Defects in semiconductors consist of two major types, namely, point defects and extended defects^[1-4]. Types of point defects include substitutional or interstitial foreign atoms, vacancies, anti-sites, or their complexes, and these defects usually involve one or a few atoms at most. Conversely, extended defects, such as threading dislocations or grain boundaries, can involve thousands or more atoms that are displaced from their regular lattice sites. One distinct difference between the two types of defects lies in that one single extended defect may exhibit much higher density of states than that of many point defects jointly for a moderate level of point defects. Since point defects can often be saturated with high carrier density, they will have greater impact for applications involving low carrier densities, whereas extended defects will be more detrimental for high carrier density applications^{[5, 6]}. In practice, these two defect types will often play competing roles in a device. For example, increased carrier density may lead to larger carrier diffusion lengths, which will worsen the adverse effects of extended defects, since they may deplete a large number of carriers within the diffusion length^{[5, 7]}. In addition, point defects typically degrade device performance, whereas extended defects will not only have the same effect but they may also be responsible for the ultimate, fatal device failure. Empirical correlations between the formation of so-called dark-line defects and degradation of GaAs semiconductor lasers were reported decades ago^[8]. However, obtaining direct evidence is usually more difficult. It was shown recently that a simple dislocation defect could mutate under high-density optical injection conditions into a far more detrimental dark-line defect^[5]. Neverthless, it still remains to be shown how the defect mutates under electrical injection in a practical device.

There are many different methods available for characterizing the structure and properties of individual defects correlatively^{[5, 9-19]}. In particular, advanced microscopy techniques can nowadays provide detailed knowledge about structural defects in semiconductor materials and devices. Nevertheless, carrying out operando characterization at an individual defect level is often highly challenging or even impossible. Thus, determining the impact of specific types of defects on device performance and failure mechanisms is rarely answered systematically nor unambiguously. Some open questions include the following:

(i) How is device performance affected by an individual defect? Many such defects, including their atomic structure, can be imaged but the defects are typically not studied in functioning devices.

(ii) In what way does the impact of the defect depend on the device operation conditions? For instance, some defect types are likely to be more problematic at low carrier density, whereas others might be more detrimental at high carrier density. Such conditions could, for example, apply to a solar cell operating at one sun vs. 1000 suns or to an LED operating at 10 mA vs. 1000 mA.

(iii) How does the impact vary from one defect to another? One specific type of defect might be thought to be worse than another, but it is impractical to construct specific defects with known microscopic structure in a device and then compare their impact.

This type of information would enable a much more concrete understanding of the crucial role of defects in device operation, particularly providing the knowledge for assessing the need and possible strategies for removing or mitigating their impact.

Despite the long history of semiconductor defect research, the efforts have not always been very effective. Defect studies have mostly been performed in a parallel mode. One wafer is divided into multiple pieces used separately for structural, optical, and electrical characterization, as well as device fabrication and testing, as illustrated in Fig. 1: confocal photoluminescence (PL) imaging in one piece identified some defects that quenched the PL signal; high-resolution Raman imaging in another piece showed enhanced Raman signal at certain defect sites; high-resolution transmission electron microscopy (HRTEM) in yet another piece observed a dislocation defect and determined its detailed structure; and one more piece was fabricated into a solar cell on which an individual defect was located, for example by PL imaging, and this defect was further found to significantly reduce the photovoltaic efficiency when a laser beam was focused at the defect location. This commonly adopted approach of defect study clearly yields a lot of interesting information about the defects present in a particular material, and the general practice appears to be efficient. Unfortunately, it is incapable of yielding conclusive information about how the specific defects might affect the device performance or provide definitive answers to the crucial questions posed above. The major limitation of this approach is that either the defect investigated by structural characterization techniques is not the same defect that is affecting the device performance (i.e., they are presumed to be similar defects but are nevertheless taken from different pieces of the material) or the defect is not characterized under normal device operating conditions.

There are in fact many well-known defect characterization techniques that provide qualitative assessments of material quality, such as chemical etching where the observed etch pit density is used as a measure of defect density and often correlates well qualitatively with device performance. However, it is usually unclear what effects a defect corresponding to a specific etch pit would have if the material were used for fabricating a real device, for instance, a solar cell. Attempting to address this question, we decided to correlate PL imaging, chemical etching, and scanning electron microscopy (SEM) imaging of CdTe epilayers by carrying out the measurements in the same order as they are mentioned^[13]. Interestingly, it was observed that the PL dark spot density was much higher than the etch pit density. It was tempting to explain the discrepancy between the two densities in terms of the difference in sensitivity of the two techniques. However, unexpectedly, despite the lower density, etch pits did not always match the locations of the dark spots visible in PL imaging^[13]. These findings suggest the need to determine which defects are actually impactful in a real device. Although the PL imaging results in this specific study were somewhat closer to the effect of defect recombination in real devices, the PL process is in general unable to reveal how carriers are generated, injected, or extracted in real electronic or optoelectronic devices, such as transistors, photo-detectors, solar cells, and LEDs.

This background has inspired us to embrace a more comprehensive approach to defect study, namely applying an array of correlative and spatially-resolved optical, electrical, and structural characterization techniques, all at the individual defect level but applied sequentially, i.e., in a series mode^{[20, 21]}, as illustrated in Fig. 2. Specifically, after a device is fabricated, individual defects on the device are identified. Then, the impact of individual defects is thoroughly investigated by applying different techniques as needed in a proper order, such as electroluminescence (EL), PL, micro-Raman, and micro-scale illuminated current–voltage (I–V) measurements^[22]. As the last step, the atomic-scale defect structures for the same defects are determined using HRTEM. It has been shown previously that defects may be modified or mutate during device operation^{[8, 23]} or under high density photo-excitation^[5]. To investigate defects in their as-grown states, care should be taken to avoid unintended modifications to the defect structure by limiting the illumination light density or injection current level to below any mutation thresholds, so that the native defect structure can be obtained by HRTEM. Our past efforts were in this category. On the other hand, for the purpose of investigating the possible role of the defects in the device failure mechanism, one can instead monitor the material properties and device characteristics at the defect sites (e.g., local spectroscopy signature, I–V characteristic under illumination, temperature) by continuously increasing the carrier injection level to reach the critical point.

In this focused review, we describe some examples from our recent work^[20] to illustrate how in operando and correlative characterization at the single defect level, but implemented in the series mode, can provide not only a comprehensive understanding of the defect at a fundamental level but also unprecedented insight into the real impact of an individual defect on a real device (e.g., a GaAs solar cell). The findings enable us to explicitly answer those questions above that are otherwise impossible to address in the traditional parallel approach to defect characterization. We will also comment on an in operando micro-Raman 3-D thermometry approach^[24] that can be used for providing a high-spatial resolution probe of high power semiconductor devices with high-temperature accuracy, such as LED and power electronic devices, and further expand the capability of the general approach.

Because of the relatively low density of dislocation defects (on the order of 10³ cm^–2 or lower) in the GaAs epilayer, it is inefficient and even challenging to identify individual defects using PL via raster scan. In a simple GaAs double heterostructure (DH) sample (e.g., GaInP/GaAs/GaInP), a PL imaging system was normally used to first identify the approximate locations of individual defects over a macroscopic area in the sample^{[5, 25]}. Then, high-spatial-resolution PL mapping was used to determine its accurate location, followed by detailed optical investigation. However, in the finished device (the details of the device structure can be found in the original paper^[20]), the PL signal from the active layer (i.e., the absorber), which was the layer of the main interest, typically became much weaker than that in the simple DH sample. The primary reason for the weakened PL signal was that the n⁺p GaAs junction depleted the photo-generated minority carriers in the p-region or the active layer, which made it more difficult even using the PL imaging technique to locate individual defects. EL imaging was instead used for surveying a large device area, because it typically provided higher sensitivity and efficiency in locating individual defects that were detrimental to the electrically injected or photogenerated carriers. As shown in Fig. 3(a), EL imaging located a few potential threading dislocation defects (labelled as A, B, C) that exhibited quenched EL emission on a cell area named #5-2 (#5 for cell number, #5-2 for one cell area separated by electrode fingers). To ensure that these sites were not due to some unintended surface blemishes, we compared the EL image with an optical image Fig. 3(b), which indicated that the observed dark features in the EL image (indicated by the red dots in the optical image) were not surface effects but were generated from the interior of the sample. In this area, some surface blemishes were visible Fig. 3(b), which were actually unintended carbon-like parasitic depositions that did not match the exact defect locations^[20]. Once the approximate defect locations were identified, they could be interrogated more closely using various high-resolution optical characterization techniques.

Observations with high-resolution PL and Raman mapping scans were then used to confirm that the suspects located by EL imaging are indeed dislocation defects, because dislocation defects tend to exhibit certain distinctive features in PL and Raman^{[5, 26]}. Fig. 3(c) shows the PL mapping result at the PL peak wavelength of a general site (~868 nm) near the triplet in Fig. 3(a), with the comparison between the PL spectra of the largest defect site (#5-2A) and a general site given in Fig. 3(d), under an excitation density of 5.6 × 10⁴ W/cm². As expected, the PL intensity is much weaker at the defect site than the general sites. Interestingly, at the defect site the PL peak energy is found to be blue-shifted to 863 nm compared to 868 nm at a general site. This finding is fairly common to similar defects observed in this and other GaAs samples, although the amount of peak shift varies. The possible reason for the shift could be the existence of a compressive in-plane strain field in the vicinity of the defect, although exact mechanism remains unclear. Fig. 3(e) shows Raman mapping result near the largest defect #5-2A at the peak position of the defect site, together with Raman spectra from the defect site and a defect-free site, shown in Fig. 3(f), measured under the same condition as for the PL mapping. The defect-free site shows two GaAs Raman modes: 296 cm⁻¹ for longitudinal optical (LO) mode and 268.4 cm⁻¹ for transverse optical (TO) mode, while the defect Raman shows a sharper, stronger and red-shifted LO mode at ~293 cm⁻¹ and a slightly weaker TO mode at 269.1 cm⁻¹. Note that the intrinsic GaAs Raman modes are reported to be 268.1 (TO) and 291.4 cm⁻¹ (LO)^[27] and our measurements for a GaAs: Cr sample yielded 268.0 and 291.5 cm⁻¹. At the first glance, the Raman results were counter intuitive: the spectrum of the defect site is closer to that of the intrinsic GaAs. In fact, the differences in intensity and frequency shift of the LO Raman mode can be attributed to the difference in carrier density between the defect and defect-free sites^[26]: because the above-bandgap excitation generates an electron plasmon and the measured Raman signal results from the coupled mode of the LO phonon and plasmon (LOPP)^[28], the lower carrier density at the defect site makes its Raman spectrum more resembling that of the intrinsic GaAs.

A dark I–V characteristic reflects the average property of the device, which obviously cannot probe individual defects. However, under the illumination of a focused beam, the inhomogeneity of the device can be investigated. By raster scanning the laser beam while monitoring the short-circuit current of a PV cell, one can have a so-called laser-beam-induced current (LBIC) map^[22]. One can further park the laser beam at one location to measure the light I–V characteristic. Such obtained light I–V characteristic can be rather similar to that of the light I–V characteristic of the cell as a whole illuminated by light, as in the case of the GaAs solar cells that were studied. However, for the CZTSe thin-film cells studied previously^[22], it was found that local light I–V curves were very different from that of the whole cell light I–V curve.

It is not always possible to make direct comparison between the macroscopic and microscopic I–V characteristics. The reason is that the device is operated under very different conditions between the two cases. Only a very small area is illuminated in the microscopic measurement, which is approximately equivalent to a solar cell with a small active area that is under illumination and a large electrode area, the whole non-illuminated area, that is under forward bias, thus resulting in much higher dark current. Hence, the microscopic light I–V curve tends to show reducedV_oc and FF, which was very significant for the CZTSe cells^[22] but much lesser in the GaAs devices, because the latter had an overall lower dark current and a larger shunt resistance than other thin-film solar cells.

Fig. 4 shows the microscopic light I–V characteristics measured at the location of the largest defect #5-2A identified in Fig. 3(a) under different illumination powers. For reference, the light and dark I–V curves for cell #5 as a whole are shown in Fig. 4(a). The dark I–V curve indicates a well-behaved diode characteristic. The light I–V curve is obtained using a solar simulator with a power density of 850 W/m², which gives rise to I_sc = 0.125 mA, V_oc = 893 mV, and P_max = 0.0892 mW. Using the total active device area (the sum of all the strips) of 0.767 mm² or P_L = 0.652 mW, the energy conversion efficiency of this small cell is η = P_max/P_L = 13.7%, and the fill factor is FF = P_max/(I_scV_oc) = 0.799. Figs. 4(b)–4(d) show results obtained using a 532 nm focused laser beam (with a 50xLWD/ND0.5 lens) under three laser powers (roughly over three orders in magnitude), P_L = 213, 18.5, and 1.82 µW, respectively. Key cell parameters were calculated from the measured data given in Table 1.

All cell parameters are found to degrade at the defect site, regardless of the illumination power level, but the extent of degradation depends on the illumination density. From the high to low power, the reductions in efficiency are 72.3%, 66.5%, and 68.2%, respectively, for the three laser powers. For this particular defect, the variation in the magnitude of efficiency reduction does seem to depend on the illumination density. Because the laser spot size is much larger than the defect core, theI−V characteristic of the “defect site” actually reflects an average effect within the laser spot size which is comparable to the lateral carrier diffusion length in this device. These results provide semi-quantitative assessment of the impact of an individual defect in a solar cell. However, it is non-trivial to evaluate the impact of an individual defect exactly in a device, which depends on the impacted area relative to the total illuminated area. Despite the significant efficiency reductions, the ratios are nevertheless much smaller than those in PL, because PL is measured under the open-circuit condition where the nonradiative recombination is much more competitive than the radiative recombination. In fact, the GaAs materials comparable to the one used in this solar cell exhibited relatively low internal quantum efficiency in PL at the low excitation density (e.g., 1 sun)^[6], but they could, however, yield reasonably good cell efficiencies, benefiting from the strong carrier extraction power under the non-open-circuit condition.

The light I–V characteristic of a well-behaved solar cell can be described by the standard model^[29]:

$ I={I}_{\mathrm{L}}-{I}_{0}{(\mathrm{e}}^{\frac{q\left(V+I{R}_{\mathrm{s}}\right)}{nkT}}-1)-\frac{V+I{R}_{\mathrm{s}}}{{R}_{\mathrm{s}\mathrm{h}}} , $  (1)

where I_L is the photo-generated current, I₀ is the dark reverse saturation current, V is the photo-voltage generated by the solar cell, n is the diode ideality factor (n = 1 for an ideal diode), R_sh is the shunt resistance (ideally infinity), and R_s is the series resistance (ideally zero). Additionally, q is the electron charge, k is the Boltzmann’s constant, and T is the absolute temperature. Eq. (1) is not, strictly speaking, applicable for the focused illumination. However, since the obtained I–V curves under focused illumination appear similar to a typical well-behaved solar cell, they can be fitted well using Eq. (1). The fitting curves are plotted in Figs. 4(b)–4(d), and show excellent agreement with the measured data. The effective n, R_sh and R_s values are obtained from the fitting, and are listed in Table 1. Quantitative comparisons of I₀ and n between either uniform and focused illumination or defect-free and defect site are not straightforward to make, other than noting that I₀ is significantly larger for focused illumination, because the dark area is much larger for the latter case. A more important finding between the defect-free and defect sites is that the latter consistently yields lower shunt resistance under the same illumination power, and more so for higher illumination power. This trend can be explained as follows: firstly, the dislocation defect provides a shunt to the photogenerated carriers; secondly, the effect is more significant for high excitation density, because the effect of the point defects is saturated, and the dislocation defect becomes more prevalent^{[5, 7, 25]}.

Several threading dislocation defects were characterized by PL, Raman, and I–V characteristic under the same conditions, as shown in Fig. 5, where the results are compared for the three neighboring defects shown in Fig. 3(c), #5-2A, #5-2B, and #5-2C, and another defect, #5-3A, from area #5-3. Figs. 5(a) and 5(b) are, respectively, the PL and Raman mapping results near defect #5-2B & C. Figs. 5(c) and 5(d) are the same results for defect #5-3A, and they are qualitatively similar to the corresponding results for #5-2A, Figs. 3(c) and 3(e). Figs. 5(e) and 5(f) compare, respectively, the PL and Raman spectra of the four defects and the defect-free site. Evidently, as shown in Fig. 5(e), the PL peak positions of the GaAs peak are rather different for different defects, but all are blue-shifted from the defect-free value: 858–867 nm or δE = 17–2 meV relative to 868 nm of the defect-free site. Within #5-2A, #5-2B, and #5-2C, the amount of reduction in PL intensity seems to correlate with the visual size of the defect with the largest defect #5-2A, shown in Fig. 3(c), also exhibiting the most intensity reduction. Fig. 5(f) compares the Raman spectra of the four defects. There is qualitatively an anti-correlation between the PL intensity and LO mode Raman intensity among the four defects, which can be explained by the opposite dependencies of the carrier density for PL and Raman ^[26]. Figs. 5(g) and 5(h) depict their I–V characteristics under two illumination powers (2.1 and 225 μW) for comparison between the four defects. The severity of impact is in the order of A, B, and C, with A being the strongest for the three defects in area #5-2, whereas the impact for #5-2A and #5-3A are comparable.

Once optical and optoelectronic characterization of individual defects are completed, the next logical step is to investigate the microscopic structure of the same defects. However, this specific step has rarely been taken in the past defect research. In our study, the defect locations, defined as the lowest PL intensity positions, were carefully noted under high magnification optical microscope imaging (with about 0.1 μm accuracy), typically with respect to some nearby recognizable surface feature, so that they could be located again using the SEM in the dual-beam FIB system that was used for preparing the cross-sectional TEM samples. Several defects were investigated. Results are highlighted here for two of them: #5-2A and #5-3A.

Fig. 6 shows low magnification TEM and atomically-resolved scanning TEM images of #5-2A. The PL image in Fig. 3(c) had already indicated that this defect was laterally more extended than most of the other defects. A thin specimen slice was extracted using the FIB milling system, from the device at the approximate location indicated on the PL image included with Fig. 6(a). This image reveals that this defect cluster is mostly confined to the GaAs emitter layer and penetrates only a short distance downwards into the GaAs absorber layer, although it extends laterally by more than one micron. However, the GaInP window layer above the GaAs layer in this region seems to be free of defects, as shown in Fig. 6(b), which is an enlargement of the region indicated by the yellow box in Fig. 6(a). This defect is clearly a massive cluster combining many individual defects. Some of these areas were examined more closely, and high-resolution images were taken from the four areas indicated in Fig. 6(a). Areas 1 and 2 include the GaInP/GaAs interface. The structures of the area 1 and 2 are similar, as shown in Fig. 6(c) for area 1, which indicates that the top region of the device (part of the GaInP window layer) is amorphized. The area 3, shown enlarged in Figs. 6(d)–6(f), has a stacking fault, where the end of the stacking fault marked by a black square in Fig. 6(e) ends in a 30° partial dislocation while the other end terminates in a 90° partial dislocation. An enlarged view is given in Fig. 6(f). According to detailed analysis of a high-angle annular-dark-field (HAADF) STEM image of the same area (not shown here), the single atomic column at the core of the 30° partial dislocation (marked by the white arrow) consists of arsenic. Area 4, as shown in Figs. 6(g) and 6(h), exhibits a 60° dislocation near the top of the image, and the extra half-plane of paired columns is indicated by the white line in Fig. 6(g). An enlarged view of this dislocation is shown in Fig. 6(h): the Burgers circuit is shown in white, with the resulting Burgers vector shown in red. The extra half-plane of atoms is marked with the black line.

Fig. 7 shows the results of the microscopic characterization of defect #5-3A, which is structure-wise distinctly different from #5-2A, and has been shown to have comparable impact on the optical properties and device performance (Fig. 5). In the optical image of the device (not shown), there is a white spot right next to the EL dark spot at defect #5-3A. The low-resolution TEM image in Fig. 7(a) reveals that the white spot corresponds to a pit-like feature in GaAs near the device surface, which is covered over by the GaInP window layer, while the dark spot in the EL map is caused by the defect cluster adjacent to the pit. This defect cluster extends mostly along {111}-type directions from the surface of the solar cell down into the absorber material for at least 2 μm, but not apparently all the way through to the back-surface confinement layer. However, it is far less extended laterally compared to #5-2A. It includes many structural defects, which are mostly concentrated in the top 1 μm of the absorber layer. Fig. 7(b) is a HAADF image showing a major intersection of defect clusters and Fig. 7(c) is a corresponding large-angle BF (LABF) image taken at higher magnification. Fig. 7(d) is a LABF STEM image of the area in the white square in Fig. 7(c), and shows a single 30° partial dislocation that terminates an intrinsic stacking fault. The dislocation type is identified in Fig. 7(d) by drawing a Burgers circuit around the defect, as shown by the red arrows. The resulting projected Burgers vector is indicated by the red arrow. The extended stacking fault is terminated by a single, unpaired atomic column indicated by the white circle in the figure, which is identified as corresponding to As.

Each of the defect clusters was unique, yet they shared common characteristics. Visual comparison of the different defect clusters showed that every cluster was dominated by a collection of extended {111}-type stacking faults, mostly located near the top of the device in the GaAs emitter and absorber layers. In contrast to the common belief that dislocation defects in the GaAs epilayer originate from the substrate that happens to have a comparable defect density, these defects were typically initiated within the absorber layer at around 0.5 to 1.5 μm from the lower GaAs/GaInP interface, and passed through the GaAs emitter and the GaInP window layer. However, one defect cluster (#5-2A), with relatively short stacking faults, remained mostly confined to the n⁺ GaAs emitter layer, leaving the GaInP window layer nearly free of defects, and only extended down into the GaAs absorber layer over a short distance (~0.25 μm). It might seem counter-intuitive that the GaInP PL signal is often weaker from the non-defective GaInP site (#5-2A) than from the defective GaInP site (#5-3A), as shown in Fig. 5(e). The explanation could be that the carriers in the former case are mostly captured by the underneath GaAs layer, whereas the carriers are localized to the GaInP layer itself in the latter case.

Although we did not observe significant change (enhancement) of the TO mode intensity at the defect site, as previously suggested resulting from relaxation in selection rule due to defect induced disordering^{[30, 31]}, we observed a strongly enhanced TO Raman mode at the location of the pit, shown in Fig. 7(a), which has {111}-like surfaces for which TO Raman scattering is allowed. We note that the lateral extensions of the defects identified in these epitaxially-grown solar cells are much smaller than the dislocation defects observed in GaAs ingots, which are intended to be used as substrates, with lateral sizes in the order of 100 μm^{[30, 31]}. Moreover, dislocation defects originated from the substrate, e.g., in SiC^[19], are often found to be much larger in lateral size.

Overall, these observations reveal that those defects showing major impact on optoelectronic performance (e.g., reduction in PL intensity and photo-current) were most often associated with a cluster of extended defects rather than a single extended defect, as previously assumed based on optical data alone. Those extended defects that exhibited weaker impact on the material properties and device performance have not been carefully examined. Quite likely they were simpler defects, which remain to be investigated.

One further step towards advancing this correlative approach to defect characterization would naturally be to perform density-functional theory (DFT) level structural and electronic calculations using the HRTEM results as inputs. However, realistically, the defect complexes, such as those discussed above (e.g., #5-2A and #5-3A), are too challenging for the most capable DFT codes to deal with at this time. Nevertheless, one could begin with those well-defined sub-structures for the defect modeling.

The series approach illustrated above has provided unique and unprecedent insights for both practical knowledge of the adverse effects of defects and fundamental understanding of defect physics at the single defect level. The impact of the dislocation defect on the solar cell performance parameters tends to be more significant at higher illumination density, because of the competing roles of point and extended defects. In addition to reduction in photo-current and open-circuit voltage, the dislocation defects also reduce the shunt resistance, due to enhanced recombination loss through the defect states. The degree of impact varies significantly among different defects, and the trend is mostly consistent between the spectroscopy signatures (e.g., PL intensity) and optoelectrical characteristics (e.g., I–V characteristic). In the future, the same approach can be used to probe devices under high injection current or high illumination density to induce structural modifications of the defects, and then to investigate the effects.

Several techniques are available to measure device temperature, from perhaps the most straightforward method – IR imaging^[32] to more sophisticated methods, such as lock-in thermography^[33], thermo-reflectance^[34], and Raman thermometry^[35]. These techniques are typically only capable of or were only used for probing the surface temperature of a device. An in operando micro-Raman based 3-D thermometry has recently been developed with diffraction-limit spatial resolution and high temperature accuracy of ±2 ^oC^[24]. While measuring Raman signal from the volume below the device surface is not difficult with a confocal micro-Raman system when a below bandgap laser is used, the salient aspect of this technique is its ability to suppress the interference of the below bandgap emission of an InGaN LED when it is operated under CW injection. The general idea of the technique is to drive LEDs under a continuous current for a relatively long period of time to allow the device to reach a steady operational state, then switch the power off briefly to take the Raman spectrum in the middle of the off-time window. This “split-time-window” method takes advantage of the fact that the electronic relaxation time is much faster than the thermal relaxation time so that the luminescence decays quickly after the device is turned off but the thermal distribution remains nearly the same as in a truly continuous mode during the Raman data collection window. Additionally, in the past, Raman frequency shift was often used as a temperature calibration, which could be problematic, because temperature change also leads to stress relaxation that also contributes to the Raman frequency shift. Therefore, the calibrated relative intensity ratio of Stokes and anti-Stokes Raman scattering is used as the temperature metric in the new approach.

Fig. 8 shows the LED device and the 3-D locations that were measured in this demonstration. Fig. 8(a) shows the optical image of the device, and Fig. 8(b) the cross-section SEM image. As shown schematically in Fig. 8(c), a total of 60 locations were measured on one vertical cross section area at different depths from the sapphire substrate/GaN buffer layer interface to the top surface of the device.

The results are shown in Fig. 9 under 350 mA current (for a 1 mm size LED chip). Fig. 9(a) depicts the measured temperature contour in the x–z cross section. The overall temperature profile looks like a columnar shape along the vertical direction. Assuming the LED junction temperature is the same as the top surface temperature, we find that at 350 mA the average temperature difference is only approximately 2 °C between the top and bottom of the GaN epilayer, as shown in Fig. 9(b). However, we note that the magnitude of temperature fluctuation within the same depth can be significantly larger than the overall laterally averaged variation in the depth direction, because for various reasons the device simply does not have an axial symmetry to justify treating it as a 1-D system. Within the cross section, the maximum temperature of 86.2 °C (location no. 4) is measured at the top surface and a minimum temperature of 74.8 °C (no. 1) is measured at the bottom surface close to the sapphire substrate. Raman spectra of a few extreme points (no. 1 to no. 4) are included in Fig. 9(a). Moreover, we find that, despite some fluctuation, on average the GaN temperature near the top of the patterned sapphire pillars is lower than that of the region between the top sites at the same height. Fig. 9(c) plots the temperatures from 32 locations for each of the two types of sites and the averaged difference is approximately 5.7 °C. We note that the density of the hot spots is much lower than the typical defect density in this type of material, as shown by the TEM image from a similar device, Fig. 8(d).

Clearly it is of great interest to further investigate the microscopic structure of the identified hot spots using HRTEM, as we have done for GaAs solar cells to correlate the spatially resolved optoelectronic characterization with structural analysis^[20]. One can continue monitoring the temperature distribution to even higher injection current to assess the device failure mechanism, as well as measure the temperature effect of the spatially resolved efficiency droop^[36].

The 3-D Raman thermometry is equally applicable to other devices, such as monitoring highly non-uniform heating, potentially the failure mechanism, in power electronic devices. More effectively, one can apply a similar correlative approach demonstrated for the GaAs solar cells to power electronic devices based on III-nitrides, SiC, and Ga₂O₃, which likewise face device fabrication, performance, and reliability challenges related to extended defects. For example, we recently observed the morphology of failed GaN-on-GaN vertical p–i–n diode devices after reverse-bias electrical stressing to breakdown, and found large surface craters that were ~15–35 μm deep with lengthy surface cracks, and substantial concentrations of threading dislocations around the cracks and near the crater surfaces^[37]. Moreover, as illustrated in Fig. 10, we have recently found two different breakdown mechanisms in stress-tested devices. By applying the proposed methodology to GaN power devices during biasing and stress testing, we should be able to directly monitor development of the current leakage process and identify the underlying cause(s). Moreover, by continuously measuring the relevant physical parameters at locations showing excessive temperature rise, it should be possible to determine the critical point when irreversible damage occurs.

Using GaAs solar cells as prototype devices, we have demonstrated how a series approach to device characterization, which offers comprehensive, in operando and correlative information about defects in a semiconductor device, can lead to answers to the important questions that would otherwise never be answered with certainty. We anticipate that this novel approach, coupled with the various spatially-resolved characterization techniques described in this review, can be more broadly applied to investigate the physics of defect and the roles that they play in real devices under different operation conditions, and to assess practical device reliability issues.

This work was partially supported by ARO/Electronics (Grant No. W911NF-16-1-0263). YZ acknowledges the support of Bissell Distinguished Professorship at UNC-Charlotte.