The Use of Gravity Filtration of Carbon Nanotubes from Suspension to Produce Films with Low Roughness for Carbon Nanotube / Silicon Heterojunction Solar Device Application

Featured Application: Transparent electrodes are critical in many applications. This work probes an approach to make highly conducting, highly transparent electrodes using carbon nanotubes. Abstract: The morphology of carbon nanotube (CNT) ﬁlms is an important factor in the performance of CNT / silicon (CNT / Si) heterojunction solar devices. Films have generally been prepared via vacuum ﬁltration from aqueous suspensions. Whilst this enables strong ﬁlms to be formed quickly, they are highly disordered on the micron scale, with many charge traps and gaps forming in the ﬁlms. It has been previously established that lowering the ﬁltration speed enables more ordered ﬁlms to be formed. The use of slow gravity ﬁltration to improve the morphology of CNT ﬁlms used in the CNT / Si device is reported here. It was found that slow ﬁltration causes signiﬁcant macroscale inhomogeneity in the CNT ﬁlms, with concentrated thick regions, surrounded by larger thinner areas. By using atomic force microscopy (AFM), scanning electron microscopy (SEM), and polarised Raman spectroscopy, it was determined that there was no large improvement in directional organisation of the CNTs on the microscale. However, the ﬁlms were found to be much smoother on the microscale, with arithmetic and root mean square average height deviation values roughly 3 times lower for slow-ﬁltered ﬁlms compared to fast-ﬁltered ﬁlms. A comparison was performed with CNT-Si solar cells fabricated with both slow and fast-ﬁltered single-walled CNTs (SWCNT) ﬁlms. It was found that slow ﬁltration can produce similar photovoltaic results with thinner ﬁlms. The results demonstrate that ﬁlm morphology, even without improved CNT alignment, can lead to signiﬁcant improvement in device performance in some applications. However, slow ﬁltration did not form ﬁlms of uniform light transmittance over an extended area, causing an increase in the variation in performance between individual devices compared to fast-ﬁltered ﬁlms.


Introduction
Carbon nanotubes were first incorporated into photovoltaic devices as nanoscale fillers in a polymer matrix to provide an electron transport path, or as a transparent conducting electrode for hole collection [1][2][3][4][5]. In these applications, the nanotubes did not participate in the generation of a photocurrent; they merely assisted in the transport of already generated photoelectrons [2][3][4]. In 2007 formation of smoother films. Slow filtration is also less labour-intensive and does not require the use of dangerous chemicals normally used to align CNTs. This would make slow filtration-a preferable method on both a cost and environmental basis. For CNT alignment to occur in slow filtration three conditions must be satisfied.
The surfactant concentration must be below the critical micelle concentration (CMC). The CNT concentration must be below a threshold value. The filtration process must be controlled at a low speed [34]. The alignment of the CNTs in suspension is driven by electrostatic interactions between anisometric particles in a colloidal solution as investigated by Onsager [35]. Song et al. determined that long MWCNTs form a lyotropic liquid crystalline phase [36] and Engel et al. used this property to form rows of aligned CNTs via solvent evaporation [31].
In this work, we produce and compare CNT films filtered quickly under vacuum and filtered slowly under gravity. We applied these films in CNT/Si solar cell devices. The films produced using gravity were smoother but more inhomogeneous over a large scale meaning solar cells using these films yielded more inconsistent results.

Materials and Methods
CNT suspensions were prepared using previously published approaches [37][38][39]. A surfactant solution was formed by adding Triton X-100 to water at a concentration of 1% v/v. This mixture was bath sonicated (≈50 W RMS (root mean squared Watts) (Elmasonic S 30 H, Elma Schmidbauer, Singen, Germany) for 20 min before dry CNTs were added to the mixture at a concentration of 0.2 mg mL −1 .
To suspend the CNTs in the solution, the mixture was bath sonicated in 3-4 20 min periods. The bath water was replaced after each period to keep the bath at around room temperature. Once the suspension was formed, it was centrifuged at 17,500× g for 1 h to remove material that was not well dispersed. The top 2/3 of the centrifuged mixture was collected and the remaining 1/3 was centrifuged for an additional hour to increase yield, with the top 2/3 collected from this second run. The centrifuge used was a Beckmann-Coulter Allegra X-22 (Brea, CA, USA).
Fast-filtered CNT films were produced from suspension using vacuum filtration. A Buchner flask was attached to a water aspiration vacuum via tubing, and a glass frit was placed in the top with a rubber bung providing the seal. A series of two filtration membranes were placed on the frit. The bottom membrane was a mixed cellulose ester (MCE) template membrane (VSWP, Millipore, 0.025 µm pore size, Merck Millipore Ltd., Tullagreen, Carrigtwohill, Co., Cork, Ireland). This membrane was patterned with the shape desired for the final film. For device production a pattern consisting of 4 small holes was used to make 4 identical films. From these films one would be attached to glass and used for light transmittance and conductivity testing whilst the other three were used in device production.
The second (top) membrane is an MCE membrane (HAWP, Millipore, 0.45 µm pore size, Merck Millipore Ltd., Tullagreen, Carrigtwohill, Co., Cork, Ireland). A solution of approximately 250 mL of ultra-pure water containing an aliquot of CNT suspension was filtered through this set up. The flow rate of the solution was faster through the patterned holes in the lower membrane than through the 0.05 µm pores. Thus, the nanotubes were deposited in the same pattern as the lower template paper. After the initial filtration, an additional filtration of 250 mL of ultra-pure water was performed in order to ensure any CNT or CNT/GO residue was washed from the glassware onto the film. Additionally, this step could assist in washing some surfactant from the CNTs in the film. Slow-filtered films were produced from a similar apparatus only on a small scale and with no patterned second membrane. In addition, there was no vacuum applied and the solution filtered through the membrane solely due to gravity.
CNT films were attached to a variety of substrates (glass, silicon, or gold/silicon solar device substrates). The nanotube film was cut (using a shim hole punch) from the excess MCE membrane and was placed, nanotube side down, onto the desired substrate. The film was then wetted with ultra-pure Appl. Sci. 2020, 10, 6415 4 of 18 water, with any excess water siphoned away via a pipette. The film was moved to the desired location on the substrate and was then dried by applying manual pressure to the wet film with absorbent paper. This pressure has the added advantage of ensuring any air pockets trapped between the film and the substrate were forced out prior to attachment. A piece of Teflon was placed over the film, followed by a piece of glass to form a substrate/CNT film/Teflon/glass sandwich. The Teflon was placed to prevent the MCE/nanotube film attaching to the upper layer of glass rather than the substrate. Once the sandwich was formed it was heated at 80 • C for 15 min. This heating dries the film, leading to strong attachment to the substrate. After heating, the film was left to cool for 2-5 min to prevent the film detaching from the substrate when the MCE membrane was dissolved. To dissolve the MCE, the substrate with attached film was submerged in a bath of acetone for 30 min. Following this, the acetone was replaced, and the substrate was again submerged for 30 min with stirring. The substrate was then washed with fresh acetone and dried under a N 2 gas flow. This process gives strong attachment to the substrate, with manual scratching only succeeding in removing small portions of the film where the scratching took place. For the devices presented in this report, arc-discharge single-walled CNTs (SWCNTs) were used.
The photovoltaic device substrates (see Figure 1) were produced by coating pieces of n-doped (phosphorous) silicon wafer (Active Business Company, Brunnthal, Germany) with a negative photoresist (AZ-1518, MicroChemicals, Ulm, Germany) by spin coating at 3000 rpm for 30 s using a Laurell WS-400B-6NPP/LITE Spin Coater (Laurell Technologies Corp., North Wales, PA, USA). The resist was soft-baked at 100 • C for 50 s and was exposed to UV light through a patterned mask before developing in trimethyl ammonium hydroxide (MicroChemicals, Ulm, Germany) for 30 s. The patterned silicon was coated with 5 nm of Cr and 145 nm of Au before an acetone bath was used to dissolve the remaining resist. The result was an array of device substrates which consisted of a gold top electrode with a circular hole 0.079 cm 2 in area in the middle. This hole is the active area, prior to CNT film attachment a drop of buffered oxide etch (6:1 ratio of 40% NH 4 F and 49% HF) was used to dissolve the 100 nm oxide layer on the surface of the silicon. After CNT film attachment (see schematic in Figure 1b), the reverse side of each device was manually scratched with a diamond-tipped pen (to remove the 100 nm oxide layer), a GaIn eutectic was applied, and the device was attached to a steel plate with the eutectic acting as a contact between the silicon and the steel and as an adhesive. The devices were tested as prepared, after a 2% HF treatment, after allowing a droplet of SOCl 2 to evaporate from the surface, and after another 2% HF treatment.
Appl. Sci. 2020, 10, x FOR PEER REVIEW 4 of 18 second membrane. In addition, there was no vacuum applied and the solution filtered through the membrane solely due to gravity. CNT films were attached to a variety of substrates (glass, silicon, or gold/silicon solar device substrates). The nanotube film was cut (using a shim hole punch) from the excess MCE membrane and was placed, nanotube side down, onto the desired substrate. The film was then wetted with ultrapure water, with any excess water siphoned away via a pipette. The film was moved to the desired location on the substrate and was then dried by applying manual pressure to the wet film with absorbent paper. This pressure has the added advantage of ensuring any air pockets trapped between the film and the substrate were forced out prior to attachment. A piece of Teflon was placed over the film, followed by a piece of glass to form a substrate/CNT film/Teflon/glass sandwich. The Teflon was placed to prevent the MCE/nanotube film attaching to the upper layer of glass rather than the substrate. Once the sandwich was formed it was heated at 80 °C for 15 min. This heating dries the film, leading to strong attachment to the substrate. After heating, the film was left to cool for 2-5 min to prevent the film detaching from the substrate when the MCE membrane was dissolved. To dissolve the MCE, the substrate with attached film was submerged in a bath of acetone for 30 min. Following this, the acetone was replaced, and the substrate was again submerged for 30 min with stirring. The substrate was then washed with fresh acetone and dried under a N2 gas flow. This process gives strong attachment to the substrate, with manual scratching only succeeding in removing small portions of the film where the scratching took place. For the devices presented in this report, arcdischarge single-walled CNTs (SWCNTs) were used.
The photovoltaic device substrates (see Figure 1) were produced by coating pieces of n-doped (phosphorous) silicon wafer (Active Business Company, Brunnthal, Germany) with a negative photoresist (AZ-1518, MicroChemicals, Ulm, Germany) by spin coating at 3000 rpm for 30 s using a Laurell WS-400B-6NPP/LITE Spin Coater (Laurell Technologies Corp., North Wales, PA, USA). The resist was soft-baked at 100 °C for 50 s and was exposed to UV light through a patterned mask before developing in trimethyl ammonium hydroxide (MicroChemicals, Ulm, Germany) for 30 s. The patterned silicon was coated with 5 nm of Cr and 145 nm of Au before an acetone bath was used to dissolve the remaining resist. The result was an array of device substrates which consisted of a gold top electrode with a circular hole 0.079 cm 2 in area in the middle. This hole is the active area, prior to CNT film attachment a drop of buffered oxide etch (6:1 ratio of 40% NH4F and 49% HF) was used to dissolve the 100 nm oxide layer on the surface of the silicon. After CNT film attachment (see schematic in Figure 1b), the reverse side of each device was manually scratched with a diamond-tipped pen (to remove the 100 nm oxide layer), a GaIn eutectic was applied, and the device was attached to a steel plate with the eutectic acting as a contact between the silicon and the steel and as an adhesive. The devices were tested as prepared, after a 2% HF treatment, after allowing a droplet of SOCl2 to evaporate from the surface, and after another 2% HF treatment. The testing was performed by applying contacts to the steel plate and the gold top contact. An applied voltage was varied from 1 V to −1 V and the current measured and converted to a current density measurement based on the active area of the device. This was done in complete darkness to test the diode properties of the device and under illumination by a 150 W Xenon lamp (Newport, The testing was performed by applying contacts to the steel plate and the gold top contact. An applied voltage was varied from 1 V to −1 V and the current measured and converted to a current density measurement based on the active area of the device. This was done in complete darkness to test the diode properties of the device and under illumination by a 150 W Xenon lamp (Newport, Irvine, CA, USA) shone through an AM1.5G filter (Newport, Irvine, CA, USA). The light intensity was kept constant using a silicon reference cell (PV Measurements, NIST, Gaithersburg, MD, USA). The measurements were performed using a Keithley 2400 SourceMeter (Keithley Instruments, Inc., Cleveland, OH, USA) running through a custom program written in LabView.
Transmittance percentage values for CNT films attached to glass was determined by measuring the transmittance spectrum for a given film from 300 to 1100 nm using a Cary 60 UV/Vis spectrophotometer.
The transmittance values at 450 nm and 800 nm averaged. The sheet resistance of these films was determined using a four-point probe (Keithlink, New Taipei City, Taiwan) attached to a Keithlink Gwinstek data acquisition unit. A set of four measurements were taken for three different orientations across the CNT film and the resultant 12 measurements were averaged to give the sheet resistance for the film. For experiments on glass, HCl was used in lieu of HF as HF would have dissolved the glass substrates.
Atomic force microscopy (AFM) was performed in air with a Bruker multimode 8 AFM with a Nanoscope V controller (Bruker Corporation, Billerica, MA, USA) in standard tapping mode. The AFM probes used were silicon HQNSC15/AIBS Mikromasch (Tallinin, Estonia) probes (nominal tip diameter and spring constant is 16 nm and 40 N m −1 respectively). Set-point, scan rate, and gain values were chosen to optimise image quality. The AFM scanner was calibrated in x, y, and z directions using silicon calibration grids (Bruker model numbers PG: 1 µm pitch, 110 nm depth and VGRP: 10 µm pitch, 180 nm depth, Bruker Corporation, Billerica, MA, USA). AFM topography images have been flattened, and thickness and roughness measurements determined using Nanoscope Analysis 1.4 (Bruker Corporation, Billerica, MA, USA).
Scanning electron microscopy (SEM) images were obtained using an Inspect FEI F50 SEM (Hillsboro, OR, USA) with beam spot size and high voltage adjusted to optimise image quality. The image resolution was generally 2048 × 1768 with an imaging speed of 30 µs.
Polarised Raman spectroscopy was performed using a Horiba XploRA (Kyoto, Japan) scientific confocal Raman microscope. The laser used was a Nd:YAG 532 nm laser and the power used at the sample was 1.2 mW. A ×50 magnification objective was used. The data was collected at 25 locations in 5 × 5 grid 2 spectra were collected per location with a 30 s integration time per spectrum. The grating used was 1200 grooves per mm which gives spectral resolution of approximately 3 wavenumbers.

Surfactant and CNT Concentration
It was important to ensure that the sodium dodecyl sulfate (SDS) concentration was below the critical micelle concentration (CMC) to ensure that the SDS properly assists in suspending the SWCNTs [34]. The CMC of SDS is 8.2 mM [40].
It was found that the CNT concentration used in these experiments was 0.2 mg mL −1 .

Flow Rate
A variety of conditions were trialed to obtain the ideal flow rate of 1-2 mL h −1 (0.017-0.033 mL min −1 ). The final process was to perform a brief vacuum filtration with just water to flatten the membrane and then to filter the CNT suspension completely under gravity, with no applied vacuum. An example of the flow rates of arc-discharge single-walled CNTs (SWCNTs) is shown in Figure 2a,b. Over time, the flow rate approaches the ideal parameters.  Figure 2c shows the macroscale inhomogeneity of CNT films filtered under gravity. This represents a common problem observed for all slow-filtered films. As can be seen, the dark regions of higher CNT concentration are smaller in area to the lighter regions. As such, for solar device application and future imaging, the lighter areas were used as more small films could be cut from this area and the transmittance and film conductivity could be more reliably calculated. The films produced by He et al. [34] were visually homogeneous across the whole film. The likely reason for this is that in their research a low, controlled vacuum was applied, whereas, in these experiments, gravity was used to pull the CNT suspension through the membrane. Figure 3 shows the microscale morphology of SWCNT films fabricated with both slow and fast filtration at different volumes. The only change visible with increasing SWCNT suspension volume (and thus SWCNT concentration) is an improvement in surface coverage. It is clear that directional SWCNT alignment was not achieved in any of the samples with the SWCNTs exhibiting random orientations in all of the slow and fast-filtered samples. However, a stark difference is observed in film roughness with slow filtration when compared to fast filtration. Large buckled SWCNT regions are seen in the fast-filtered films whereas these are not present in the slow-filtered films. From this alone, it is likely that slow-filtered films are preferable for use in CNT/Si devices, as smoother films will provide a better contact between the SWCNT film and the silicon substrate. In addition, polarised Raman spectroscopy was performed on slow-filtered films at two different orientations (see Appendix A). This showed no difference in peak intensity when the orientation angle was changed, further confirming the lack of CNT alignment in slow-filtered films.

Film Morphology
As observed from SEM images ( Figure 3) the fast-filtered films appeared much rougher on the microscale than slow-filtered films. The smoother a SWCNT film, the better contact it will make with the silicon surface and will thus lead to an increase in photocurrent production. The roughness of the film was analysed using AFM. Two areas on a film of each filtration type were imaged by AFM and the images are shown in Figure 4. Additionally, the root mean square average height deviation (Rq) and the arithmetic average of the surface height deviations (Ra) were taken from the image and are shown in Table 1, together these are quantitative measurements of the roughness of the film.  Figure 2c shows the macroscale inhomogeneity of CNT films filtered under gravity. This represents a common problem observed for all slow-filtered films. As can be seen, the dark regions of higher CNT concentration are smaller in area to the lighter regions. As such, for solar device application and future imaging, the lighter areas were used as more small films could be cut from this area and the transmittance and film conductivity could be more reliably calculated. The films produced by He et al. [34] were visually homogeneous across the whole film. The likely reason for this is that in their research a low, controlled vacuum was applied, whereas, in these experiments, gravity was used to pull the CNT suspension through the membrane. Figure 3 shows the microscale morphology of SWCNT films fabricated with both slow and fast filtration at different volumes. The only change visible with increasing SWCNT suspension volume (and thus SWCNT concentration) is an improvement in surface coverage. It is clear that directional SWCNT alignment was not achieved in any of the samples with the SWCNTs exhibiting random orientations in all of the slow and fast-filtered samples. However, a stark difference is observed in film roughness with slow filtration when compared to fast filtration. Large buckled SWCNT regions are seen in the fast-filtered films whereas these are not present in the slow-filtered films. From this alone, it is likely that slow-filtered films are preferable for use in CNT/Si devices, as smoother films will provide a better contact between the SWCNT film and the silicon substrate. In addition, polarised Raman spectroscopy was performed on slow-filtered films at two different orientations (see Appendix A). This showed no difference in peak intensity when the orientation angle was changed, further confirming the lack of CNT alignment in slow-filtered films. Appl. Sci. 2020, 10, x FOR PEER REVIEW 7 of 18  As observed from SEM images ( Figure 3) the fast-filtered films appeared much rougher on the microscale than slow-filtered films. The smoother a SWCNT film, the better contact it will make with the silicon surface and will thus lead to an increase in photocurrent production. The roughness of the film was analysed using AFM. Two areas on a film of each filtration type were imaged by AFM and the images are shown in Figure 4. Additionally, the root mean square average height deviation (R q ) and the arithmetic average of the surface height deviations (R a ) were taken from the image and are shown in Table 1, together these are quantitative measurements of the roughness of the film.

Film Morphology
Ra values for fast-filtered film were both 3-4 times higher than the slow-filtered film. Additionally, the Z-Range was also roughly 3 times higher. By observing the AFM images, large ridges can be seen in the fast-filtered film which dwarf the highest of 'mounds' visible in the slow-filtered film. These ridges are a product of the turbulent water flow caused by the vacuum filtration method and are the cause of the higher Z-range, Rq, and Ra values.
Slow and fast-filtered films were produced for use in CNT/Si devices. Both sets of films were produced with the same volume from the same SWCNT suspension. However, as slow filtration is known to produce non-homogeneous films it cannot be assumed that both fast and slow filtration setups will produce films of equal transmittance. The slow-filtered film had a transmittance of 78.5% whilst the fast-filtered film had a much lower value of 51.2%. This is a significant observation as a difference in transmittance of this size would be expected to give a large PCE difference. It is thus apparent that there is a difference in the light transmittance vs. film conductivity for SWCNT films produced via slow filtration. To further demonstrate this difference, the sheet resistance for both sets of films was measured and plotted here as Figure 5.   Both the AFM images themselves and the quantitative roughness data and Z-Range demonstrate clearly that the slow-filtered films are significantly flatter than fast-filtered films. The R q values and R a values for fast-filtered film were both 3-4 times higher than the slow-filtered film. Additionally, the Z-Range was also roughly 3 times higher. By observing the AFM images, large ridges can be seen in the fast-filtered film which dwarf the highest of 'mounds' visible in the slow-filtered film. These ridges are a product of the turbulent water flow caused by the vacuum filtration method and are the cause of the higher Z-range, R q , and R a values.
Slow and fast-filtered films were produced for use in CNT/Si devices. Both sets of films were produced with the same volume from the same SWCNT suspension. However, as slow filtration is known to produce non-homogeneous films it cannot be assumed that both fast and slow filtration setups will produce films of equal transmittance. The slow-filtered film had a transmittance of 78.5% whilst the fast-filtered film had a much lower value of 51.2%. This is a significant observation as a difference in transmittance of this size would be expected to give a large PCE difference. It is thus apparent that there is a difference in the light transmittance vs. film conductivity for SWCNT films produced via slow filtration. To further demonstrate this difference, the sheet resistance for both sets of films was measured and plotted here as Figure 5.

Initial CNT/Si Device Comparison
Slow and fast-filtered films of arc-discharge SWCNTs were produced from 400 µ L aliquots in 10 mL of water and were attached to a CNT/Si device substrate with an active area 0.079 cm 2 . For the fast-filtered films, the film areas were homogeneous on the macro scale and thus any area of the films could be used in CNT/Si devices. As discussed above, slow-filtered films are not homogeneous on the macroscale and thus it is important to keep consistent which areas of a film are chosen for solar device usage. All slow-filtered films have a small area of high CNT density and a larger area with a much lower density, for solar device applications, the high SWCNT density area was never used in order to keep the devices as consistent as possible.
It is apparent that, on average, the devices with fast-filtered films outperform those with slowfiltered films with an average PCE of 7.21% vs. 6.51% although this includes a low performing 2.26% device in the slow-filtered batch (see Figure 6). Without this outlier device the slow-filtered films average rises to 8.64% albeit from only 2 cells. To address the large offset between devices in the slowfiltered samples a further set of 3 devices were fabricated and tested. This data is shown in Appendix B. The additional set of devices did not serve to improve the stability of the average PCE, with one device failing to produce a photocurrent and two other devices of 6.51% and 2.25%. If the three top performing devices from the slow and fast sample sets are averaged the slow-filtered films outperform the fast-filtered films with an average of 7.93% compared to 7.21%. However, the existence of several poorer performing devices in the slow-filtered film sets highlights the inhomogeneity of the films produced. The full sets of device data after final 2% HF etch can be seen in Table 2. It is noticeable that, for most parameters, the slow-filtered devices best performance was superior to or equal to the best performance of the fast-filtered devices. However, for every parameter (except shunt resistance) the variation between individual data points was greater for the devices with slow-filtered films. The high variability between devices produced with slow-filtered SWCNT films is likely due to the macroscale inhomogeneity of the films (Figure 2c). Table 2. Solar device data for fast-filtered and slow-filtered films, bold text is the best-recorded result with the average and standard deviation in brackets.

Initial CNT/Si Device Comparison
Slow and fast-filtered films of arc-discharge SWCNTs were produced from 400 µL aliquots in 10 mL of water and were attached to a CNT/Si device substrate with an active area 0.079 cm 2 . For the fast-filtered films, the film areas were homogeneous on the macro scale and thus any area of the films could be used in CNT/Si devices. As discussed above, slow-filtered films are not homogeneous on the macroscale and thus it is important to keep consistent which areas of a film are chosen for solar device usage. All slow-filtered films have a small area of high CNT density and a larger area with a much lower density, for solar device applications, the high SWCNT density area was never used in order to keep the devices as consistent as possible.
It is apparent that, on average, the devices with fast-filtered films outperform those with slow-filtered films with an average PCE of 7.21% vs. 6.51% although this includes a low performing 2.26% device in the slow-filtered batch (see Figure 6). Without this outlier device the slow-filtered films average rises to 8.64% albeit from only 2 cells. To address the large offset between devices in the slow-filtered samples a further set of 3 devices were fabricated and tested. This data is shown in Appendix B. The additional set of devices did not serve to improve the stability of the average PCE, with one device failing to produce a photocurrent and two other devices of 6.51% and 2.25%. If the three top performing devices from the slow and fast sample sets are averaged the slow-filtered films outperform the fast-filtered films with an average of 7.93% compared to 7.21%. However, the existence of several poorer performing devices in the slow-filtered film sets highlights the inhomogeneity of the films produced. The full sets of device data after final 2% HF etch can be seen in Table 2. It is noticeable that, for most parameters, the slow-filtered devices best performance was superior to or equal to the best performance of the fast-filtered devices. However, for every parameter (except shunt resistance) the variation between individual data points was greater for the devices with slow-filtered films. The high variability between devices produced with slow-filtered SWCNT films is likely due to the macroscale inhomogeneity of the films (Figure 2c). films were again inhomogeneous on the macroscale and the obviously thicker regions were avoided when sections were cut for use in the solar cells. An attempt was made to produce slow and fastfiltered films of roughly equivalent transmittance. Figure 7 shows that both sets of devices have the same maximum PCE of around 5.4%. In this case, however, the slow-filtered films produced a higher average PCE, with a higher average Jsc, and Voc, but a lower average FF.  As per the films used previously, the slow-filtered film segments used for solar device application were still thinner than the fast-filtered films. There was some variance between the films and so two films were used in each case. The fast-filtered films were found to be between 85% and Figure 6. J/V properties for devices with 400 µL aliquot slow and fast-filtered arc-discharge SWCNT films. Where PCE is Percent Current Efficiency, J sc is Short-Circuit Current Density, V oc is Open Circuit Voltage, and FF is the Fill Factor. Table 2. Solar device data for fast-filtered and slow-filtered films, bold text is the best-recorded result with the average and standard deviation in brackets. An important observation from Table 2 is that the best performing device with a slow-filtered film produced a PCE of 9.6%, which is higher than the best performing device with a fast-filtered film of 8.1%. However, the large inhomogeneity in the devices with slow-filtered films led to a significantly lower average of PCE (5.7% vs. 7.2%) and a much higher error value of 3.2% vs. 0.8%. Thus, slow-filtered films have the potential to improve the performance of the CNT/Si device but the inhomogeneity in performance, likely due to the macroscale variant in CNT density, makes the improvement difficult to achieve.

Fast-Filtered Films Slow Gravity Filtered Films
The slow-filtered films used for solar devices let through 30% more light than the fast-filtered films but have twice the sheet resistance. Together, the sheet resistance and transmittance can be combined to produce a ratio of DC electrical conductivity (σ dc ) and optical conductivity (σ OP ). This ratio is a good way to directly compare thin SWCNT films with differing resistance and transmittance properties [41,42]. The ratio is defined relative to the transmittance as shown in Equation (1) and this equation can be rearranged to directly give the σ dc :σ OP ratio with an experimental transmittance at 550 nm and an experimental sheet resistance (Equation (1)), both of which have been achieved for the slow and fast arc-discharge SWCNT films.
where T is transmittance at 550 nm in this case, R s is the sheet resistance, µ 0 (4π × 10 −7 N A −2 ) is the free space permeability and ε 0 (8.854 × 10 −12 C 2 N −1 m −2 ) is the free space permittivity. Equation (1) shows that a high σ dc :σ OP ratio will be produced for a film with low sheet resistance and high light transmittance. Thus, a higher σ dc :σ OP ratio indicates a more optimal film for use in a CNT/Si device [43]. Table 3 shows that the σ dc :σ OP ratio increases with treatment, indicating an improvement in overall film performance. This is expected as the treatment improves the film conductivity, the thionyl chloride doping also removes the VHS transitions visible in the 600-800 nm regions of the spectrum and thus reduces light absorbance (increasing transmittance). Overall, Table 3 also shows that the σ dc :σ OP is higher for the fast-filtered films than for the slow-filtered films. This indicates that the fast-filtered films are superior for CNT/Si device use as their lower sheet resistance outweighs their increased light absorbance. Table 3. σ dc : σ OP ratio calculations for slow and fast films used in solar device production. Note that HCl was used in lieu of HF as HF would have dissolved the glass on which the CNT films were attached.

CNT/Si Devices with Films of Similar Transmittance
Due to the difference between the transmittance values, another experiment was performed with CNT films with similar transmittance of about 95%. Whilst thinner films were used, the slow-filtered films were again inhomogeneous on the macroscale and the obviously thicker regions were avoided when sections were cut for use in the solar cells. An attempt was made to produce slow and fast-filtered films of roughly equivalent transmittance. Figure 7 shows that both sets of devices have the same maximum PCE of around 5.4%. In this case, however, the slow-filtered films produced a higher average PCE, with a higher average J sc , and V oc , but a lower average FF.
As per the films used previously, the slow-filtered film segments used for solar device application were still thinner than the fast-filtered films. There was some variance between the films and so two films were used in each case. The fast-filtered films were found to be between 85% and 95% and the slower filtered films were from 95% to 97%. It is expected that this difference in thickness for the fast-filtered films accounts for the variance in device performance with these films. It is known that, when the transmittance of the CNT films approaches 90% and above that the devices will decline in performance as the ability of super thin films to transfer current is poor as the CNT network becomes patchy as thickness decreases. This explains the high variability in PCE for devices with fast-filtered films (Table 4) but it is surprising to see that the devices with slow-filtered films performed well and with low variation considering that they all had transmittances above 90%. This performance again highlights the importance of the film smoothness. The sheet resistance values for these films were all an order of magnitude higher than that used previously. In addition, the values were much more varied and values on the order of ×10 6 Ω were occasionally reported. These issues are all likely due to how thin the films were, thinner films will conduct poorly compared to thick films and will have more charge gaps and holes throughout the network. The outlying data points were removed when average sheet resistance was calculated. The sheet resistance and transmittance measurements were combined to produce σ dc :σ OP ratios for the slow and fast-filtered films to allow for a better comparison.  As per the films used previously, the slow-filtered film segments used for solar device application were still thinner than the fast-filtered films. There was some variance between the films and so two films were used in each case. The fast-filtered films were found to be between 85% and   Table 5 shows a different result to Table 3, with the slow-filtered films producing higher σ dc :σ OP ratios than the fast-filtered films. Both fast-filtered films produced σ dc :σ OP ratios significantly lower than the thicker fast-filtered films in Table 3 which is expected given the thinness of the films. This indicates that the slow-filtered films were superior thin conductive films to the thin fast-filtered films. On the macroscale, films produced via the slow filtration method were less homogeneous than with standard filtration methods, with small areas of high CNT density surrounded by larger areas of low CNT density. The data in Appendix C clearly shows that over smaller areas the slow-filtered films are uniform. Fast-filtered films, on the other hand, were homogeneous on the macroscale to the naked eye.
For improved reproducibility for the slow filtered films, some degree of stirring during gravity filtration could be used. However, this may also decrease the benefits seen on the microscale for slow filtration as it could disrupt the natural ordering of the CNTs in suspension. Alternatively, the use of a controllable low vacuum system instead of gravity filtration may lead to improved macroscale homogeneity. The low vacuum would ensure that a uniform and controllable filtration rate, which would provide tunable alignment and film thickness. The importance of the filtration rate was demonstrated by He et al. [34].

Conclusions
It was apparent that smoother slow-filtered films (T = 80%) have the potential to produce better performing CNT/Si devices than devices with fast-filtered films (T = 50%). However, inhomogeneity in the results meant that, on average, devices with fast-filtered films performed better. When thinner CNT films were used, slow-filtered films were found to produce CNT/Si devices that performed at a similar level to devices produced with fast-filtered films despite the slow-filtered films being thinner than the fast-filtered films. This indicates that the smooth nature of the films allows for an improved CNT-Si contact and improved charge production and extraction capability despite a thinner film being used. The slow-filtered SWCNT films were found to produce superior σ dc :σ OP ratios at high transmittance (>90%), indicating that when the CNT films are very thin slow-filtration produces a better thin film conductor.
This is an important conclusion as it has been determined from previous research that the performance of the CNT/Si device rapidly drops off at film transmittance above 90% as the sheet resistance increases [44]. Clearly, highly transparent and still highly conducting films are of great value and this work shows that a slow filtration approach has the potential to produce high quality transparent films. With improvements in reproducibility, it is clear that the CNT films could be used as thin film conductors for other applications such as gas sensing, display screens, and OLEDs.

Appendix A
In the work by He et al. [34] polarised Raman spectroscopy was used to quantify the degree of alignment within slow-filtered SWCNT films. By varying the polarisation of the incident laser and the detection polarisation between horizontal and vertical, the intensity of the characteristic SWCNT peaks would vary wildly depending on whether the laser was polarised perpendicular or parallel to the CNT alignment direction. This technique was attempted with one of the slow-filtered SWCNT films. However, the Raman setup used lacked the capability to switch between horizontal and vertical polarisations. Thus, the incident laser polarisation was kept horizontally polarised and the sample rotated to observe any alignment effects. Three spectra were taken by recording non-polarised, horizontally polarised, and vertically polarised Raman scattering. It was observed that the intensity of the spectra changed significantly when repeated scans at the same location were performed. To account for this, 25 spectra were produced and averaged over a 5 × 5 grid with each spectrum produced from two separate scans. An averaged spectrum was produced for the slow-filtered film and the fast-filtered film.
Appl. Sci. 2020, 10, x FOR PEER REVIEW 14 of 18 Figure A1. Raman spectra with different detector polarisation for slow-filtered and fast-filtered films. These spectra show the characteristic D (1340 cm −1 ), G (1600 cm −1 ), and G' (2700 cm −1 ) peaks for SWCNTs [45,46]. Figure A1 shows the Raman spectra of a slow filtered film recorded with the sample at two different angles 45 degrees apart (a 90 degree rotation of the sample may accidently lead to the same angular difference between laser polarisation and sample alignment, thus 45 degree separation was chosen) with different detection polarisations. As well as the characteristic D, G, and G' peaks, smaller peaks are visible at around 1000 cm −1 , 2400 cm −1 , and 2900 cm −1 . The 2400 cm −1 peak may be due to highly oriented pyrolytic graphene, which displays a small peak at 2400 cm −1 [45], whilst the 2900 cm −1 peak may be due to a G + D peak [46]. The presence of these peaks indicates that there are graphene-like structures in the SWCNT film. Glass microscope slides display a peak at a little over 1000 cm −1 when a 532 nm laser is used [47] (as was used here) thus this peak is likely caused by the glass substrates. Glass can also show a small peak at 2400 cm −1 which may also be the cause of the 2400 cm −1 peak in Figure A1. Peak intensities and the ratios between peak heights for the nonpolarised spectra and the polarised spectra of the slow-filtered SWCNT film are tabulated in Table  A1. The peak intensities do not change appreciably with rotation of the sample, indicating there is no preferred alignment direction of the carbon nanotubes in the film.  Figure A1. Raman spectra with different detector polarisation for slow-filtered and fast-filtered films. These spectra show the characteristic D (1340 cm −1 ), G (1600 cm −1 ), and G' (2700 cm −1 ) peaks for SWCNTs [45,46]. Figure A1 shows the Raman spectra of a slow filtered film recorded with the sample at two different angles 45 degrees apart (a 90 degree rotation of the sample may accidently lead to the same angular difference between laser polarisation and sample alignment, thus 45 degree separation was chosen) with different detection polarisations. As well as the characteristic D, G, and G' peaks, smaller peaks are visible at around 1000 cm −1 , 2400 cm −1 , and 2900 cm −1 . The 2400 cm −1 peak may be due to highly oriented pyrolytic graphene, which displays a small peak at 2400 cm −1 [45], whilst the 2900 cm −1 peak may be due to a G + D peak [46]. The presence of these peaks indicates that there are graphene-like structures in the SWCNT film. Glass microscope slides display a peak at a little over 1000 cm −1 when a 532 nm laser is used [47] (as was used here) thus this peak is likely caused by the glass substrates. Glass can also show a small peak at 2400 cm −1 which may also be the cause of the 2400 cm −1 peak in Figure A1. Peak intensities and the ratios between peak heights for the non-polarised spectra and the polarised spectra of the slow-filtered SWCNT film are tabulated in Table A1. The peak intensities do not change appreciably with rotation of the sample, indicating there is no preferred alignment direction of the carbon nanotubes in the film. Appendix B Figure A2. J/V properties for additional devices with slow filtered 400 µ L films additional devices.

Appendix C
The AFM data in Figure 3 can also be used to produce height distribution plots in an attempt to determine the thickness of the films. The width of the resulting distribution can be another way to compare film roughness. Figure A3. Height distribution plots produced from AFM images in Figure 3. The red data represents an estimate of the thickness of the films by measuring the height difference between the lower tail of the curve and the centre. The black data shows the full width half maximum for each distribution. Figure A3 shows that there is a significant difference between both the film thickness (from lowest height to average height) and film roughness (as determined from the full width half maximum measurements) for the fast-filtered films and the slow-filtered films. The roughness data backs up that shown in Table 1 Figure A2. J/V properties for additional devices with slow filtered 400 µL films additional devices.

Appendix C
The AFM data in Figure 3 can also be used to produce height distribution plots in an attempt to determine the thickness of the films. The width of the resulting distribution can be another way to compare film roughness.

Appendix C
The AFM data in Figure 3 can also be used to produce height distribution plots in an attempt to determine the thickness of the films. The width of the resulting distribution can be another way to compare film roughness. Figure A3. Height distribution plots produced from AFM images in Figure 3. The red data represents an estimate of the thickness of the films by measuring the height difference between the lower tail of the curve and the centre. The black data shows the full width half maximum for each distribution. Figure A3 shows that there is a significant difference between both the film thickness (from lowest height to average height) and film roughness (as determined from the full width half maximum measurements) for the fast-filtered films and the slow-filtered films. The roughness data backs up that shown in Table 1 Figure A3. Height distribution plots produced from AFM images in Figure 3. The red data represents an estimate of the thickness of the films by measuring the height difference between the lower tail of the curve and the centre. The black data shows the full width half maximum for each distribution. Figure A3 shows that there is a significant difference between both the film thickness (from lowest height to average height) and film roughness (as determined from the full width half maximum measurements) for the fast-filtered films and the slow-filtered films. The roughness data backs up that shown in Table 1 further confirming that slow-filtered CNT films are much smoother than fast-filtered films. The film thickness data shows that the fast-filtered films are roughly three times as thick as the slow-filtered films. It was expected that the fast-filtered films would be thicker than the slow-filtered films as the samples were cut from homogeneous films, whereas the slow-filtered film samples were cut from the thinner areas of the inhomogeneous slow-filtered films (Figure 2c).