Identifying and investigating spatial features in InGaAs solar cells by hyperspectral luminescence imaging

Electroluminescence (EL) and photoluminescence (PL) imaging have been well-established as highly useful methods for the characterization of photovoltaic materials and devices, enabling the visualization of cracks, defects, grain boundaries, and grid interruptions,^1–5 as well as the calculation of quantities such as local voltage and minority carrier lifetimes in cases where the spectral composition can be assumed to be relatively uniform across the sample.⁴ 

Hyperspectral EL and PL imaging adds high-resolution spectral data on a pixel-by-pixel basis to conventional luminescence images. As it is a relatively new technique, the applications are still being explored. Often, this means using averaged or selectively chosen hyperspectral data to find quantities that are more commonly calculated from global spectral luminescence measurements or developing methods to create maps of such quantities across the sample.^6–8 It is particularly noteworthy that the spectral resolution allows for full calibration of the data, regardless of the luminescence spectrum or its variability across the sample.

The high-resolution spectral data obtained with hyperspectral imaging also enable a more in-depth analysis of features that can be identified with conventional imaging. It can reveal features that would not be identified at all in conventional images, even when separate spectral luminescence measurements are performed.

In this work, we explore these capabilities, applying temperature-dependent hyperspectral luminescence imaging to the study of dilute-alloy InGaAs solar cells. We find and identify local areas with a high concentration of radiative defects, which drain carriers and reduce cell performance. We also identify distinct patterns of material variation that influence both radiative and non-radiative recombination over a full sample. The extensive data contained in a small number of measurements provide insights that can be used to identify the origin of such features and how fabrication processes and device design may be adjusted to improve device performance.

Experiments were performed on 2 × 2 cm² InGaAs commercial solar cells with a p-type base and an n-type emitter as depicted in Fig. 1. The in-house measured power conversion efficiency (PCE) of typical cells under air mass 1.5 G is ≈ 24%. The In concentration was estimated to be around 2% by optical modeling. Hyperspectral EL and PL measurements were performed on several devices, with the results presented here being representative of the phenomena observed in all samples. This paper presents hyperspectral image cubes—the 3D data structures containing a spectrum at each pixel of an image—taken under the PL mode with a 20× microscope objective and under the EL mode with a 20 mm field of view and a 5× microscope objective. Other magnifications are available and have also been studied.

Figure 1 also shows schematics depicting the operation of the hyperspectral system and its main components [(a) and (b)]. The hyperspectral filter box contains several gratings that are used to acquire images over a wide spectral region. For these devices, with relatively high radiative efficiency, approaching 2%, the acquisition time for an image cube depends on the integration time and wavelength steps. With 2 nm spectral scanning over the 800–1000 nm spectral region used for obtaining these images, the acquisition time for one image cube containing about 100 images is about 3 min. In addition to a UV/VIS camera, covering the spectral range from 400–1000 nm, a second short-wave infrared camera is also available for measurements in the 1000–1600 nm range but was not utilized for this study. For PL, a 532 nm laser was used. The laser light illuminates the sample uniformly over the entire field of view and, therefore, no sample or beam scanning is required. Raw image cubes were corrected with a dark cube subtraction and calibrated by a method described previously.^9,10 A low-profile, liquid-nitrogen cooled optical cryostat was used for temperature-dependent measurements.

We first consider hyperspectral PL image cubes taken with a 20× microscope objective, containing one of the numerous dark areas that have been observed on most devices.¹¹ Figure 2 shows frames taken from calibrated image cubes measured near room-temperature (300 K) and at a low-temperature (90 K). The first image at each temperature corresponds to the peak of the spatially averaged PL signal and shows a dark area. At longer wavelengths, a small region in the center of the dark area (circled) is observed to have a strong luminescence, as shown in the second image at each temperature. Finally, a map of the PL peak luminescence wavelength at both temperatures shows that within this region, the dominant luminescence signal is at a noticeably lower energy (longer wavelength) than in the surrounding regions in the device. This has been observed near the center of most dark spots on such devices and indicates a locally high concentration of a sub-bandgap radiative defect.

Figure 3 shows the PL spectra extracted from the full temperature range of hyperspectral image cubes, both from the local defect discussed above and from the clear area to the left of the metal finger (dark vertical line) visible in Fig. 2. In the clear area, it is apparent that several peaks are present. At each temperature, we can decompose the measured spectrum into band-to-band (BB) and defect-related luminescence. These observations are different than our previous report of a local radiative sub-bandgap transition around pinhole sites in rear junction GaAs solar cells.⁵ In those devices, the BB and the local sub-bandgap peaks remained distinct from each other throughout the temperature range of the study. Here, four distinct defect transitions are intertwined with the BB peak and do not remain independent throughout the temperature range as we will show. We use a model for the band-to-band emission developed by Katahara and Hillhouse¹² that takes into account the material bandgap, absorption coefficients, a disorder that leads to tail states in the bandgap, temperature, and the quasi-Fermi level splitting. Figure 4 includes the bandgap used to fit the temperature-dependent PL spectra with this model. This can be well fit by Varshni’s model of bandgap evolution with temperature,¹³ 

in which E₀ is the bandgap at 0 K and α and β are fit parameters, with E₀ = 1.49 eV, α = 8.07 × 10⁻⁴ eV/K, and β = 407 K fitting these data. These values are similar to those reported previously for GaAs¹³ but, as expected with the inclusion of In, produce a slightly smaller bandgap.

We use a Gaussian function to fit each of the four distinct defect peaks that appear over different temperature ranges. Viewing spectra at different temperatures in both linear and logarithmic scales makes these peaks clear, as shown in Fig. 3(b). Defect peaks are labeled 1 through 4 in order of increasing energy. At room temperature, peaks 1 and 2 are present at 1.32 and 1.36 eV, respectively. Peak 1 has a very low peak luminescence and is not detectable at lower temperatures. Peak 2 is present throughout the temperature range, but the intensity relative to the band-to-band and total luminescence decreases. A third defect peak appears at 280 K and increases in intensity over the next few temperature steps. By 170 K, a fourth defect peak is visible, which increases in intensity strongly as the temperature drops further. We will note that this peak may be present in the 190–210 K range but close enough in location to be indistinguishable from peak 3. These peak locations are plotted in Fig. 4 to enable a more obvious tracking across temperatures.

The change in the intensity of these defect peaks compared to the band-to-band luminescence confirms that they are different in nature. Based on our current suite of absolute PL measurements, we do not believe that these near band-edge states are responsible for inducing any additional non-radiative recombination (Shockley–Read–Hall) in our devices beyond what is typically present in GaAs devices. A separate calculation of the external radiative efficiency (ERE) in our devices shows values of ≈1.6%, consistent with GaAs solar cells of similar PCE (around 24%). Previous work on InGaAs, including dilute alloys, also shows the presence of near band-edge defect peaks, which can arise from the structural defects that may form due to lattice mismatch or thermal stresses.^14–16 

The integrated photon flux (sometimes referred to as the intensity throughout this paper) from the band-to-band luminescence and peak 4 is shown in Fig. 5. The drop in band-to-band luminescence intensity below 170 K can be attributed to the redistribution of carriers to defect levels associated with the sharply increasing peak 4.¹⁷ The interdependence of peak intensity limits the ability to separately analyze the temperature-dependent integrated intensity of each peak (as we do below with the local defect peak). It also suggests a multi-center mode of PL quenching for peak 4.¹⁷ Further understanding of the origin of these near-edge defects would likely require additional non-luminescence-based experiments combined with impurity band density functional theory calculations and is beyond the scope of this work.

The PL measurements in the local defect region are fit with a Gaussian curve [Fig. 3(a)], which replicates the peak and high-energy side well. The low-energy tail decays more slowly than the Gaussian curve. Therefore, when calculating the integrated intensity of this peak (Fig. 5), we use the measured low-energy tail and the high-energy half of the Gaussian fit. The short, wide high energy shoulder seen in Fig. 3 is what can be detected of the clear-area peaks in the sampled local defect region and is otherwise ignored in this analysis.

Measurements performed by varying the excitation intensity at both room temperature and 77 K show no systematic change in peak energy position, indicating that a free-to-bound type transition is likely responsible for this luminescence line.¹⁸ Additionally, a previous identification of donor–acceptor-pair transitions in these material systems confined such interactions to temperatures lower than what we measure here.¹⁹ 

The activation energy of a defect can be accessed by the peak position or by the evolution of the integrated intensity with temperature. The difference between the local defect PL peak energy and the bandgap extracted from the band-to-band luminescence varies from 37 meV at 77 K to 49 meV at 230 K and above (Fig. 4). This variation may be due to changes in the distribution of carriers within the conduction band, as we will later see that non-uniformity of the alloy fraction is indicated in this sample.

The defect activation energy also influences the temperature dependence of the integrated luminescence, I_PL, as in the model developed by Levcenko et al., to account for the multiple processes that can occur involving a defect level,

where A and α are constants and N is the ratio of the density of states in the valance band to carrier density,²⁰ also assumed constant here.

A least square fit of the integrated intensity of the local defect peak to Eq. (2) is shown in Fig. 5 and suggests an activation energy of 31.2 ± 6.7 meV, although a slightly wider range of activation energies can provide good visual fits as well.

To identify this defect, we take advantage of the literature on radiative defects in GaAs, which is extensive compared to that on dilute InGaAs. Given the low In alloy fraction in these devices, and a bandgap much closer to that of GaAs than to that of In_0.5Ga_0.5As, we can apply these data to these samples with the proper considerations. For effective comparisons to the literature, and to align with the fit of temperature-dependent integrated intensity, we consider the peak position at 77 K and how it relates to the band edge. It has been noted that the position of shallow defects tends to be fixed relative to the nearer band edge, while those of deep defects are independent of the location of the band edges.²¹ At 77 K, this transition is located at 1.444 eV and the bandgap was found to be 1.481 eV. The bandgap of GaAs at this temperature is 1.51 eV,¹³ a shift of 37 meV. We, therefore, look for a defect in GaAs that produces a radiative transition with a peak intensity between 1.44 and 1.48 eV.

Numerous studies have identified a radiative transition in GaAs in this range,^22–26 attributed to native defects: a gallium antisite (Ga_As) or gallium vacancy (V_Ga) in n-type materials and the arsenic vacancy (V_As) in p-type materials. Given the defect peak energy just below the bandgap energy of InGaAs, this transition likely originated in the InGaAs p-type base as opposed to the InGaP n-type emitter. Additionally, fitting the integrated luminescence as a function of excitation energy I_ex at a fixed temperature to the relation

results in an exponent a ≈ 1.48. This dependence has previously been found to be quadratic in an n-type material featuring Ga_As defects and linear in a p-type material featuring V_As defects, making our results more consistent with the latter.²² 

Hyperspectral imaging, therefore, allows us to identify the defect causing dark spots observed in a luminescence image. Doing so with separate luminescence imaging and spectral measurements would be difficult at best, given the small area that contains a high concentration of this defect compared to the device and dark spot sizes.

The observation of this V_As defect in the p-type base is different than our previous report of a gallium antisite defect in an n-type GaAs emitter layer formed around pinholes that disrupt the lattice structure.⁵ Additionally, unlike the previous report where no extensive dark regions were observed around the local sub-bandgap defects, here we show very large dark areas, up to hundreds of μm in size, around these local defects. Since the local voltage at the sub-bandgap defect sites is lower than the surrounding area,¹¹ a local draining of carriers takes place in the vicinity of these sites, leading to these suppressed luminescence regions. Paired with the knowledge of the fabrication processes of these devices, including details and history of the tools, this information can be used to improve the growth processes and reduce or eliminate the resulting dark PL spots, which reduce the cell’s overall external radiative efficiency, resulting in unnecessary photovoltaic conversion efficiency losses.^10,27

Hyperspectral image cubes taken with a 20 mm field of view (under EL mode) at room temperature are shown in Fig. 6. Figure 6(a) shows the total luminescence, similar to what would be obtained in a conventional EL imaging measurement. Figure 6(b) shows the luminescence at 912 nm, on the low-energy side of the 890 nm main peak. At this energy, a grid-like pattern can be seen that is much less obvious (and in some devices, not noticeable) in the total luminescence image, or closer to the peak energy. Previous studies have used spectral PL measurements from multiple locations in a sample to analyze uniformity,^16,19 but such methods are not guaranteed to catch all variations and would require a prohibitively large number of measurements to reveal a pattern such as the one seen here. Viewing the detailed pattern [Fig. 6(b)] assists in the identification of the source of variation.

Using a moderate, 5×, magnification image [Fig. 7(a)], we select an area from a “dark” and “bright” region in the sub-peak energy range to analyze the cause of this variation. First, we note that this variation, indeed, originated inside the base InGaAs layer (and not a superposition of emissions from any other layer) because the emission characteristics between spots A and B are identical except for the slight redshift between them as shown in Fig. 8(a). Second, no other layer as identified in the device structure [Fig. 1(c)] has a luminescence response close to the 1.38 eV peak shown here. To assist in the selection of these spots, we map the intensity at 912 nm compared to that at 890 nm in Fig. 7(b), to remove the influence of the metal grid on local luminescence levels. The spectra from each location are shown in Fig. 8, along with the decomposed peaks. We find that the brighter (higher intensity at 912 nm compared to 890 nm) areas correspond to a slight shift to lower bandgap energies, which becomes more pronounced at lower temperatures. The defect 4 peak also exhibits such a shift. Mapping the peak luminescence wavelength at 77 K [Fig. 8(b)] makes it clear that this shift in bandgap and peak location does, indeed, follow the grid-like pattern.

One explanation for the shift in bandgap is the existence of a variation in the indium alloy fraction. However, another high-resolution materials technique such as high-resolution x-ray diffraction may be needed to confirm this hypothesis. We also see variations in the intensity of local defect peaks. The bandgap shift and uneven influence on defect peaks indicate that this pattern is not due to a rear optical structure. Such a feature could potentially unevenly reduce photon loss through the rear surface. This would result in a better photon recycling, but the luminescence we detect would still primarily come from transitions occurring near the front surface (as discussed in Sec. III A), obscuring any spectral dependence of the rear surface reflectance. Instead, this pattern may be related to a fabrication process, for example a shape of a heating element, or to an electronic structural design, such as through-holes for contacts that can affect local concentrations of diffused species.

With luminescence hyperspectral imaging, we can easily view the variation in radiative recombination across the device, but we can also consider if there is variation in the concentration or characteristics of non-radiative centers. In Fig. 9, we show the Arrhenius plot of the integrated intensity of the total luminescence signal (band-to-band and all defect-related peaks) for the two locations. Fitting these curves to the equation

in which C₁ and C₂ are constants associated with activation energies E_A1 and E_A2, allows us to characterize the dominant non-radiative centers in each location.^15,28 The luminescence in spot A shows a good fit with a single non-radiative center with an activation energy of 122 meV, while two centers provide a better fit in spot B. Assuming that one of the centers is the same as that found in spot A, we find a second center with E_A = 50 meV. Additionally, the constant C₁ is higher in spot B, indicating a larger influence on quenching, likely from a higher concentration of these centers. All fit parameters are shown in Fig. 9.

The increased non-radiative recombination at room temperature in the “brighter” areas will only become more significant at the higher temperatures that devices tend to operate at in real-world conditions. This increased non-radiative recombination in these areas will, of course, reduce the overall performance of the solar cell, and using hyperspectral imaging to assist in diagnosing the cause provides an opportunity for improved efficiencies.

We have demonstrated the use of hyperspectral luminescence imaging on dilute-alloy InGaAs solar cells to identify and investigate dark spots and the locally high concentration of radiative defects at their center, as well as systemic variations of the luminescence pattern across the device. We were also able to identify the specific defect at the center of the dark spots. Thus, hyperspectral imaging allows for a more thorough understanding of features seen in a standard luminescence image, along with the identification of spatial features that would not be seen without the full spectral resolution. The identification of specific defects causing increased recombination, and spatial patterns that may correlate with specific fabrication processes, opens the door to adjustments to design and fabrication that can result in improvements to device performance.

The authors would like to thank E. Greco and R. Campesato of Cesi S.p.A for device structure information and discussions and Kenneth Schmieder and Margaret Stevens of the U.S. Naval Research Laboratory for helpful discussions. The authors also acknowledge the continuous and timely support provided by the technical staff at Photon etc. to NIST researchers for the optimization and maintenance of the hyperspectral imaging instrument.

NIST disclaimer: Certain commercial equipment, instruments, software, or materials are identified in this article to specify the experimental procedure adequately. Such identification is not intended to imply recommendation or endorsement by the National Institute of Standards and Technology nor is it intended to imply that the materials or equipment identified is necessarily the best available for the purpose.

The authors have no conflicts to disclose.

Brianna Conrad: Conceptualization (equal); Data curation (lead); Formal analysis (lead); Methodology (equal); Validation (lead); Writing – original draft (lead). Behrang H. Hamadani: Conceptualization (equal); Project administration (lead); Resources (lead); Supervision (lead); Writing – review & editing (equal).

The data that support the findings of this study are available from the corresponding author upon reasonable request.