Impacts of the Hole Transport Layer Deposition Process on Buried Interfaces in Perovskite Solar Cells

The perovskite solar cell (PSC) is a next-generation photovoltaic device. Here, Wang et al. identify the 3D structures of two PSCs deposited by different methods. This study suggests that the deposition methods not only inﬂuence the ﬁlm’s own properties but also those of its connected layers. SUMMARY Perovskite solar cells (PSCs) are one of the emerging solar cell technol-ogies with high conversion efﬁciency. Several deposition methods had been applied for preparing their hole transport layer (HTL). However, there are few direct evidences to demonstrate whether HTL and its interfaces in PSCs have been inﬂuenced by the deposition methods. In this study, the 3D morphology of PSCs has been reconstructed by focused ion beam-scanning electron microscopy from the PSCs in which HTLs are deposited by different methods. The compositional distribution of HTLs is unveiled as well. All these associated layers and interfaces display obvious morphological and compositional differences that are attributed to the HTL components’ solubility differences in the precursor solvent. Our investigation demonstrates the PSCs that HTL fabricated by dynamic spin-coating method have higher efﬁciency, better ﬁlm uniformity, and less interfacial roughness than the static spin-coating-based devices.

As a p-i-n or n-i-p device, a PSC normally has three functional layers: perovskite absorption layer; electron transport layer (ETL); and hole transport layer (HTL). The qualities of these layers have strong impact on the device performance. Higher phase purity, greater density, and increased film uniformity can yield lower series resistance, higher shunt resistance, less current density-voltage (J-V) hysteresis, higher efficiency, and better long-term device stability. 18 Importantly, the above techniques normally supply 2D information for the device crosssection (SEM, FIB, and (S)TEM), 2D information of the top surface (SPM and SEM), or a 3D projected chemical composition in 2D using X-ray-based techniques (EDX and XRF). The 3D morphological information of the buried layers and interfaces, such as metal/ HTL, perovskite/ETL, or perovskite/HTL interface, in PSCs at a high resolution remains hidden. Moreover, due to limited field of view at high resolution, one risk is inaccurately extrapolating intrinsic properties of samples from abnormal features. A more important fact is that these interfaces are crucial for the PSC's stability and efficiency. For example, in an n-i-p planer structure PSC, oxygen and water are diffused into the metal/HTL interface, 32,33 moisture-induced device degradation starts from the HTL/perovskite interface, 33 and photon-induced perovskite degradation happens at the ETL/perovskite interface. 34 The volumetric reconstruction of these buried interfaces and layers should illuminate detailed morphological information and its impacts on perovskite-based device performance.
Spin-coating is widely applied for depositing PSC functional layers in small area devices. As shown in Figure 1A, there are two different spin-coating methods, static spin-coating (or ''static dispense'') and dynamic spin-coating (or ''dynamic dispense''). In a static spincoating, the solution is loaded and fully wetted on the substrate before spinning process starts. After that, the solution and substrate are spun together to obtain the uniform film. In a dynamic spin-coating, the substrate is spinning at a desired speed. Solution is then quickly loaded (<2-3 s) onto the rotating substrate to finish film deposition. For the perovskite films, compared to static spin-coating, several recent studies show that dynamic spin-coating can enhance the device performance. The successive loaded precursor solution during the dynamic spin-coating for perovskite film deposition helped with retaining the perovskite intermediate phase 35,36 and facilitated the cation ions intermixing. 37 As a result, high-quality perovskite films were obtained.
As for the HTL in PSCs, one of the most common materials combinations is N 2 ,N 2 ,N2 0 ,N2 0 ,N7,N7,N7 0 ,N7 0 -octakis(4-methoxyphenyl)-9,9 0 -spirobi[9H-fluorene]-2,2 0 ,7,7 0 -tetramine (Spiro-OMeTAD), 4-tert-butylpyridine (tBP), and lithium bis(trifluoromethanesulfonyl)imide (LiTFSI). In this combination, Spiro-OMeTAD is the main hole transport material, LiTFSI is the Spiro-OMeTAD oxidizer to improve the hole mobility and conductivity, 38 and tBP is the layer's morphological controller, which can homogenize the LiTFSI distribution by forming the tBP-LiTFSI complexes. 33,39 Due to the low solubility of LiTFSI in chlorobenzene (solvent for the HTL), LiTFSI needs to be dissolved in acetonitrile first. The LiTFSI-acetonitrile solution, tBP, and Spiro-OMeTAD are dissolved in chlorobenzene together as the HTL solution. In previous literature, it has been reported that tBP and acetonitrile can dissolve PbI 2 , which is the precursor for perovskite. 33,39,40 As a result, the perovskite layer can be damaged. To allay this concern in an n-i-p PSC, the HTL is usually deposited by dynamic spin-coating. 41 Thus, it is expected that the minimized direct interaction time between tBP and acetonitrile to perovskite should alleviate the perovskite partially dissolved by the organics. However, the field lacks direct evidence to demonstrate this principle.
Due to the limited field of view, high-resolution 2D imaging may be unable to show the difference between the PSCs with an HTL deposited by static spin-coating (s-PSCs) and dynamic spin-coating (d-PSCs). As shown in Figure 1B, the two crosssection SEM images were acquired from the same s-PSC. The left image shows more voids in the perovskite layer than that of the right image. In Figure 1C, the two cross-section SEM images were acquired from the same d-PSC, which also displayed different morphology at perovskite layers. If comparing the left two images in Figures 1B and 1C, there are more perovskite layer voids in the s-PSC. In contrast, if comparing the right two images in Figures 1B and 1C, the opposite conclusion is drawn: there are more perovskite layer voids in the d-PSC. Such samples require 3D characterization to evaluate the difference between s-PSC and d-PSC voids and morphology.
In this research, FIB-SEM 3D reconstruction is applied to PSCs to observe their buried layers and interfaces. The 3D structures for s-PSC and d-PSC are reconstructed and compared at high resolution. It shows that the different HTL deposition methods have obvious impacts on the morphology of the perovskite layer, HTL, gold layer, and their interfaces. Additionally, liquid chromatography-mass spectrometry (LC-MS), angleresolved X-ray photoelectron spectroscopy (AR-XPS), and ultraviolet photoelectron spectroscopy (UPS) are applied to characterize that the HTL spin-coating methods yielded different compositional ratios, interfacial structures, and energy levels of the HTLs and their associated interfaces. We propose a mechanism based on the precursors solubility difference in organic solvent to explain why these two film deposition methods lead to the obvious compositional and morphological differences. Collectively, this study suggests that the researchers need to carefully select the deposition methods to obtain the desired film composition and morphology. For the multi-components' HTL deposition in PSCs, this study demonstrates the dynamic spin-coating method is better than static spin-coating method.

3D Reconstructions of PSCs
In this study, both s-PSCs and d-PSCs are planar structure devices. The device architecture consists of SnO x electron transport layer, CH 3 NH 3 PbI 3 perovskite layer (with extra CH 3 NH 2 as Lewis base additive), Spiro-OMeTAD HTL (with LiTFSI and tBP additives), and 80-nm gold layer. 42 Except for the HTLs, other layers in s-PSCs and d-PSCs have identical fabrication conditions. 20 devices for each condition had been measured (their efficiencies with forward scan and reverse scan were displayed in Figure S1). As shown in Figures 2A and S1 and FIB milling processes used to prepare cross-section images for PSCs. 38,43 This is expected to limit ion-beam-induced sample morphological change. 44 FIB-SEM carries other benefits for PSCs imaging: the entire thickness for a typical PSC is within 3 mm (in cross-section view without counting the conductive substrate), which means that even a small FIB milling current ($pA) is enough to efficiently micro-machine the slices. In our experiments, the time for processing of all the slices and image acquisition was finished within 2 h-this was even shorter than the time for using FIB-SEM to prepare a PSC TEM lamella ($4.5-6 h). 33,38 In contrast, if a device has thick layers (>5 mm), FIB-SEM requires a longer acquisition time and larger

OPEN ACCESS
Cell Reports Physical Science 1, 100103, July 22, 2020 5 milling current ($nA). In our previous research, the technique was applied for lithium metal anode. Due to the large film thickness and beam sensitivity of Li metal, the samples had to be milled at a cryogenic condition in order to minimize the large ion beam current-induced morphological change. 45,46 On the other hand, compared with other 3D reconstruction techniques, FIB-SEM also has obvious advantages: The cross-section image resolution for a FIB-SEM 3D reconstruction is the same as SEM. The resolution for FIB milling direction is 10 s of nanometer, which is limited by the thinning thickness. It means the majority of the morphological information in PSCs can be observed. TEM 3D reconstruction may show even higher resolution; however, due to the low beam stability for PSCs in TEM, the technique needs to be further developed to minimize the damaging effects. 44 The 3D images were reconstructed in Thermo Scientific Avizo Software 2019.1. All functional layers and voids were labeled as shown in Figures S2C and S2D. After stacking all labeled slices together, the device 3D structures were obtained.  2D). Especially for the HTL/perovskite interfaces, the Ra and root mean square (RMS) of s-PSC are almost twice that of d-PSC. This can be attributed to the fact that the dynamic spin-coating minimized the etching effects from tBP and acetonitrile in HTL solution to perovskite layer. During dynamic spincoating, the fast spin-coating revolutions per minute enable the substrate to get rid of excess amount of tBP and acetonitrile immediately; as a result, the relative smooth perovskite film was protected. On the other hand, in s-PSC, the HTL solution needs to wet the perovskite substrate in order to cover the entire surface before spin-coating. During this period of time, the tBP and acetonitrile can partially dissolve the perovskite layer. Additionally, the perovskite porosity demonstrates d-PSC had fewer voids than s-PSC ( Figures 2E and 2F). This could be explained for the same reason-the dynamic spin-coating can prevent the tBP and acetonitrile from etching the perovskite layer. Moreover, compared with the s-PSC, the HTL additive-free s-PSC in Figure S3 demonstrated less roughness at the HTL/perovskite interface. It further indicates the HTL additive etching is the main reason to cause the morphological change in s-PSC, although dynamic spin-coating can minimize this impact.
For comparison, top-view ex situ SEM images were taken ( Figure S4) Figures S4A and S4B, the morphology of a freshly prepared perovskite layer ( Figure S4A) was changed after treatment with chloroform ( Figure S4B). As a result, although in Figures S4C and S4D, the ex situ SEM of the perovskite films with the HTL deposited by dynamic spin-coating showed fewer pinholes than the static spin-coating sample, it cannot represent the actual perovskite film morphology. Chloroform had already changed the perovskite film morphology as an artifact. In contrast, in Figure  On the other hand, voids were observed in the s-PSC HTL although in the d-PSC was not ( Figures 2E and 2F). It indicates that the tBP or acetonitrile etching to perovskite could be one of the factors on the formation of HTL voids. During the etching, bubbles may form in the HTL solution, and this manifests as voids in deposited HTL. As displayed in Figure 3A, previous research proposed two types of ''imperfect'' Spiro-OMeTAD HTL 3D structures. The first can be termed the ''honeycomb model.'' [47][48][49] In this model, HTL has channel-like pinholes across the entire film without twisting.  entire HTL to connect to perovskite without twisting, the channel-like structures in HTL will be observed by FIB-TEM. On the other hand, if the bubble-like structures were observed in HTL, Swiss cheese model could be correct. However, in Figure 3A, there is also a third possibility: if the pinholes in HTL are twisted as coils, the similar morphology should be observed as the Swiss cheese model in the 2D cross-section FIB-TEM. The HTL with twisted pinholes across the HTL can be termed as the ''anthill model.'' With the help of 3D reconstruction, the accurate HTL 3D model is able to be observed. As displayed in Figure 3B, the dynamic spin-coating HTL has a dense and void-free film. Regarding the static spin-coating HTL, only the bubble-like voids exist in the HTL, and the actual 3D structure for the imperfect HTL in PSC agrees with the Swiss cheese model.
The 3D reconstructions for s-PSC and d-PSC demonstrated obvious morphological differences on their functional layers and interfaces. To understand why different HTL spin-coating methods can show huge impacts on the devices morphology, several characterizations were applied to probe the compositional information of the HTLs.

Compositional and Energy Band Differences in HTLs
To understanding why HTL spin-coating methods could influence the morphology of PSCs, several characterizations were applied to compare the differences between the HTLs. First the film thicknesses prepared at different revolutions per minute were measured (deposition time was 20 s; the revolutions per minute ranged from 500 to 5,000). As shown in Figure 4A, almost all static spin-coating HTLs had larger film thicknesses than the dynamic spin-coating samples. The SEM and 3D reconstruction results in Figures 1B, 1C, and 2B also display the same phenomena that the s-PSC has a thicker HTL than d-PSC. It can be attributed to a process where successive liquid droplets during the dynamic spin-coating may re-dissolve the deposited materials on the films.
In the meantime, the HTLs' compositional ratios were characterized by high-performance liquid chromatography-MS (HPLC-MS). The results were summarized in Figure 4B. More detailed information for the HPLC-MS results were shown in Figure S5 and Table S4. As a reference, the solution composition for HTL spin-coating has a 4:2:1 tBP:Spiro-OMeTAD:LiTFSI molar ratio. According to Figure 4B and Table S4, both of the films have higher LiTFSI:tBP and LiTFSI:Spiro-OMeTAD ratios than the precursor solution. The tBP:Spiro-OMeTAD:LiTFSI ratios are 0.75:0.75:1 for static spin-coating HTL and 0.6:0.6:1 for dynamic spin-coating HTL. It means that the LiTFSI was enriched during the spin-coating process. The dynamic spin-coating HTL has even higher LiTFSI ratio than that of the static spin-coating HTL. This trend has been further evidenced by the HTLs energy diagram. As shown in Figures 4C and  4D, the valence band, work function, and band gap for the HTLs deposited by different spin-coating methods were characterized by ultraviolet spectroscopy (UPS) (22.12 eV HeI source) and Tauc-Plot UV-Vis. To rule out the influence of Spiro-OMeTAD oxidation reaction, the films were deposited in Ar environment glove box and transferred to UPS by an N 2 environment glove box. The energy band diagram based on Figures 4C and 4D were displayed in Figure 4E. It shows that the valence band for dynamic spin-coating HTL is more close to Fermi level, which indicates this film has been more p-doped than the static spin-coating HTL. Previous researches reported LiTFSI is the p-dopant in Spiro-OMeTAD-based HTL. 38,50 Without obviously changing the band gap ( Figure 4D)

Article
As discussed before, the tBP:Spiro-OMeTAD:LiTFSI molar ratio in solution is changed from 4:2:1 to 0.75:0.75:1 (static spin-coating HTL) or 0.6:0.6:1 (dynamic spin-coating HTL) after film deposition. Compared to the precursor solution, this ratio change illustrated that the deposited films have materials loss and the loss order is tBP > Spiro-OMeTAD > LiTFSI. Moreover, the dynamic spin-coating sample has more materials loss than the static spin-coating sample. On the other hand, the material's loss order is the same as the HTL components' solubility in chlorobenzene, which is the solvent for HTL solution. In chlorobenzene, as a liquid component, tBP was miscible with the solvent, although Spiro-OMeTAD was soluble. However, LiTFSI has less solubility in chlorobenzene, and its dissolving process in chlorobenzene required acetonitrile and tBP. LiTFSI was dissolved in acetonitrile first (the LiTFSI-acetonitrile solution was shown as visibly blurred liquid droplets in chloroform). After adding tBP, the formation of tBP-LiTFSI complexes can finally help LiTFSI to be homogenized dissolving in chlorobenzene. 33,39 So in general, the HTL components' solubility in chlorobenzene is tBP > Spiro-OMeTAD > LiTFSI. It indicates that the compositional changes in these deposited films may be induced by the materials' solubility difference in chlorobenzene. Furthermore, tBP, as the only volatile liquid component, can decrease even more due to the evaporation during spin-coating.
Based on these observations and concerns, a mechanism was proposed in Figure 5 to demonstrate the compositional difference of the HTLs deposited by different spin-coating methods: in the static spin-coating case, the wetted substrate contains 4:2:1 molar ratio of tBP, Spiro-OMeTAD, and LiTFSI in chlorobenzene. During the spin-coating, the solution liquid droplets, which are rich in tBP and Spiro-OMeTAD, will be swept out of the substrate. Due to the low solubility in chlorobenzene, compared to tBP and Spiro-OMeTAD, less LiTFSI will be swept out of the substrate. As a result, the ratio of LiTFSI is much higher on the deposited film with respect to tBP and Spiro-OMeTAD. As for the dynamic spin-coating HTL, the liquid droplets are continuously loaded onto the substrate in the first several seconds while the substrate is spinning. During the process, the earlier loaded solution droplets have formed as film, which has higher LiTFSI ratio. However, the later loaded solution droplets can re-dissolve the film and swept away the tBP-and Spiro-OMeTAD-rich droplets. Less of LiTFSI in the film will be re-dissolved for its low solubility in chlorobenzene. The re-dissolved tBP and Spiro-OMeTAD will be swept away with even less amount remaining on the film. As a result, the dynamic spin-coating HTL has even higher LiTFSI molar ratio, due to the initial-deposited film are re-dissolved by the Article successive loaded solution. This re-dissolved process also is confirmed by Figure 4A, which indicates that the dynamic spin-coating HTL has less film thickness than static spin-coating HTL.
The proposed mechanism in Figure 5 is able to explain the compositional differences of the HTL, which deposited by different spin-coating methods. Moreover, the roughness and porosity differences for s-PSC and d-PSC on the HTL, perovskite, and HTL/perovskite interface had been attributed to the etching effect from tBP and acetonitrile to perovskite. However, according to Figures 2C and 2D, the exterior/Au and Au/HTL interface for s-PSC and d-PSC also demonstrate different roughness and morphology. Regarding these two interfaces, the s-PSC also has larger surface roughness than d-PSC. So next, the difference for s-PSC and d-PSC's Au layer (exterior/Au interface) and Au/HTL interfaces will be discussed.

Au Layer (Au/Exterior Interface) and Au/HTL Interface in PSCs
According to Figures 6A and 6B, the top-view SEM images indicate that there is almost no obvious morphological difference between the top surfaces of static and dynamic spin-coating HTL. The morphological change happens after Au deposition. As Figures 6C and 6D displayed, the Au on dynamic spin-coating HTL maintains the similar morphology of the HTL, whereas the Au has island-like structures on the static spin-coating HTL. It agrees with the 3D reconstruction results in Figures  2C-2F. The Au layers in Figures 6C and 6D were deposited together. Moreover, the instrument is equipped with a cooling system to keep the substrate at 20 C during the deposition. Therefore, the Au difference for s-PSC and d-PSC should not be attributed to the thermal-induced effect.
To investigate the reason for the Au morphological difference, angle-resolved XPS (AR-XPS) was applied to investigate the Au/HTL interfaces in s-PSC and d-PSC. As shown in Figure 7A, 5 nm of Au was deposited on these HTLs before the characterization to ensure the HTL buried interface could be detected by AR-XPS because the maximum probing depth in XPS is $10 nm. Each sample was tilted at 0 and 60 takeoff angles, respectively. In AR-XPS, the maximum probing depth is 3lcos q, where the q is take-off angle and l is the materials electrons inelastic mean free path. At 0 , all compositional information is the same as normal XPS. For minimizing to collect the elastic scattering electrons from deeper thickness in AR-XPS, the largest tilting angle for the experiments was restricted to 60 , though the instrument (AXIS-Supra) can be tilted to even a higher angle (80 ). 51 When tilting at 60 , the maximum probing depth is $5 nm, which means the compositional information is more close to the Au/HTL interface. The survey spectra (all peaks were labeled) and XPS spectra of N 1s, O 1s, and S 2p were displayed in Figures S6 and S7. It has to note that, though the samples contain LiTFSI, the Li signal cannot be detected due to the low relative sensitive factor of Li 1 s peak, and also, the peak is overlapping with Au 5p.
In Figure 7B, for both samples the F 1s AR-XPS spectra have two components, TFSI À at 688.7 eV and F À at 685.2 eV. It had been reported that the F À might from the X-ray-induced TFSI À decomposition. 47 A controlled XPS experiment was conducted for the HTL sample, which measured over 10 times. As displayed in Figure S8, the Spiro-OMeTAD-based HTL has a ''fluoride testing window'': within 5 times measurements in our condition, the F À amount was not changed. After that, a linear positive correlation was observed between the F À amount and the XPS measurement time that indicates the X-ray-induced TFSI À decomposition happened. It means that the F À is the intrinsic component in the HTL if the measurements controlled within 5 times, and the all AR-XPS results in this study meet the requirements. Under this concern, within the F testing window, the formation of F À in HTL can be attributed to the partial decomposition of LiTFSI during the HTL spin-coating. As shown in Figure 7B, both HTLs have more F À components when close to the Au/HTL interface because the 60 AR-XPS results have higher F À ratio. From the bulk to Au/HTL interface, the HTL that deposited by dynamic spin-coating has a steeper F À change (from bulk to surface, 4.6%-25.4%) than static spin-coating (from bulk to surface, 12.7%-33.6%). On the contrary, in Figure 7C, the C 1s spectra display the opposite trends. In the spectra, the ratio of sp 2 carbon (C=C) at 283.8 eV to sp 3 carbon (C-H and C-C) at 284.6 eV can be applied for evaluating the Spiro-OMeTAD oxidation. It is because the oxidized Spiro-OMeTAD has fewer p electrons than the pristine one. The higher of (C-H and C-C):(C=C) ratio means the more oxidized Spiro-OMeTAD exists. In Figure 7C, from the bulk to surface, the (C-H and C-C):(C=C) ratio for static spin-coating HTL changed from 0.29 to 1.44, although the values change for dynamic spin-coating HTL is from 0.89 to 1.36. It means the oxidized and pristine Spiro-OMeTAD is more uniform distributed in the dynamic spin-coating HTL.
The AR-XPS C1s and F 1s spectra indicate that the dynamic spin-coating HTL has a more steep distribution of F components and uniform distribution of the oxidized and pristine Spiro-OMeTAD than that of static spin-coating HTL. This could be attributed to that the dynamic spin-coating sample has the components re-dissolve process and LiTFSI enrichment during the spin-coating. As a result, all species gradient had been changed and remixed, although some of the TFSI À enriched on the surface and decomposed as F À .
To further observe the Au/HTL interface for the HTLs, the components' distribution was reconstructed by comparing the value ln (I 60 /I 0 ) for each component in the AR-XPS. 52  Article Figure 7D. In the relative depth plot results, the static spin-coating sample has more widespread data points, which indicates the components have gradient distribution. Moreover, the organic species, such as CN, C-H, and C-C in static spin-coating sample are on top of F À ; during the gold deposition, the Au vapor might embed with these components and result in a higher roughness Au layer. In dynamic spincoating HTL, F À is located at the top surface of the HTL. The inorganic component might be able to prevent the Au diffused into the HTL during the e-beam evaporation. As a result, the morphology of Au was the same as the HTL without deposition as the 3D reconstruction results ( Figure 2) and SEM images ( Figure 6) displayed. In general, the dynamic spin-coating changed the components' distribution in HTL, which enables the HTL to maintain its roughness during the Au deposition.
In summary, FIB-SEM 3D reconstruction was applied for unveiling the buried layers and interfaces in perovskite solar cells at high resolution. With the help of the technique, two types of PSCs were compared into which HTLs had been deposited by static and dynamic spin-coating. Both devices' buried interfaces and layers were able to be imaged by the 3D reconstruction and showed different surface roughness and porosity. The roughness changes on HTL/perovskite interfaces were attributed to the etching effect of the HTL components. In HTL, the 3D reconstruction results suggest the Swiss cheese model could be the right structure for the HTL with defects. The compositional differences for the HTLs prepared by different spin-coating methods were investigated by HPLC-MS. It suggests that the dynamic spin-coating method can enrich LiTFSI in the HTL. A mechanism was proposed to explain why the spin-coating methods can lead to the compositional ratio change in HTL. It indicates that the solubility for each component in the HTL solvent and the re-dissolve during dynamic spin-coating results in the difference. The 3D reconstruction results also displayed the morphological difference on Au/HTL and exterior/Au interfaces between the two spin-coating methods. It was further characterized by the AR-XPS, which demonstrates the different spin-coating methods can change the concentration gradient for each component on the Au/HTL interface. The dynamic spincoating HTL has a ''firm'' interface to resist the morphological change from the Au deposition. Researchers may pay attention to choose ''right'' spin-coating methods once the precursor solution contains multiple components that have different solubility in the spin-coating solvent. These results also suggest 3D reconstruction by FIB-SEM can be a versatile and indispensable technique for perovskite device characterizations, especially for the morphological information of buried layers and interfaces.

EXPERIMENTAL PROCEDURES
Resource Availability Lead Contact Information and requests for resources and materials should be directed to the lead contact, Ying Shirley Meng (shirleymeng@ucsd.edu).

Materials Availability
Materials synthesized in this manuscript can be obtained by request to the lead contact.

Data and Code Availability
The published article includes all datasets generated or analyzed during this study.

Materials
All reagents, unless otherwise specified, were purchased from Sigma-Aldrich. The PSC has an n-i-p planar architecture, with ITO Glass/SnO x /CH 3 NH 3 PbI 3 /doped Spiro-OMeTAD/Au. The ITO (indium-doped tin oxide) glass slides were cut to 5 cm 2 (2 cm by 2.5 cm). The slides were etched by Zn powder with HCl with a 1 cm by 2 cm area to prevent short circuiting. They were cleaned by ultrasonication in 2% Hellmanex solution (in water), deionized water, isopropanol, and acetone, sequentially for 15 min. After drying the substrates and with 10-min oxygen plasma treatment, the slides were ready for materials deposition. The HTL solution contains 17.7 mL LiTFSI-acetonitrile (520 mg/mL) solution, 20 mL tBP, and 0.08 g Spiro-OMeTAD (Merck) in 1 mL chlorobenzene solvent. The solution was prepared in an Ar environment glove box with a <0.1 ppm water level. For static spin-coating HTL, 80 mL of HTL solution was spread on the substrate before starting spin-coating. Then, the films were spin-coated at 4,000 rpm for 20 s. For dynamic spin-coating HTL, 80 mL of HTL solution was loaded within the first 2 s once the substrate started spinning at 4,000 rpm, the entire spin-coating time is 25 s. 80 nm of Au was deposited on the devices by e-beam evaporator.

Perovskite Solar Cells 3D Reconstruction
For 3D reconstruction, the data acquiring is an automatic process in FIB-SEM (FEI Scios DualBeam FIB-SEM). As shown in Figure 2B, in the experiment, a 5 mm by 5 mm Pt protective layer was deposited first in FIB-SEM, with a 500-nm thickness to prevent the sample damaged by ion beam. After that, three sides (front [8 mm 3 2 mm 3 5 mm], left and right [3 mm 3 10 mm 3 5 mm]) around the Pt layer were milled to ensure the acquired region is distinguishable. A fiducial marker with an ''X'' shape (2 mm 3 500 nm 3 500 nm) was milled near the image acquisition region for the software (Auto Slice and View) to identify the proper position. During the image acquisition process, the software can automatically mill, focus, and take image for each slice at the assigned cross-section region sequentially. For each device, 40 slices (30-nm thickness for each slice) were acquired along with the red arrow in Figure 2B. The FIB milling current is 5 pA at 30 kV, and the SEM imaging current is 0.1 nA at 5 kV.
After acquiring all slices of images in FIB-SEM, the results were reconstructed in Thermo Scientific Avizo Software 2019.1. All functional layers and voids in the software were labeled as shown in Figures S1A and S1B. To measure roughness, we measured deviation in Z from the median of a layer's minimum and maximum height in Z for every XY coordinate. The distances were calculated from surfaces calculated from segmentation data acquired through the segmentation editor after a non-local means filter on the raw data. Voids were omitted (by digitally filling them in) for

OPEN ACCESS
Cell Reports Physical Science 1, 100103, July 22, 2020 15 surface roughness calculations. Surface roughness calculations were calculated per interface and per device in Excel.

Characterization
The performances of PSCs were tested with a solar simulator with a 150 W xenon lamp (Solar Light SL07265; equipped with an AM1.5G filter; calibrated with a standard Si solar cell to simulate AM1.5 illumination [100 mW cm À2 ]) and a Keithley 2400 source meter.
SEM images were taken by FEI Scios DualBeam FIB-SEM; the current is 0.1 nA at 5 kV.
Angle-resolved X-ray photoelectron spectroscopy (AR-XPS) and ultraviolet photoelectron spectroscopy (UPS) were performed using a Kratos AXIS Supra with Al Ka anode source (1,486.6 eV) operated at 15 kV and 10 À8 Torr chamber pressure. The data were calibrated with the hydrocarbon C1s peak (284.8 eV) and processed by CasaXPS. In AR-XPS, the samples were coated with 5-nm thickness of Au and were measured at 0 and 60 .
UV-Vis spectra were carried out in an absorption mode on a Lambda 1050 UV-Vis spectrometer.
High-performance liquid chromatography-mass spectrometry was carried out in Thermo LCQdeca; 10 slides of spin-coated HTLs on 5 cm 2 ITO substrates in each condition were dissolved by chloroform to minimize the sample errors.
The porosity and roughness for the 3D reconstruction PSCs were calculated by Avizo.