B. For the spectrum displayed here, P/B ≃ 900 (the same data set as in Fig. 2).

Generally, for SDDs, P/B ≃ 300–500 (Eggert et al., 2003; De Geronimo et al., 2010) is typical under illumination of the whole detector active area. Note that this ratio is dependent on the energy resolution of the detector because the fluorescence line width sharpens and P increases with increased resolution, while the background B is not dependent on the resolution. Hence, reporting P/B ratios also requires metadata affecting the energy resolution to be reported. These include detector conditions (for example the SDD chip temperature), electronics parameters (for example the pulse peaking time) and the total count rate received. An alternative could be to use (Kappen et al., 2001)

In this case, the dependence on the resolution would be lifted. Both P/B and (P/B)′ are unit-less quantities. While (P/B)′ decouples charge loss from detector operations, the disadvantage of this definition is that it depends on the integral boundaries of both the peak and the background.

As mentioned above, P/B ≃ 10² has been reported for fully illuminated SDDs. Rejecting charge sharing, this value is expected to drastically improve. Using local illumination with a small beam away from edges, P/B ≃ 10³–10⁴ can be achieved (Kappen et al., 2001; Eggert et al., 2003). Covering edges with a sufficiently absorbing material to avoid absorption of photons near the edges is thus a viable way to improve spectroscopic performance. De Geronimo et al. (2008, 2010) employed a 125 µm thick molybdenum mask (240 µm wide) for a 16-element SDD and reported an improvement of P/B from 300 to 5000. For a seven-element SDD, Hansen et al. (2008) used a 450 µm thick zirconium mask, realizing P/B ≃ 10³.

While the peak-to-background ratio provides a measure of spectral performance, it is also useful to consider the number of events (N_j) generated in a detector element j as a result of photon absorption in an adjacent element i, with N_i events recorded in that element. A corresponding cross-talk ratio from element i into element j may be defined aswith c_ij ∈ [0, 0.5]. At the limits, c_ij = 0 if all events are registered in i (no cross-talk) and c_ij = 0.5 if N_i = N_j (full cross-talk; half of the total events registered in both i and j). The latter case is expected for photons hitting the detector exactly on the boundary between elements i and j. c_ij values >0.5 are physically not defined as this would mean N_i > 0.5(N_i + N_j) and thus N_i > N_j, which would imply that upon illuminating element i the majority of events are registered in the adjacent element j.

Cross-talk investigations for segmented XAS detectors remain relatively rare. Rossington Tull et al. (1998) showed that a 300 µm thick silicon strip detector with a strip pitch between 100 and 300 µm gave c_ij ≃ 16% for 100 µm wide strips, decreasing to c_ij ≃ 6% for 300 µm wide strips (for 5.9 keV photons and homogeneous illumination). Bucher et al. (1996) used an HPGe detector comprising an array of 2 × 2 square elements of 100 mm² each. Irradiating 500 µm away from an element border using a beam of 100 µm in size (5.9 keV) led to cross-talk c_ij < 10%. Characterization of a seven-element SDD (Kappen et al., 2001) showed c_ij < 0.08% at 150 µm distance from the border between elements under local illumination with a 10 µm beam (E = 10 keV) and c_ij < 0.1% at 50 µm distance from the border. Results from a new multi-element germanium detector array with small 1 mm² elements and CMOS preamplifiers have been presented (Tartoni et al., 2015), giving c_ij ≃ 1.5% for illumination with a 10 µm beam 150 µm away from an element border (E = 15 keV). Accordingly, local illumination in the centre of an element resulted in a high value for P/B of ∼10³.

It is evident that c_ij and P/B depend on the detector geometry. For very small detector elements, the area close to the element borders increases relative to the rest of the active area. On the other hand, making detector elements smaller increases the total count-rate capability of the detector system, which can be desirable, especially at high-flux XAS beamlines and in space-constrained environments such as in situ apparatuses. While it is possible to develop geometric models to optimize detector element size, thickness and shape, the discussion becomes complex in detail. For example, near element borders, c_ij and P/B are functions of the angle of incidence of the photons onto the detector and the photon attenuation length in the detector material. For detectors subtending large solid angles, the angle of incidence can be strongly dependent on the detector-to-sample distance, meaning that in practice a detector geometry model to optimize c_ij and P/B does not have a unique solution for XAS applications because a fixed sample-to-detector geometry cannot be assumed for XAS.

However, photon-rejecting masks offer a practical solution to eliminate or minimize these geometric dependencies and to reduce charge sharing/loss and improve peak to background by a factor of 10–20, as discussed above. Beyond employing masks, it is possible to reject charge sharing in multi-element detectors entirely by using compound detectors which bundle single-pixel detectors into macroscopic assemblies. In this case, charge losses around element borders still need to be considered in order to improve P/B ratios and thus optimize spectroscopic performance and lower the limits of detection for XAS.

Escape events are a special case of charge loss (see Section 3) and have been known and studied for some time (see, for example, McCutchen, 1957; Palms et al., 1968; Christensen, 1979; Rosner & Mingay, 1983; Rossington et al., 1992). These events occur when the material comprising the active detector substrate is excited by the incoming photons and some of the resulting fluorescence photons leave (escape) the detector. As a fluorescence detector will ultimately measure the charge generated upon absorption of the primary photon, escape events lead to some of the charge being lost through radiative loss of the escaping photon. Therefore, escape events manifest as `escape peaks' in the recorded fluorescence spectrum (see Fig. 5), with the escape peak occurring at a specific energy below its parent peak.

Figure 5 Silicon Kα escape peaks with their parent peaks in a fluorescence spectrum recorded with a 55Fe source (the same data set as in Fig. 2).

Table 2 shows the separation energies for silicon and germanium; these are simply the corresponding K and L fluorescence line energies.

The escape probability, and thus the escape-peak intensity, decreases with increasing primary photon energy (Rossington et al., 1992) as photons are absorbed at greater depths in the detector and are less likely to escape from the detector volume. For example, from 5 to 15 keV the attenuation length of photons falling into an SDD increases from 18 to 442 µm, whereas the attenuation length of Si Kα escape photons is fixed at 12 µm. Typical escape-peak intensities relative to the parent peak are summarized in Table 3 [based on the literature (Palms et al., 1968; Christensen, 1979; Rossington et al., 1992; Papp & Campbell, 2001; Lépy et al., 2003; Papp, 2003; Shariff et al., 2004; Thompson, 2009) and observation]. For simplicity, the energy dependence is captured only qualitatively by quoting practical ranges.

The intensity of escape peaks also depends on the detector geometry. Photons can escape through the detector entrance window, the sides (if photons are absorbed close to the edge) or the reverse side (for high-energy photons and sufficiently thin detectors). Using collimators around detector edges is useful to suppress escape events from the sides. For instance, the attenuation length of Ge Kα escape photons in germanium is 50 µm; more than 99% of escape events will originate from a depth up to five attenuation lengths (250 µm). For a cylindrical detector of 10 mm in diameter, the corresponding outer shell of 250 µm thickness makes up 10% of the detector volume. Evidently, the collimator approach relates closely to reducing the charge sharing discussed in Section 3 (for further discussion of collimator geometry, see also Durak & Özdemir, 2001; Can & Bilgici, 2003).

It is interesting to note that escape peaks have found practical applications in other fields, such as investigations of scintillator response curves (Khodyuk et al., 2010), evaluation of fluorescence yields in monocrystalline germanium (Casnati et al., 1984) or energy calibration in gamma-ray detection (see, for example, MacKenzie & Campbell, 1972).

It is also worth noting that escape peaks do not artificially increase the event rate that the detector electronics have to process. In the absence of escape effects, the events would have been registered in the parent peak and the total count rate would be the same as in the presence of escape peaks. As such, escape peaks are already captured as part of detector dead time and its correction.

For the purpose of XAS, escape peaks can be considered to be unwanted while fundamentally unavoidable. There is an energy dependence of the escape intensity across a typical EXAFS scan range; this dependence is strictly monotonic, except for the case of fluorescence XAS at the absorption edge of the detector material itself. In either of the cases, the effect on normalization of EXAFS spectra is difficult to assess empirically, given that escape peaks cannot be avoided. This may call for further theoretical work to be performed, drawing on previous efforts to simulate detector spectral responses (see, for example, Campbell et al., 1998, 2001; Papp, 2003; Shariff et al., 2004). For further discussion, the energy dependence will be assumed to be linear over the range of a typical XAS scan. Thus, the fraction of escape events from the parent peak contributes only a linear term to the XAS background, and for moderately concentrated samples the variation of escape intensity is dominated by the increase in the parent peak whilst scanning through the absorption edge. In these cases, the presence of escape peaks only scales the spectrum linearly.

There are, however, instances where escape peaks may be detrimental to XAS. One special case occurs when an escape signal and another fluorescence signal pile up in such a way that they contribute to the fluorescence line of interest. For example, using a germanium detector to analyze traces of cadmium in a manganese-rich sample, we would find pile-up events (sum peaks) resulting from Mn Kα and Ge K escape that, above the Cd K edge, would move with the incident photon energy into the region of interest around the Cd Kα fluorescence line. The probability for this to occur would depend on the total count rate (see Section 2) and may make an ∼0.5% contribution to the Cd Kα fluorescence. Depending on the cadmium concentration, this could lead to background distortions that do not normalize properly. However, in cases such as this a practical remedy would be to use 100 µm thick aluminium foil to filter out 95% of the fluorescence from manganese.

While this example is quite specific, and perhaps artificial, similar issues can arise when the escape peak alone moves into or through a fluorescence peak of interest. This only becomes a problem for very dilute systems coupled with strong scatter peaks resulting in appreciable escape signals. The specifics are dependent on the sample and beamline setup (air and other scatter paths). As a practical guide, a heavier element (X) present at 10 p.p.m. in a light matrix (M) may result in a fluorescence response with I(X)/I(M) ≃ 10⁻². Using an SDD, the escape peak intensity is about 5 × 10⁻³, meaning that in this example the escape peak from the elastic peak contributes about 50% to I(X). This contribution is energy-dependent as the escape peak moves with E and thus can distort the XAS spectrum background, leading to normalization problems. Similar arguments apply to germanium detectors, where the Ge L escape peak occurs at E = 1.19 keV, although the Ge L escape intensity is ∼10 times less than that of silicon.

Importantly, however, problems only potentially arise at some absorption edges, depending on the scan range (XANES versus EXAFS) and the element of interest. Figs. 6 and 7 summarize possible conflicts arising from silicon escape. For instance, XANES at the Cu K edge is not affected as the silicon escape peak stays well below the Cu Kα fluorescence line. For EXAFS, however, the escape peak will move into the Cu Kα line for the last 200 eV of the EXAFS scan.

Figure 6 Interference of the silicon escape peak with fluorescence peaks for XANES at selected K and L3 absorption edges (scan range −70 eV ≤ E0 ≤ +300 eV).

Figure 7 Interference of the silicon escape peak with fluorescence peaks for EXAFS at selected K and L3 absorption edges (scan range −150 eV ≤ E0 ≤ +1000 eV).

These issues can place fundamental limits on XAS at the dilute level. Further research and documentation of these non­linearity effects on XAS spectra is required to fully evaluate the linearity boundaries. However, with escape peaks being intrinsic features of germanium and silicon fluorescence detectors, mitigation of the effect may require either advanced data processing (modelling of individual fluorescence spectra and separation of escape intensities) or employing secondary fluorescence analyzers through methods such as high-energy resolution fluorescence detection (HERFD).

In the modern context, XAS needs to respond to the demands of researchers to study elements that are present at very low concentrations (p.p.m. to p.p.b.) or to follow reactions and processes in operando on a timescale of minutes to seconds. Enabling XAS at the p.p.m. to p.p.b. level is particularly relevant for, for example, studying biological systems at realistic natural levels or investigating environmental trace contaminants close to legislative limits. Performing in operando XAS with some time resolution is relevant, for example, for monitoring reactions in catalysts at moderate concentration levels (for example 1000 p.p.m.). In some cases, specific experimental setups are space-constrained, thus also requiring detectors with a fairly small footprint. In this coupling, detectors with high fluorescence count-rate capabilities (∼10⁶ counts s⁻¹; Mcps) at good energy resolution and a high peak-to-background ratio are key to answering challenging questions and addressing those issues that help to improve the way we live, work and eat.

It is important to deal with detector nonlinearities to ensure the integrity and quality of XAS data and promote their translation into research outcomes. Of the nonlinearities discussed in this chapter, dead time has the most wide-ranging influence on XAS data; in the extreme case it can lead to the extinction of XAS amplitudes. The dead-time behaviour of a detector system depends on operator choices, and within reasonable limits the nonlinearities arising from dead time can be corrected. Alongside this correction, the very high count-rate regime (several Mcps) thus also demands a careful consideration of pulse pile-up to maintain appropriate system linearity. Modern digital pulse-processing systems use, for example, deconvolution methods to recover piled-up pulses (Scoullar et al., 2011) or pattern-matching techniques to improve pile-up rejection (Mott, 2010). The diffusion of such or similar systems to, at least, high-flux XAS beamlines is an important element of XAS development globally.

The other nonlinearities discussed generally have a less significant influence on XAS. While escape peaks are an intrinsic property of the detector material and thus are unavoidable, they are only relevant in specific cases and primarily at low concentration (p.p.m. range) of the element of interest. Contributions from pile-up at moderate count rates (a few hundred thousand counts s⁻¹) are typically small (∼0.1% or below) and can be controlled through sufficiently fast pulse-processing electronics and by operating in count-rate regimes that are adequate for the electronics at hand. Nonlinearities due to charge sharing are mitigated by physically implementing a mask on the detector surface. To the beamline user, charge sharing should thus be of no concern.

In general, nonlinearity solutions or algorithms implemented in hardware should be preferred over modelling and `post-acquisition' software, because this removes nonlinearity issues closer to the source and enables the user to focus the research.

Future perspectives of dealing with detector nonlinearities are likely to be governed by count-rate capability, and thus pulse processing needs to satisfy scientific demands. High-flux environments or faster measurements to follow processes in operando will continue to drive the demand for faster photon detection while maintaining high energy resolution. There are fundamental time limitations in the photon detection and processing chain that will ultimately constrain the ability to accept higher event rates. Charge-collection times in the detector active material and corresponding minimum pulse rise time are one limit; ADC sampling speeds are another. The solution will likely come in the form of detectors with higher degrees of parallelization, possibly into the hundreds of detector elements, to maintain high energy resolution at comparably low event rates per detector element. Corresponding development efforts will require sustained investment directed at the mid to long term. General market-driven investment alone may not be the right mechanism due to the overall market size and time frames involved. Similar to imaging detector development, development nucleated at synchrotrons or other large-scale research institutions coupled with commercialization at a later stage may be more promising.

For the XAS community, the challenge will likely be to continue driving the push for detector development and, ideally, find overlap with other scientific communities to increase critical mass and open market opportunities to make investment in development as attractive as possible.

Note: This chapter was originally written in 2018 and should be read in that context.