Making Record-efficiency SnS Solar Cells by Thermal Evaporation and Atomic Layer Deposition

Substrate selection cleaning

Oxidized Si wafers are used as substrates. The substrates are the mechanical support for the resulting solar cells, and their electrical properties are not important. Si wafers are preferred to glass because commercially purchased Si wafers are typically cleaner than commercially purchased glass wafers, and this saves time in substrate cleaning. Si substrates also have higher thermal conductivity than glass, which leads to more even heating during growth and annealing. With commercially purchased glass wafers it was found that it was necessary to clean the substrates with detergent, including a manual rub with gloved fingers followed by a warm ultrasonic bath, in order to remove all visible traces of contamination, and even then substrate cleanliness could not be guaranteed. It was experimentally confirmed that the choice of glass or Si substrates has no impact on device performance. However, this comparison was performed with devices were in the 2% – 3% efficient range, and repeated comparisons will be worthwhile as the baseline efficiency improves.

Typically 10 or more substrates are cleaned at a time using a custom designed quartz wafer carrier. This yields more reproducible results than handling wafers individually with tweezers.

Mo sputtering

It was decided to deposit the Mo back contact (the cathode) in-house rather than purchase Mo-coated substrates after disappointing results were obtained with the substrates available from several different vendors. With purchased substrates (sputtered Mo on glass) problems were encountered with cleanliness, delamination, or both. The Mo film is deposited by DC magnetron sputtering in two layers, one for adhesion to the Si / SiO₂ substrate and one for high conductivity, according to the results published by Scofield et al.¹³

Typical Mo films are 720 nm thick with a sheet resistance of approximately 1 Ω/sq. The sheet resistance is measured on a sacrificial substrate using a four point probe system after each Mo growth run. In addition, the adhesion to the substrate is tested using a scalpel blade. Well-adhered films can withstand scratching by hand with a scalpel at moderate pressure without de-laminating. Poorly adhered films will de-laminate with only slight pressure. It was observed that a short substrate plasma cleaning step in the sputtering chamber before Mo growth is important to obtain good adhesion. Typical parameters for this cleaning step are 20 mTorr Ar, 20 W RF, and 60 sec.

SnS growth by thermal evaporation

SnS is evaporated from source powder using a commercially purchased low temperature effusion cell with a pyrolytic boron nitride crucible with volume 32 cm³. When the source is first loaded it is a fine powder and dark grey in color. Typically 3 4 g of powder is loaded. With new source powder, the source temperature is approximately 540 °C in order to maintain a growth rate of 1 Å/sec on a substrate heated to 240 °C. With an increasing number of growth runs, the source temperature must be increased to maintain a constant growth rate. When the required source temperature reaches 610 °C the powder is exchanged.

Closely connected to the source temperature is the issue of SnS flakes. When the growth rate is increased to over 10 Å/sec, large SnS flakes can be observed in the growing film. It is unknown whether these originate from the main source, or from secondary sources such as the effusion cell shroud or the source shutter.

One remarkable observation is that when a batch of source powder is exhausted it leaves behind a white, porous residue at the bottom of the crucible. X-ray diffraction confirms that this is SnO₂. The weight of this residue is typically 0.01 g.

SnS growth by ALD

A critical parameter in this process is the pressure of the nitrogen carrier gas that flows to the Sn(amd)₂ precursor. The pressure is maintained near 250 Torr, but may vary a little bit from time to time. Given that the volume ratio of the Sn(amd)₂ precursor bubbler and N₂ gas trap is 5:1, the pressure inside the bubbler is around 50 Torr. If this value becomes too large, the evaporation of Sn precursor will be significantly suppressed. On the other hand, if the pressure inside the bubbler is too small, there would not be a sufficient pressure drop between the bubbler and the reaction furnace (which has a pressure of ~10 Torr) to enable a smooth gas flow. Either of these scenarios will result to inadequate Sn precursor in the ALD reaction. An indication of this problem is that film thickness near the outlet of the reaction furnace is much thinner than that near the inlet. During each deposition, the pressure inside the reactor is monitored to make sure that the system pressures sit in the correct range.

To ensure a uniform temperature distribution within the reaction zone, the inlet and outlet of the hot-wall deposition tube are wrapped with heating tapes. Under the heating tapes, a pair of thermocouples are inserted to measure the temperature. A non-uniform temperature distribution inside the reaction zone will result to different SnS film morphologies at different regions. At higher deposition temperature, films tend to be rougher and have a lighter color. At low temperatures below 200 °C, films have a higher reflectivity, as can be examined by eye.

SnS annealing in an H₂S tube furnace

The purpose of the annealing step is to optimize the morphology, crystallinity, and electrical properties of the SnS films. For the TE-grown solar cells, the annealing step is performed in a dedicated tube furnace. This 2” diameter quartz tube furnace is capable of flowing mixtures of 4% H₂S (balance N₂), 4% H₂ (balance N₂), pure N₂ and pure Ar. Temperatures are controlled by external Nichrome heating elements and monitored using a quartz-encased thermocouple located in the hot zone. Gas is evacuated using an oil pump filled with inert oil. Seals are made using H₂S resistant elastomers. Typical base pressure is 8 20 mTorr.

The annealing temperature of 400 °C is a balance between secondary grain growth and film re-evaporation. In principle, higher annealing temperature might be beneficial for device performance, and could be achieved without significant film loss by using higher furnace pressure. This is the subject of active investigation.

SnS surface passivation with a native oxide

The purpose of the passivation step is to reduce the density of electronic trap states at junction between the absorber and the buffer layers, and to serve as a diffusion barrier to prevent unwanted mixing of the constituent elements of the absorber and buffer layers.¹⁴ It has been observed that samples processed with this oxidation step have higher V_OC values than samples processed without.

At this time, the oxidation step has not been extensively studied and is probably not optimized. Using X-ray photoelectron spectroscopy results (not shown) it is estimated that the oxide should be less than 1 nm thick for good performance and to avoid current blocking.

Deposition of the transparent conducting oxide (TCO), indium tin oxide (ITO)

Prior to this point care is taken to control the total air exposure of the samples at each step. However, after the buffer layer deposition air exposure is no longer limited and the samples are stored and transported in ambient air.

Prior to this point, all depositions have been “blanket” films, covering the entire substrate. From this point on the depositions are patterned to define individual devices. For both the ITO and metallization steps the depositions are defined using laser-cut metal shadow masks. For the ITO deposition it is very important that the area of the deposited pad is sharply defined by the shadow mask. If the area is not sharply defined, for example due to flexing of the mask in the mask aligner, then the active area of the resulting devices may be significantly larger than the nominal size of 0.25 cm². This can lead to erroneous over-reporting of the current density.

Metallization

The metallization pattern is designed to enable the illumination spot of the quantum efficiency measurement tool to fall completely on the ITO without overlapping with the metal. This constraint results in 2 fingers separated by 1.5 mm, each 7 mm long, and a 1 x 1 mm² contact pad, see Figure 4. This pattern is less than optimal from a device-performance standpoint. A pattern optimized for device performance would use more fingers with smaller spacing.

Ag has been used for TE-grown cells and Ni / Al has been used for ALD-grown cells. This division is historical and is not grounded in an experimental outcome. It is possible that Ni / Al provides superior resistance to corrosion during long storage times. In fact, it has been observed that Ag contacts have a tendency to corrode during extended storage in air (for example, longer than a year).

Device characterization

Finished devices are characterized using current-voltage (“J–V”) measurements collected and external quantum efficiency (EQE) measurements collected with and without white light and voltage bias. To-date solar cells have been measured by contacting individual devices one-at-a-time, using probe micromanipulators in a typical probe station configuration. The solar simulator and EQE systems were physically disconnected, so the sample needed to be re-contacted for each measurement. As a result, it would take approximately 4 – 5 hr to measure J–V and EQE on all 11 devices.

An integrated high-throughput test station that combines J–V and EQE using a single sample chuck and a probing card that contacts all devices simultaneously was recently installed at MIT, see Figure 5. The electrical connections are controlled by a programmable multiplexer, and the motorized X-Y chuck stage is computer controlled. In this way, J–V and EQE measurements can be performed sequentially on all 11 devices in under 1 hr.

For routine device testing an area-defining light aperture is not used. Therefore the active device area may be under-estimated, resulting in over-estimates of the current density. However an aperture is used by certification laboratories, often resulting in lower efficiencies (c.f. Figures 7, 9 and 10). Using an area-defining light aperture is always desirable, but for routine testing it is often neglected due to practical concerns such as to minimize physical contact with the top of the sample. Systematic error due to under-estimating the active device area can be mitigated by using larger device areas. For the work described here, the rather small (0.25 cm²) size was chosen as appropriate for early stage technology development (at earlier stages an even smaller device of 0.03 cm² with no metallization was used). Now that the devices are in the range of 4% efficient and are repeatable, it is worth increasing to a size of 1 cm² or more.

In addition to the standard characterization techniques described above, on occasion samples are also tested using techniques including temperature-dependent J–V, Suns-V_OC, capacitance-voltage profiling, and lock-in thermography. These techniques are used to understand and quantify specific loss mechanisms such as interface recombination and series resistance loss.

The significance of sharing device fabrication protocols

In publications on inorganic thin film PV the results of device tests are never (with experience) accompanied by sufficient experimental details to reproduce the experiment. This situation leads to unnecessary frustration among individual researchers, and hampers the progress of the whole field. This situation also makes it difficult to compare the procedures described herein to those used by other research groups. The techniques described in this manuscript were developed with the aid of numerous conversations with researchers in thin film PV (mainly in the United States), and a lot of trial-and-error. The authors hope that this work helps others avoid unnecessary frustrations, and sets a precedent for details reporting of experimental methods in thin film PV.

Future applications of the described protocol

The protocol described herein is used to establish a “baseline” SnS solar cell. The most important feature of a baseline fabrication protocol is repeatability; the absolute efficiency number is less important. In the authors’ experience, repeatability is the key attribute that will enable ongoing research to improve efficiency, such as by improving surface passivation or reducing bulk defect density. Without a repeatable baseline protocol it is extremely difficult to judge the effects of changes to the fabrication protocol. This is fully described in the section below on hypothesis testing.

Ongoing and future work on SnS solar cells will leverage the baseline protocol described here to optimize the individual fabrication steps with the goal of increasing device efficiency. Of particular interest are the H₂S annealing and the surface passivation steps, since these directly impact the defect density in the bulk of the absorber and at the p–n junction.

Data ensembles enable hypothesis testing

In a field that idolizes champion efficiencies, it is tempting to overlook the ensemble — the >99% of (non-champion) devices — and the useful information it provides. This section motivates ensemble data analysis, and presents facile approaches for visualizing and extracting useful insights from ensemble data. It is presupposed that the reader has a working understanding of experimental statistics (hypothesis testing), and is comfortable calculating a Gaussian distribution, standard deviation, standard error, and 95% confidence intervals for a given data set.

Simply put, ensemble data analysis is the study of variability, which when reduced, enables better hypothesis testing. Variability, colloquially “noise,” obfuscates the “signal” during hypothesis-driven process engineering and scientific research. As noise increases, more experiments are rendered inconclusive. Inconclusive experiments are a waste of time, resources, and optimism. Ensemble data can help reduce variability in two ways:

First, ensemble data reveal process non-uniformities in space and time. This type of variability is systematic (e.g., caused by temperature- or flow-rate-gradients within a given thin-film deposition chamber), yielding a clear spatially resolved pattern of performance variation, exemplified by Figure 12. The spatial or temporal variation embodies the “fingerprint” of the offending process step. In-situ metrology and control samples can help identify and troubleshoot sources of systematic process variability.

Second, ensemble data reveal “luck-of-the-draw” or “common-cause” variance, i.e., statistical variability that affects all ensemble elements equally. This variability is more difficult to troubleshoot, because it is the aggregate result of multiple coupled processing steps. Common-cause variance can best be minimized by stringent process controls and standard operating procedures at each step — admittedly a challenging proposition in a fast-changing and minimally staffed academic environment. Nevertheless, the following exercise illustrates why reducing common-cause variance is essential.

The impact of common-cause variance exemplified: A friendly, fictional scientific competition between Dr. Meticulous and Dr. Messy: Angela and Nessi are researchers from two different laboratories. They are engaged in a friendly scientific competition to test the hypothesis that Process B generates better devices than a well-accepted baseline Process A. Both researchers employ a standard hypothesis-testing procedure, which assumes that common-cause variance results in Gaussian efficiency distributions (more representative statistical distribution functions include log-normal for distributions without outliers, and the more statistically robust log-Cauchy-Lorenzian for distributions with strong outliers).

Angela is known to her colleagues as “Dr. Meticulous.” She endeavors to reduce process variability. Angela does not share her beakerware with others, employs chamber pre-conditioning routines before thin-film deposition, incorporates control samples with every fabrication run, and prefers IC-grade silicon substrates with thermal oxide surfaces instead of the more variable glass surfaces. She produces baseline (Process A) devices with a “true” (i.e., actual) mean efficiency of 10% and a true standard deviation (σ_true) of 0.5%. Device fabrication and measurements are time consuming, and she is only able to fabricate and measure six devices (N = 6) per process.

In another lab, Nessi is hard at work. To her colleagues, Nessi is known as “Dr. Messy.” Her fabrication and metrology tools are located in a shared-use facility. When it’s her turn to use them, she doesn’t take the necessary precautions to ensure low common-cause variance. But because of her sloppiness, her true standard deviation is 1.5% absolute (3× greater than Angela’s). This higher σ_true reflects less well-controlled experimental conditions. However, because Nessi uses higher-purity feedstock materials, her baseline Process A achieves a true mean of 10.6%. Nessi fabricates and measures N = 6 devices per process.

Let’s suppose that Process B improves “true” device efficiency by 10% relative (i.e., 10% improves to 11%; 10.6% improves to 11.7%). Both Angela and Nessi apply the central limit theorem to the N = 6 devices they each fabricate, as shown in Figure 13. Note that the “true” distributions (Figures 13A, D, G) are hidden to the researchers; they only observe their experimental data (Figures 13B, E, H) and the resulting Gaussian fits, standard errors, and confidence intervals (Figures 1C, F, I).

On one hand, Angela’s tighter process control (lower variability, smaller σ_true) allows her to reject the null hypothesis, concluding with >95% confidence that Process B is superior to Process A (Figure 13C). On the other hand, Nessi, who has higher σ_true, cannot conclude that Process B is better than Process A (Figure 13F) with N = 6. Even though Nessi is lucky to make two devices with efficiencies that are higher than any of Angela’s (Figures 13B, E), Angela is winning the race to publish the scientific paper that will revolutionize the way her field thinks about Process B.

Nessi realizes that she needs to increase her confidence intervals, which requires her to reduce her standard measurement error (SE), i.e.,

.    [1]

Nessi can pursue one of two approaches to match Angela’s 3× smaller SE: Nessi can reduce her true variability (σ_o) by a factor of 3, or increase N by a factor of 3² = 9. Nessi gains access to a high-throughput device-measurement apparatus, increasing N by 9×. This improvement offsets her 3× greater process variability, and she is able to detect a statistically significant difference between Processes A and B, rejecting the null hypothesis with >95% certainty (Figure 13I). She is back in the race for publication.

Higher baseline efficiencies increase the odds of successful hypothesis testing: Focusing on the equation for SE (Equation 1), it can be seen how increasing baseline performance increases the odds of successful hypothesis testing. If the aforementioned Process B were tested on a 5%-efficient baseline device instead of 10%, the absolute efficiency improvement would be only 0.5% instead of 1%. Assuming σ_o is unchanged, the minimum number of samples required to reject the null hypothesis increases by 4×. Thus, improving baseline device performance has the same mathematical effect as reducing σ, i.e., a 1-to-1 improvement in confidence intervals.

Final word: Reducing standard error is essential to minimize the risk of inconclusive hypothesis tests. Standard error can be reduced by decreasing common-cause variance, manifest in σ_o, which results in a 1-to-1 reduction in SE. Improving baseline performance has an equivalent effect. One can also increase sample size N, but this will have a weaker impact on SE because of the square root (increasing N also increases the risk of systematic variation).

The importance of applying experimental statistics is widely accepted in biology and physics (c.f. standing statistics committees at all high energy experiments).¹⁵ To improve the quality of data reporting in PV, it is recommended that researchers pay attention to the 99% of devices they fabricate, and adopt hypothesis testing with data ensembles.