DOI: 10.1039/C8SE00215K
(Paper)
Sustainable Energy Fuels, 2018, Advance Article

Silvan Suter^{a},
Marco Cantoni^{b},
Yannick K. Gaudy^{a},
Simone Pokrant^{c} and
Sophia Haussener*^{a}
^{a}Laboratory of Renewable Energy Science and Engineering, EPFL, Station 9, 1015 Lausanne, Switzerland. E-mail: sophia.haussener@epfl.ch; Tel: +41 21 693 3878
^{b}Interdisciplinary Centre for Electron Microscopy, EPFL, Station 12, 1015 Lausanne, Switzerland
^{c}School of Engineering, htw saar, Goebenstrasse 40, 66117 Saarbrücken, Germany

Received
9th May 2018
, Accepted 7th September 2018

First published on 10th September 2018

Semiconductors with complex anisotropic morphologies in solar to chemical energy conversion devices enhance light absorption and overcome limiting charge transport in the solid. However, structuring the solid–liquid interface also has implications for concentration distributions and diffusive charge transport in the electrolyte. Quantifying the link between morphology and those multi-physical transport processes remains a challenge. Here we develop a coupled experimental-numerical approach to digitalize the photoelectrodes by high resolution FIB-SEM tomography, quantitatively characterize their morphologies and calculate multi-physical transport processes for the exact geometries. We demonstrate the extraction of the specific surface, shape, orientation and dimension of the building blocks and the multi-scale pore features from the digital model. Local current densities at the solid–liquid interface and ion concentration distributions in the electrolyte have been computed by direct pore-level simulations. We have identified morphology-dependent parameters to link the incident-light-to-charge-transfer-rate-conversion to the material bulk properties. In the case of a structured lanthanum titanium oxynitride photoelectrode (E_{g} = 2.1 eV), with an absorbance of 77%, morphology-induced mass transport performance limitations have been found for low bulk ion concentrations and diffusion coefficients.

Here we propose a combined experimental-numerical approach for the morphological and transport characterization of complex 3D photoelectrodes. We acquired the exact 3D geometry of morphologically complex photoelectrode films destructively and ex situ using FIB-SEM tomography. This tomography provides high resolution (voxel size down to 4 × 4 × 4 nm^{3}), with the ability to measure large sample volumes (hundreds of cubic microns). Precise ion beam milling and advanced automated tracking systems allow for isotropic measurements, providing the same high resolution in the slicing direction as in the imaging plane. The digitalized structure enables quantification of the overall film structure of the photoelectrode, as well as the individual components of the film. Furthermore, this data can then be used in direct pore-level simulations to characterize multi-physical transport in the films and identify transport limitations. In the present study, the methodology is demonstrated on hematite ‘cauliflower’ and structured lanthanum titanium oxynitride photoelectrodes. The outcome of this study facilitates the identification of structural performance limitations and the development of morphology guidelines for performance enhancement.

Two structured photoelectrodes of different materials, deposition procedures, and characteristic feature sizes were analyzed. Sample (i) was a 650 nm thick hematite (α-Fe_{2}O_{3}) photoelectrode consisting of ‘cauliflower’ shaped pillars deposited on FTO by atmospheric pressure chemical vapor deposition (APCVD). The detailed fabrication process of this electrode and its photoelectrochemical behavior were reported by Kay et al.^{7} These authors demonstrated a maximum photocurrent of 2.2 mA cm^{−2} at 1.23 V_{RHE} in AM 1.5 G light of 1000 W m^{−2}. The smallest feature size, a single branch of a pillar, was in the range of tens of nanometers. Sample (ii) was a 7570 nm thick lanthanum titanium oxynitride (LaTiO_{2}N, LTON) photoelectrode that was fabricated by electrophoretic deposition of LTON particles on FTO, followed by a dip coating step resulting in particle necking with TiO_{2}. The detailed fabrication of LTO particles by solid state synthesis, the thermal ammonolysis step to obtain LTON particles, the deposition method, and the photoelectrochemical response were reported by Landsmann et al.^{8} These authors reported a maximum photocurrent of 0.8 μA cm^{−2} at 1.23 V_{RHE} for bare LTON, and improved it to 2.2 mA cm^{−2} at 1.23 V_{RHE} by adding a co-catalyst,^{10} both measured in AM 1.5G light of 1000 W m^{−2}. The LTON particle dimensions varied from hundreds of nanometers to micrometers, and the pores within the particles were in the range of a few nanometers. Fig. 1 shows the 3D rendering of the hematite and LTON photoelectrodes, digitalized based on the post-processed SEM cross-sections.

3.1.1 Volume fraction profile. The volume fraction profile (Fig. S2†) in the hematite film was divided into three regions: a dense base of the ‘cauliflower’ stems (z/t = 0–0.2), an intermediate layer (z/t = 0.2–0.5) with a decreasing density (85–73%), where the pillars are branching out, and a top layer (z/t > 0.5) with a sharp density drop due to the uniform heights of the ‘cauliflowers’. The LTON film showed a much more loosely packed particle network of 30% solid, with a monotone decrease at the top (z/t > 0.4) as the particle density thinned out. The plate-like particles, most standing with a single corner on the glass, led to a low volume fraction just above the FTO glass (z/t = 0.015, the FTO glass roughness had a profile height in the range of 20–120 nm). The volume fraction of the TiO_{2} phase confirmed that necking fills the roughness of the FTO surface. The calculated mean volume fractions in the films were 60.1% and 19.8% for the semiconductor phase in the hematite and LTON sample, respectively.

3.1.2 Surface area. The hematite ‘cauliflower’ structure showed a film roughness factor (surface area per flat xy-plane area) of 12.7. The LTON electrode showed a film roughness factor of 30.3, but more than half of this was accounted for by the nano-pore network within the particles. Without the surface area of the nano-pores, the film roughness factor decreased to 12.5. Our calculations also showed that the necking between the particles decreases the surface roughness only slightly, by a value of 1.3. When we considered all interfaces in contact with the electrolyte active, including the nano-pores connected to the inter-particle space, the film roughness was 18.2. For the LTON sample, some of the FTO substrate remained uncovered and contributed a value of 1.4 to the roughness factor. When evaluating the results as surface area per semiconductor volume (excluding the surface and volume contributions of the FTO), the highest value was found for the hematite sample, 3.382 × 10^{−2} nm^{−1} compared to LTON = 1.958 × 10^{−2} nm^{−1}. The specific surface of the LTON in direct contact with the percolating electrolyte was 1.133 × 10^{−2} nm^{−1}. The specific surface of the LTON without nano-pores was 0.751 × 10^{−2} nm^{−1} (Fig. S3†). This converts to hematite = 6.567 m^{2} g^{−1} and LTON = 3.169 m^{2} g^{−1}, using a bulk density of hematite = 5.15 × 10^{−12} g μm^{−3} and LTON = 6.18 × 10^{−12} g μm^{−3}. Our calculated values agree qualitatively with reported measurement results. The measured value for the surface roughness factor for hematite (measurement method not specified) has been reported to be approximately 24,^{11} and the specific surface area for LTON (BET method) has been reported to be 14.2 m^{2} g^{−1}.^{8} The FIB-SEM spatial resolution is the limiting factor for a more accurate calculation of the surface areas.

3.1.3 Size distribution of the semiconductor phase, pore space, and necking. It is difficult to experimentally measure characteristic feature sizes and to quantify their distribution, especially for PEC electrodes with complex morphologies. The opening size distribution method applied to the semiconducting phase of the segmented dataset is capable of computationally quantifying the shortest diameter of every feature and thus gives an indication of the mean path length for the minority carriers to reach the solid–electrolyte interface. For the hematite electrode, the mean feature diameter is 74 nm [d_{min} = 0 nm, d_{max} = 144 nm] (Fig. 2a). In Fig. 2b, the opening size distribution for the semiconducting phase of the LTON film is shown. The nano-pores in the individual particles were artificially filled in order to quantify the typical LTON particle size. The average characteristic feature size was 441 nm [d_{min} = 0 nm, d_{max} = 800 nm]. This is double the maximum path length a minority carrier needs to travel to reach the outer interface of a particle. However, a minority carrier may transfer to the electrolyte via the connected nano-pore network.

The pore space, filled with electrolyte in the operating PEC electrode (as long as the pore network is percolating and the capillary pressure is high enough for the electrolyte to penetrate into the smallest pores), had a mean diameter between hematite branches of 29 nm [d_{min} = 0 nm, d_{max} = 60 nm]. The mean diameter of the pore space, due to inhomogeneity of film thickness, i.e. between the thinnest and thickest film region, was 103 nm [d_{min} = 60 nm, d_{max} = 185 nm]. The pore space of the LTON film was divided into three different length scales: the nano-pores (nP) within individual LTON particles, having a mean diameter of 30 nm [d_{min} = 0 nm, d_{max} = 128 nm], the meso-pores (mP) between the individual LTON particles, having a mean diameter of 1.17 μm [d_{min} = 128 nm, d_{max} = 2 μm], and the mean characteristic diameter of the pore volume developing between the thickest and thinnest region of the film, having a mean diameter of 3.03 μm [d_{min} = 2 μm, d_{max} = 4 μm].

The opening size distribution of the necking phase (TiO_{2}) in the LTON electrode revealed that the average diameter of the necking was 93 nm [d_{min} = 0 nm, d_{max} = 256 nm]. Together with the results from the volume fraction profile, these small sizes of the necking patches indicate that, apart from the first layer of LTON particles (well connected to the FTO), the charge generated in particles higher up in the film must be transported through multiple TiO_{2} necking patches, and particles before reaching the FTO substrate.

3.2.1 LTON electrode: particle characterization. Non-overlapping ellipsoids were fitted into all the particles. Since the shape of the analyzed LTON particles was dictated by the crystal structure of the precursor oxide LaTi_{2}O_{7} (LTO), the three ellipsoid axes were associated with crystal axes of the LTO and the LTON unit cell. The resulting distribution of ellipsoid diameters along the directions of the three semi-axes were approximated by lognormal distributions (Fig. S8†). The shortest diameter of the ellipsoid, d_{s}, was confined between 100 nm and 500 nm, with a mean of 272 nm (SD = 88 nm). This result confirms and quantifies the narrow range of d_{s}, which is explained by a known cleavage plane, normal to the shortest particle diameter, with weak bonds in the initial LTO particles.^{12} The intermediate, d_{i}, and longest, d_{l}, ellipsoid diameters were spread over a larger range up to 2 μm and 4 μm, respectively (d_{i,mean} = 932 nm, SD_{i} = 427 nm; d_{l,mean} = 1789 nm, SD_{l} = 773 nm). These indicate high aspect ratio particles, quantified in Fig. 3. Overall, a high aspect ratio for d_{s}/d_{i} was found with a mean of 0.35, and a moderate aspect ratio was found for d_{i}/d_{l} with a mean of 0.55. The results were grouped according to the length of d_{i} into 4 classes of 500 nm intervals. The mean values, calculated for each class, showed a significant trend from prolate shapes to oblate shapes as d_{i} increased. The ratio d_{s}/d_{l} was fixed within a narrow range (lognormal distribution: mean = 0.18, SD = 0.09), indicating that the thicker monocrystalline particles can sustain a longer extension along the d_{l} semi-axis. These results agree with the qualitative findings of a TEM study on the particle shape for LTON,^{12} and highlight the additional advantage of strongly improved statistics. This quantitative information on particle shape delivers novel input for the understanding and control of single crystal particle growth modes in LTON.

Based on the ellipsoid fitting, the particle volume distribution was estimated by assuming cubically shaped particles (Fig. S9†). In this case, the volume of the particles can be approximated by the volume of the box surrounding the ellipsoid. The lognormal distribution function had a mean of 0.521 μm^{3}, a standard deviation of 0.544 μm^{3}, and a median of 0.338 μm^{3}.

Quantifying the bounding boxes allows the estimation of the areas of specific surface plane types known to be implicated in photoelectrode performance. Pokrant et al.^{12} correlated specific crystallographic planes with LTON particle faces and found that the biggest surface corresponds to the (100) plane, the medium surface to the (001) plane and the smallest surface to the (010) plane. Following this model, we quantified the ratio of these surface types. In an ideal case with non-touching particles, the (100) plane accounted for up to 70.57% of the overall outer particle surface, the (001) 19.33%, and the (010) 10.10%. If one knows the activity of the surface plane types, for example by ab initio calculations, then the results here are key to estimating the performance potential of LTON particles and electrodes.

Using the eigenvectors of the ellipsoid radii, the orientation of the particles with respect to the FTO substrate was determined (Fig. S10†). Due to their plate-like nature, the orientation of the particles was quantified using the smallest diameter axis (standing normal on the largest particle surface). The smallest diameter vector is perpendicular to the electrode plane for a particle lying flat on the FTO, and an elevation angle of 90° was assigned. The elevation angle was set to 0° for a particle standing on an edge (i.e. the smallest diameter vector is parallel to the FTO substrate). No trend of particle orientation as a function of the z-positon in the film was found. Only a small percentage of particles laid flat or almost flat in the film. For the most part, the normal vectors of the particles were elevated by 10° or 45°. This is a desirable result, as the incoming light is absorbed along the long axes of the particles, the minority carriers have a short way to the electrolyte, and the majority carriers have to pass through fewer TiO_{2} necking patches.

3.2.2 LTON electrode: nano-pore characterization. The pore network within the particles of sample (ii) forms during the powder ammonolysis step from LTO to LTON as compensation for the unit cell volume reduction. Its volume fraction is on the order of 14%.^{12} The inner porosity (volume fraction within the ellipsoid volume, i.e. neglecting particle boundary effects) was calculated for 68 individual particles, and an average value of 10% was obtained with a standard deviation of 0.02. This value agrees well with a calculated inner porosity of 9%, based on a complementing TEM tomography of a single LTON particle (TEM resolution of 7 × 7 × 7 nm^{3}). Our calculations predict that 34.1% of the total nano-pore volume within the particles is connected with the interparticle void space, and is thus filled with electrolyte during PEC operation.

The nano-pore orientations relative to the orientation of the hosting particle were also quantified. The semi-axes of the ellipsoid were used as the reference coordinate system (Fig. 4a). An azimuth and an elevation angle of 0° corresponded to an elongated pore directed along the x-axis, the direction of the longest particle diameter d_{l}. The calculated pore orientations were translated into crystallographic directions based on an orthorhombic LTON unit-cell with unit-cell parameters: a = 5.57137 Å ([100]), b = 7.8790 Å ([010]), and c = 5.60279 ([001]).^{13} Fig. 4b shows the 2D histogram of the relative nano-pore orientations based on all pores in 4 particles with a volume greater than 30 voxels. The volume threshold filtered out small pores for which the algorithm was not able to determine the orientation. The frequency of the elevation angles was normally distributed around 0°. Considering only the elevation angles between 0° and 5°, an overall maximum in the azimuth angle distribution was found at −30° and the corresponding local maximum at 150° was shifted by +180°. A high number of pores was elongated in the [010] direction, which corresponds to the longest semi-axis of the particle. However, the 2D histogram revealed that the pore orientations along the [011] direction belong to the most frequent elongation direction. With this result we quantitatively confirmed the qualitative TEM-based hypothesis^{12} that the nano-pore elongation direction follows a major crystal axis (Fig. S11†). We also proved that we are able to analyze complex 3D porous networks utilizing FIB-SEM tomography with a higher resolution than 10 nm.

3.3.1 Electron–hole pair generation rate and local current density. An isotropic subsample with a volume of 512 μm^{3} was used for the following calculations. The subsample was eight times larger than the representative volume (Fig. S5†) and represents well the key morphological characteristics (volume fraction profile and surface area) of the measured LTON film (Fig. S6†). To model the light absorption, 25 × 25 volume fraction profiles along the z-axis were calculated from the subsample. For each column, the generation rate per solid volume was calculated applying Beer–Lambert's law with a discretization step of 40 nm in z-direction. For the wavelength-dependent absorption coefficient (Fig. S12†), and a bandgap of 2.1 eV, 77% of the incident light of the AM1.5G spectrum was absorbed within the LTON film. This result is in agreement with the experimentally determined absorbance of 72.3%. The generation rate for front illumination was symmetrically distributed, with a peak just beyond half of the film thickness. In contrast, the generation rate profile showed a maximum in the first micrometer and then declined (Fig. S13†). This difference is directly linked to the morphology, as the first particle layers for the front illumination are much more loosely packed compared to the layers above the FTO. Assuming 100% absorbed photon-to-current efficiency, the generated charges were distributed onto the closest semiconductor–electrolyte interface. The current density distributions are shown in Fig. S14.† The most frequent current density was 0.12 mA cm^{−2} for front illumination, and 0.17 mA cm^{−2} for back illumination. As expected, the back illumination led to higher maximum local current densities (4.31 mA cm^{−2} for front illumination and 4.99 mA cm^{−2} for back illumination).

The measured absorption coefficient showed a weak wavelength dependency between 300 nm and 590 nm (bandgap). Therefore, the generation rate profile was calculated for a constant wavelength-averaged absorption coefficient of 1.4 × 10^{4} cm^{−1}. The mean deviation from the exact profile was 1.1% and 1.6%, for front and back illumination respectively. Based on this finding, the generation rate profiles and the local current densities were calculated for constant absorption coefficients ranging from 5 × 10^{3} to 1 × 10^{5} cm^{−1} and a constant bandgap of 2.1 eV. The results showed that the influence of the morphology on the generation rate profile increases with higher absorption coefficients. For an absorption coefficient of only 5 × 10^{4} cm^{−1}, the shape of the generation rate profile is already dictated by the volume fraction profile. A maximum absorbance of 97.3%, equivalent to a maximum current of 9.5 mA cm^{−2} (per flat photoelectrode area), can be reached with a high absorption coefficient of 1 × 10^{5} cm^{−1}; the remaining part is always lost due to small areas not covered by any particle. For all absorption coefficients, the maximum current density for front illumination was 9–13% lower than for back illumination. Also, for low absorption coefficients (<8 × 10^{3} cm^{−1}), evenly distributed current densities with low values (<4 mA cm^{−2}) were found, which is desirable in order to avoid degradation. High absorption coefficients (>5 × 10^{4} cm^{−1}) resulted in big fractions of the surface area having high current densities, up to 11 mA cm^{−2}.

3.3.2 Mass transport and local concentration distribution. The diffusion of reactants into the pore space of the LTON photoelectrode was studied. The same subsample as previously discussed was considered and shown in Fig. 5. The incident-light-to-charge-transfer-rate-conversion (ILCC) was calculated as a function of the intermediate reactant concentration, located above the LTON film at the interface to the macroscopic hydrodynamic boundary layer. The charge transfer rate integrated over the semiconductor–electrolyte interface is determined by the number of photons absorbed in the semiconductor and the number of reactants diffusing to the interface. The results are shown as absorbed-light-to-charge-transfer-rate-conversion (ALCC) by multiplying with the absorbance (Fig. 6a). For ALCC = 1, the mass transport did not limit photoelectrode performance for a given c_{i}, but once the transport limitation was reached, the ALCC decreased sharply with decreasing c_{i}. For back illumination, the decrease occurred at higher c_{i}, and the initial slope was steeper, because the electrolyte close to the back (where the current density is the highest) was depleted first. Similar curves with different starting points for the transport limits were calculated for various diffusion coefficients D, ranging from 10^{−4} to 10^{−6} cm^{2} s^{−1}, including the diffusion coefficients of OH^{−} (D = 5.28 × 10^{−5} cm^{2} s^{−1}) and dissolved CO_{2} (D = 2.02 × 10^{−5} cm^{2} s^{−1}) in water. The results can be approximated by the following analytical equation:

where c_{i} [mol m^{−3}] is the intermediate concentration. a and b are parameters that depend on the properties of the semiconductor, the electrolyte, the irradiance, and the morphology:

ALCC = ac_{i}e^{(−bci)} ≤ 1
| (1) |

a = Φ_{>Eg}AnD(a_{1}/α^{2} + a_{2}) [m^{3} mol^{−1}]
| (2) |

b = Φ_{>Eg}AnD(b_{1}/α^{2} + b_{2}) [m^{3} mol^{−1}]
| (3) |

Both a and b are linearly dependent on: the integral of the incoming photon flux (with energies above the bandgap), Φ_{>Eg} [m^{−2} s^{−1}], the absorbance, A [—], the number of charges transferred at the semiconductor–electrolyte interface per reactant, n [—], and the diffusion coefficient of the reactant in the fluid phase, D [m^{2} s^{−1}]. Finally, there is an inverse quadratic dependency of a and b on the absorption coefficient α [m^{−1}]. The first two parameters define the maximum possible current per irradiated flat area, and the multiplication with n defines the maximum possible reactant flux through a plane perpendicular to the irradiation direction. Any increase in this overall flux shifts the curve of the ALCC to higher c_{i}, and a decrease shifts it to lower c_{i}. In Fig. 6a, the linear dependency between the diffusion coefficient and the parameters a and b has been demonstrated by fitting a = a_{0}D and b = b_{0}D in eqn (1) to all the simulated data points (same absorption coefficient), with a RMS = 0.014 for front illumination and a RMS = 0.016 for back illumination. As shown in Fig. S16,† the ALCC highly depends on the generation rate profile, or more specifically, the absorption coefficient and the volume fraction profile. For front and back illumination, the results showed a shift to higher c_{i} for larger absorption coefficients. For the front illumination, this was mainly due to the increased generation rate. For the back illumination, the shift to higher c_{i} is much more pronounced, as a larger absorption coefficient also shifts the generated charges to lower positions in the film, thus making it more difficult for the reactants to reach. In both cases, the ALCC curves asymptotically merged with high absorption coefficients. As previously discussed, in this range of high absorption coefficients, the generation rate profile is purely dependent on morphology, and thus constant. This behavior can be described by an inverse quadratic dependency with the fitting parameters a_{1}, a_{2}, b_{1}, b_{2}. The parameters a_{1} and b_{1} are weights to control the decrease of a and b with increasing absorption coefficients. The parameters a_{2} and b_{2} characterize the effect of morphology on the ALCC. Eqn (1) to (3) were fitted to the data points in Fig. S16,† respecting the previous fit for a_{0} and b_{0}. For front illumination, the parameters a_{1} = 2.254 × 10^{−1}, a_{2} = 4.714 × 10^{−13}, b_{1} = 8.217 × 10^{−2}, and b_{2} = 1.709 × 10^{−13} were calculated with RMS = 0.032. For back illumination, the parameters a_{1} = 2.538 × 10^{−1}, a_{2} = 2.003 × 10^{−13}, b_{1} = 9.536 × 10^{−2}, and b_{2} = 6.871 × 10^{−14} were calculated with RMS = 0.041. The parameters a_{2} and b_{2} are more than two times smaller for back illumination than for front illumination, which confirms them as morphology-descriptive parameters. Smaller parameters a_{2} and b_{2} lead to a shift of the ALCC to higher c_{i}. The morphology has a much stronger effect on the ALCC for the back illumination case, as the generation rates are much deeper in the film, and mass transport limitations occur at higher concentrations.

The correlation found between c_{i} and ALCC can be combined with a macroscopic Sherwood correlation, describing the hydrodynamic boundary layer, to link the ALCC to a bulk concentration, c_{bulk}. To quantify a realistic hydrodynamic boundary layer, the LTON photoelectrode was considered to be in a PEC flow cell,^{14} and the transport of OH^{−} ions into the pore space of the LTON electrode for the oxygen evolution reaction in a water splitting device was studied (Fig. 5). For the specific case of a flow cell, the Sherwood correlation described in the Methods was used. However, eqn (1) can be linked to any other appropriate correlation in another reactor setup. The ALCC as a function of the bulk pH of an aqueous NaOH solution is shown in Fig. 6b for 3 different flow regimes: almost stagnant, Re = 4; medium velocity, Re = 143; and maximum velocity in laminar flow, Re = 2300. The bulk pH for which the mass transport through the macroscopic concentration boundary layer begins to be limiting is indicated with a dotted line. For Re = 4, the macroscopic mass transport limit was predicted to begin below a bulk pH of 12.6. The microscopic LTON structure had a minimal effect in this flow regime, and the illumination direction did not play a role. For Re = 143 and Re = 2300, the mass transport limitation was due to LTON morphology, and the difference between front and back illumination was more pronounced for higher Re. For the given case, mass limitations occurred only for relatively low bulk pH values (bulk pH < 12.6, Re = 4), and in a flow cell setup, those mass transport limitations were shown to be effectively reduced by increasing the velocity (bulk pH < 11.82, Re = 2300). However, with our multi-scale correlation, we not only predict the bulk pH below which the mass transport is limited, but also describe the ALCC as a function of the bulk pH. In addition, the local surface concentrations, c_{s}, along the semiconductor–electrolyte interface are quantified for different diffusion coefficients at specific ALCC values. Fig. 6c shows the results for D = 5.28 × 10^{−5} cm^{2} s^{−1}, D = 2.02 × 10^{−5} cm^{2} s^{−1}, and D = 5.00 × 10^{−6} cm^{2} s^{−1} at ALCC = 1, 0.8, 0.6, 0.4, and 0.2. The concentration distribution is analyzed by using histogram bins in the logarithmic scale. The surface fractions for surfaces with zero concentration are, by definition, identical for all three diffusion coefficients at a given ALCC. For decreasing ALCC values, the surface fraction of surfaces in depleted electrolyte regions increases. For both illumination directions, the depletion of the electrolyte began in regions above the FTO, which are the furthest from the bulk concentration. In the case of front illumination, a sharp increase of surfaces with zero concentration was observed from ALCC = 1 to 0.6. The charges are generated high up in the film, and almost 60% of the surfaces were depleted for a 40% decrease of the overall mass flux. The rate of surface depletion diminished afterwards for lower ALCC values. For back illumination, the surface fraction of depleted surfaces was always much lower compared to front illumination for the same ALCC. For back illumination, the peak of the generation rate was confined to the bottom of the LTON film and the ALCC was influenced by the first surface areas that were depleted. Also, the surface fraction of depleted surfaces increased linearly with decreasing ALCC. In Fig. 6c, the order of magnitude of the intermediate concentration for each ALCC and diffusion coefficient is indicated above each histogram bar. For ALCC = 1, the calculated surface concentrations were, at most, one order of magnitude smaller than the intermediate concentrations, resulting in small concentration gradients throughout the pore space of the structure. This quickly changed for ALCC = 0.8, where a wide distribution of surface concentrations was observed. The magnitude of the diffusion coefficient determined the surface fraction at each concentration. Larger diffusion coefficients resulted in a broader concentration distribution along the film thickness, and low surface concentrations (10^{−3} to 10^{−1} mol m^{−3}) had a higher surface fraction. The concentration gradient was much higher for low diffusion coefficients, and most of the surface concentrations were closer to the intermediate concentration for a given ALCC. This was observed for front and back illumination. In general, the computed results provide unique insight into the influence of morphology on local surface concentration distributions. The data can be used to study the degradation of structured semiconductor electrodes due to locally quantified pH values in which the semiconductor is not stable. The specific case of the LTON photoelectrode in aqueous NaOH electrolyte was analyzed and the surface fractions of local pH values vs. bulk pH values are shown in Fig. S17.† The ALCC was below 1 for bulk pH values smaller than 12.6 and Re = 4. For higher bulk pH values, the local pH on the electrolyte–semiconductor interface was not affected by the morphology. However, for smaller bulk pH values, e.g. 12.5, local pH values (down to pH 7) were found. For LTON, which is stable only at high pH values (above 13), we never reached a condition where morphology would locally reduce the pH and degrade the semiconductor. The introduced approach could also be applied to an electrode for CO_{2} reduction, where local pH gradients are also desired to promote the activity of a catalyst towards the production of a specific product.

The tortuosity was calculated for the same conditions as the surface concentration calculations and is shown for front and back illumination in Fig. 6d. The tortuosity was defined as the mean path length of the reactants travelling from the intermediate concentration plane to the semiconductor surface divided by the LTON film thickness. Therefore, the tortuosity can be smaller than 1, as most reactants do not travel through the whole film thickness (Fig. S18†). The path lengths were normal distributed and the error bars indicate the standard deviations. The tortuosity was generally larger for back illumination, as the key parameter is the generation rate profile calculated from the light absorption (Fig. S13†). For both illumination directions, the tortuosity decreased for smaller ALCC values due to the creation and growth of the depletion regions. For a given ALCC value, the influence of the diffusion coefficient was negligible. Despite the complex structure of the LTON film, the mean values of the path lengths did not exceed the film thickness, even in the case of no mass transport limitations.

Detailed semiconductor material distribution profiles along the electrode thickness were obtained and provide relevant information for light absorption. Surface characterizations on the exact geometry allowed decoupling of contributions from the meso-scale features and the nano-scale pores, suggesting that the connected nano-pore surfaces represent 31% of the overall surface. Advanced segmentation of the tomography data, based on machine learning algorithms which can distinguish different grey value patterns, led to the identification of the TiO_{2} necking phase in the LTON electrode and revealed that necking spans short distances between particles only. These results prove that majority carriers must be transported through multiple particles and necking patches in order to reach the FTO back-contact. An algorithm was developed which performs ellipsoid fitting into the particle-based LTON structure. This algorithm allows for the identification and characterization of individual particles in the highly interconnected particle network. Quantitative characterization of single LTON particles was then obtained by calculating frequency distributions of particle diameters and volumes, as well as particle shapes and orientations. The surface area of the particles was calculated and contributions from the characteristic particle facets, which are linked to crystallographic planes, were estimated. The high resolution of the FIB-SEM tomography gave quantitative insight into the nano-pore volumes within the LTON particles and allowed determination of the crystallographic directions, to which the pores were preferentially elongated.

Multi-physical transport characterizations were performed on the LTON photoelectrode using direct pore-level simulations. The exact 3D morphological geometry was used to investigate light absorption in the semiconductor and mass transport through the liquid space of the LTON particle network. Current densities for front and back illumination were calculated and imposed as a boundary condition at the semiconductor–electrolyte interface. The influence of the absorption coefficient on the generation rate profile was analyzed and resulting current density distributions were quantified. The mass transport calculations were done on multiple scales, simulating the purely diffusive transport of reactants through the stagnant pore space and correcting the result with a Sherwood correlation to take into account the concentration gradients in the macroscopic hydrodynamic boundary layer. The mass transport induced performance drop of the photoelectrode was described as a function of the semiconductor and electrolyte properties, the incident radiation, and ultimately linked to the morphology. Local concentration distributions on the semiconductor–electrolyte interface were quantified for various ALCC values and diffusion coefficients. For the same conditions, the tortuosity of the reactants within the LTON film was calculated.

This study presents the first in-depth geometrical quantification and multi-physical transport characterization of morphologically-complex photoelectrodes. The data acquisition with the FIB-SEM nano-tomography provides the detailed morphology-information to study the film structure, characterize the individual building blocks on multiple scales, and to run pore-level transport simulations on a representative volume. The method we developed contributes to guiding the morphology and fabrication of optimized photoelectrodes with enhanced performance.

Image Registration—despite the 3D tracking during data acquisition, translational drift occurred. This remaining drift was corrected by a registration process utilizing three notches milled into the platinum layer along the sectioning direction as alignment marks. The mark on the first image was used as a template and matched by a normalized correlation coefficient matching method to the following marks to determine the image translation and to align the slices using the plug-in Template Matching.^{17} In the case of the relatively thick LTON sample, the images were inconsistently compressed in the height of the image due to charging effects that can deflect the electron beam. This compression was compensated for by aligning the bottom of the FTO throughout the stack and by expanding or compressing the individual images accordingly. Typical compression factors were in the range of 4 to 10 voxels.

Filtering and 3D rendering—a major challenge was the removal of air bubble artifacts in the images. These bubbles formed in the epoxy during the sample preparation. In Fig. S1a,† the bubbles can be seen on the surface of the LTON particles, or in narrow gaps between the particles. A low pass filter was applied in the frequency domain, derived by a discrete Fourier transformation, to blur the pores in the particles. A histogram-based threshold segmentation and a stack erosion step, with a structuring disk of 5 pixels in diameter, was applied on the filtered images. After the erosion of the solid phase, the artifacts were disconnected from the solid phase and, by selecting only the larger connected solid volumes, a binary mask was obtained to remove the signal from the bubbles in the SEM images.

Segmentation—we used a grey value histogram-based segmentation in order to differentiate between the solid and void domains and eventually digitalize the reconstructed dataset of grey values. A coarse segmentation of the structure for the two introduced photoelectrodes was straightforward, as the epoxy in the void domain had a high carbon level with a big electron beam interaction volume leading to a low signal to noise ratio. The fine tuning of the segmentation, however, highly depended on the quality and contrast of the acquired SEM images and needed to be adjusted for each dataset individually. For sample (i), the smallest hematite features had a diameter of a few nanometers and suffered from edge effects. Edge effects appear when the scanning electron beam hits a small feature. Secondary electrons leave the feature through all surrounding edges, resulting in an increased brightness, but a less sharp solid–void interface. We used an 8 bit threshold of 125 for the segmentation. This threshold lied symmetrically between the two extreme thresholds of 100, where all grey values except the background were included, and 150, where small features started to disconnect from the main solid phase.

For sample (ii), the outer interfaces between the LTON particles and void were well defined and sharp. However, the inner pores were not properly filled with epoxy and showed weak contrast. The necking phase between the particles could be visually distinguished from the LTON by the texture and contrast in the electron beam image, and was segmented by a machine-learning algorithm trained on manually selected image features using the Trainable WEKA Segmentation plug-in.^{18} Before applying the plug-in, the contrast was improved by contrast limited adaptive histogram equalization^{19} using the Enhance Local Contrast (CLAHE) plug-in in Fiji. The quality of the segmentation could only be assessed visually and is shown in Fig. S1b.† Fig. S1c† shows the 3D rendering of the TiO_{2} necking. An animated version is available in ESI.† The individual TiO_{2} phases were labeled and connected TiO_{2} phases were given the same color. For the first time, the TiO_{2} necking was visualized and quantified, and its ability to connect the LTON particles, ideally over a length scale of multiple particles, was assessed. Once the TiO_{2} phase is subtracted, the LTON was segmented with an 8 bit threshold value of 92. An upper threshold of 106, where all the nano-pores were connected, and a lower threshold of 78, where only volumes entirely filled by epoxy were considered to be pores, were used to define error bars on the results, i.e. the sensitivity of the obtained results on the threshold accuracy.

(4) |

(X − X_{0})^{T}H(X − X_{0}) ≤ 1,
| (5) |

(6) |

For the calculation of the generation rate, a subsample of the segmented FIB-SEM data set was discretized in all three directions. For constant x = i, y = j and z = k values, the generation rate per solid volume, G_{i,j,k} was calculated as:

(7) |

(8) |

(9) |

∇(D∇c) + S_{c} = 0
| (10) |

(11) |

The void phase of the FIB-SEM dataset was meshed using a tetrahedral mesh generator^{26} and the diffusive transport equation was solved with a commercial finite volume solver.^{27} A schematic of the simulation setup is show in Fig. S15.† An intermediate concentration c_{i} was set as a top boundary condition. The side walls had symmetry boundary conditions and the bottom, which was covered by FTO had a no-flux boundary condition. The calculated local current density was applied as a reactant sink boundary condition onto the electrolyte–semiconductor interface. The boundary flux was given as a step function, being the charge flux calculated by the generation rate for a boundary concentration c_{Boundary} > 0, and being 0 for c_{Boundary} = 0. It was assumed that the photo-induced potential drop occurs only in the space-charge layer of the semiconductor and in the Helmholtz layer,^{28} thus, ions are not driven by an electric field, but only travel by diffusion. This assumption is correct for ionic concentrations in the electrolyte with a small Debye length. For our calculations, the ionic strength in most parts of the domain is high enough to keep the Debye length below 1 nm; however, for the domain regions with near depletion, the Debye length increases and contributions from the electric field through migration were neglected. The mass transport through the macroscopic hydrodynamic boundary layer in a laminar-flow parallel plate cell was taken into account by the experimentally determined Leveque-correlation:^{29}

(12) |

The tortuosity was calculated with the mean path lengths of the reactants divided by the film thickness. 2500 pathlengths were extracted from the derivative of the reactant concentration field. The start points of the pathlengths were uniformly distributed on the plane of the intermediate concentration c_{i}.

- H. R. Kim, G. Kim, S. Il In and Y. Park, Optimization of porous BiVO
_{4}photoanode from electrodeposited Bi electrode: Structural factors affecting photoelectrochemical performance, Electrochim. Acta, 2016, 189, 252–258 CrossRef CAS. - O. K. Varghese, M. Paulose and C. A. Grimes, Long vertically aligned titania nanotubes on transparent conducting oxide for highly efficient solar cells, Nat. Nanotechnol., 2009, 4(9), 592–597 CrossRef CAS PubMed.
- M. T. Mayer, C. Du and D. Wang, Hematite/Si nanowire dual-absorber system for photoelectrochemical water splitting at low applied potentials, J. Am. Chem. Soc., 2012, 134(30), 12406–12409 CrossRef CAS PubMed.
- Y. Qiu, et al., Efficient photoelectrochemical water splitting with ultrathin films of hematite on three-dimensional nanophotonic structures, Nano Lett., 2014, 14(4), 2123–2129 CrossRef CAS PubMed.
- J. Li, et al., Si nanopillar array optimization on Si thin films for solar energy harvesting, Appl. Phys. Lett., 2009, 95(3), 33102 CrossRef.
- Z. Zhang and P. Wang, Optimization of photoelectrochemical water splitting performance on hierarchical TiO
_{2}nanotube arrays, Energy Environ. Sci., 2012, 5(4), 6506 RSC. - A. Kay, I. Cesar and M. Grätzel, New benchmark for water photooxidation by nanostructured alpha-Fe
_{2}O_{3}films, J. Am. Chem. Soc., 2006, 128(49), 15714–15721 CrossRef CAS PubMed. - S. Landsmann, A. E. Maegli, M. Trottmann, C. Battaglia, A. Weidenkaff and S. Pokrant, Design Guidelines for High-Performance Particle-Based Photoanodes for Water Splitting: Lanthanum Titanium Oxynitride as a Model, ChemSusChem, 2015, 8(20), 3451–3458 CrossRef CAS PubMed.
- J. McKone and N. Lewis, Photoelectrochemical water splitting: materials, processes and architectures, ed. H.-J. Lewerenz and L. Peter, Royal Society of Chemistry, 2013, pp. 52–82 Search PubMed.
- S. Dilger, S. Landsmann, M. Trottmann and S. Pokrant, Carbon containing conductive networks in composite particle-based photoanodes for solar water splitting, J. Mater. Chem. A, 2016, 4(43), 17087–17095 RSC.
- S. C. Warren, et al., Identifying champion nanostructures for solar water-splitting, Nat. Mater., 2013, 12(9), 842–849 CrossRef CAS PubMed.
- S. Pokrant, S. Dilger and S. Landsmann, Morphology and mesopores in photoelectrochemically active LaTiO
_{2}N single crystals, J. Mater. Res., 2016, 1–6 Search PubMed. - M. Yashima, M. Saito, H. Nakano, T. Takata, K. Ogisu and K. Domen, Imma perovskite-type oxynitride LaTiO
_{2}N: structure and electron density, Chem. Commun., 2010, 46(26), 4704 RSC. - J. T. Davis and D. V. Esposito, Limiting photocurrent analysis of a wide channel photoelectrochemical flow reactor, J. Phys. D: Appl. Phys., 2017, 50(8), 84002 CrossRef.
- M. Cantoni and L. Holzer, Advances in 3D focused ion beam tomography, MRS Bull., 2014, 39(4), 354–360 CrossRef CAS.
- J. Schindelin, et al., Fiji: an open-source platform for biological-image analysis, Nat. Methods, 2012, 9(7), 676–682 CrossRef CAS PubMed.
- Q. Tseng, et al., A new micropatterning method of soft substrates reveals that different tumorigenic signals can promote or reduce cell contraction levels, Lab Chip, 2011, 11(13), 2231 RSC.
- I. Arganda-Carreras, V. Kaynig, C. Rueden, A. Schindelin, J. Cardona and H. S. Seung, Trainable_Segmentation: Release v3.1.2, 2016 Search PubMed.
- K. Zuiderveld, Contrast limited adaptive histogram equalization, Academic Press Professional, Inc., 1994 Search PubMed.
- W. E. Lorensen and H. E. Cline, Marching cubes: A high resolution 3D surface construction algorithm, ACM SIGGRAPH Computer Graphics, 1987, 21(4), 163–169 CrossRef.
- H. J. Vogel, Morphological determination of pore connectivity as a function of pore size using serial sections, Eur. J. Soil Sci., 1997, 48(3), 365–377 CrossRef.
- P. Soille, Morphological image analysis: principles and applications, Springer Science & Business Media, 2013 Search PubMed.
- M. Doube, The ellipsoid factor for quantification of rods, plates, and intermediate forms in 3D geometries, Front. Endocrinol., 2015, 6, 1–5 Search PubMed.
- M. Doube, et al., BoneJ: Free and extensible bone image analysis in ImageJ, Bone, 2010, 47(6), 1076–1079 CrossRef PubMed.
- M. Longair, Find Connected Regions Search PubMed.
- H. Friess, S. Haussener, A. Steinfeld and J. Petrasch, Tetrahedral mesh generation based on space indicator functions, Int. J. Numer. Methods Eng., 2013, 93, 1040–1056 CrossRef.
- ANSYS® Academic Research CFX, Release 17.1 Search PubMed.
- Y. K. Gaudy and S. Haussener, Utilizing modeling{,} experiments{,} and statistics for the analysis of water-splitting photoelectrodes, J. Mater. Chem. A, 2016, 4(8), 3100–3114 RSC.
- J. Qi and R. F. Savinell, Mass transfer in a laminar-flow parallel plate electrolytic cell with simultaneous development of velocity and concentration boundary layers, J. Appl. Electrochem., 1990, 20(6), 885–892 CrossRef CAS.

## Footnote |

† Electronic supplementary information (ESI) available. See DOI: 10.1039/c8se00215k |

This journal is © The Royal Society of Chemistry 2018 |