A dye-sensitized solar cell was solvated by RTILs; using optimized empirical potentials, a molecular dynamics simulation was applied to compute vibrational properties. The obtained vibrational spectra were compared with experiment and ab initio molecular dynamics; various empirical potential spectra show how partial-charge charge parameterization of the ionic liquid affects vibrational spectra prediction.
The accurate molecular simulation prediction of vibrational spectra, and other structural, energetic and spectral characteristics, of photo-active metal-oxide surfaces in contact with light-absorbing dyes is an ongoing thorny and elusive challenge in physical chemistry. With this in mind, a molecular-dynamics (MD) simulation were performed by using optimized empirical potentials for a well-representative and prototypical dye-sensitized solar cell (DSC) solvated by a widely-studied room temperature ionic liquid (RTIL), in the guise of a [bmim]+[NTf2]– RTIL solvating an N719-sensitizing dye adsorbed onto 101 anatase-titania. In doing so, important insights were gleaned into how using a RTIL as the electrolytic hole acceptor modulates the dynamical and vibrational properties of a N719 dye, estimating the spectra for the DSC photo-active interface via Fourier transformation of mass-weighted velocity autocorrelation functions from MD. The acquired vibrational spectra were compared with the experiment spectra and those sampled from ab initio molecular dynamics (AIMD); in particular, various empirical-potential spectra generated from MD provide insight into how partial-charge charge parameterization of the ionic liquid affects vibrational spectra prediction. In any event, careful fitting of empirical force-field models has been shown to be an effective tool in handling DSC vibrational properties, when validated by AIMD and an experiment.
In dye-sensitized solar cells (DSCs), the optical band gap of semiconductors is bridged by a light-absorbing, or sensitizing, dye. DSCs require continual recharging: therefore, a redox electrolyte is essential to foster this constant supply of charge (typically in the form of an I–/I3-, in an organic solvent). This facilitates the passage of holes from the sensitizing dye to the electrolyte, with injected photo-excited electrons into the metal-oxide substrate passing through to an external circuit, with eventual recombination taking place at the cathode1. A crucial aspect underpinning DSCs' positive outlook for a wide variety of real world applications originates in their straightforward manufacture, without the need for raw materials high in purity; this is in stark contrast with the high capital cost and ultra-purity required for silicon-based photovoltaics. In any event, the prospect of significantly improving the working-life timescales of DSCs by swapping less stable electrolytes with room-temperature ionic liquids (RTILs) having low volatility shows significant promise. The solid-like physical properties of RTILs combined with their liquid-like electrical properties (such as low toxicity, flammability, and volatility)1 lead these to constitute rather excellent candidate electrolytes for usage in DSC applications.
Given such prospects for RTILs in DSCs, it is hardly surprising that, in recent years, there has been a substantial boost of activity in studying DSC-prototype N719-chromophore/titania interfaces with RTILs. In particular, important work on such systems has been performed2,3,4,5, which consider a broad suite of physico-chemical processes, including the charge-replenishment kinetics in dyes2,5, the mechanistic steps of electron-hole dynamics and transfer3, and, of course, the effects of titania substrates' nanoscale nature upon these, and other, processes4.
Now, bearing in mind impressive advances in DFT-based molecular simulation, particularly AIMD6, as a highly useful prototypical design tool in materials science and particularly for DSCs7,8,9,10,11, with critical assessment of optimal functional selection being vital8,9, AIMD techniques have proven very useful previously in scrutinizing rather significant dispersion and explicit-RTIL solvation effects on dye structure, adsorption modes and vibrational properties at DSC-semiconductor surfaces. In particular, the adoption of AIMD had led to some success in attaining reasonable, semi-quantitative capture and prediction of important electronic properties, such as band gap, as well as structural binding13 and vibrational spectra14 In refs. 12-14, AIMD simulations were performed extensively on the photo-active N719-chromophore dye bound to (101) anatase-titania surface, assessing both electronic properties and structural properties in the presence of both [bmim]+[NTf2]– 12,13 and [bmim]+[I]– 14 RTILs, in addition to vibrational spectra for the case of [bmim]+[I]– 14. In particular, the rigidity of the semiconductor's surface15, apart from its inherent comparative photo-activity, led the surface to slightly alter within the AIMD simulation, which makes (101) anatase interfaces12,13,14 a suitable choice. As ref. 12 shows, the mean distance between the cations and the surface dropped by about 0.5 Å, the average separation between the cations and anions decreased by 0.6 Å, and the noticeable altering of the RTILs in the first layer around the dye, where the cation was on average 1.5 Å further from the center of the dye, were directly caused by explicit dispersion interactions in RTIL-solvated systems. Unphysical kinking of the adsorbed N719 dye's configuration was also a result of the introduction of explicit dispersion effects in vacuo. In ref. 13, analysis was conducted on whether these structural effects of explicit RTIL solvation and functional selection affected the behavior of the DSSCs, concluding that both explicit solvation and treatment of dispersion is very important. In ref. 14, with high-quality experimental vibrational-spectral data of other groups on hand, the particular effects were benchmarked systematically on both explicit [bmim]+[I]– solvation and accurate handling of dispersion established in refs. 12 & 13 on the reproduction of salient spectral-mode features; this led to the conclusion that explicit solvation is important, alongside accurate treatment of dispersion interactions, echoing earlier findings for both structural and dynamical properties in the case of AIMD modeling of catalysts in explicit solvent16. Indeed, Mosconi et al. have also performed an impressive assessment of explicit-solvation effects on DFT treatment of DSC simulation17. Bahers et al.18 studied experimental absorption spectra for dyes along with the related spectra at the TD-DFT level; these TD-DFT spectra agreed very well in terms of their computed transitions with their experimental counterparts. In addition, absorption spectra of pyrrolidine (PYR) derivatives were studied by Preat et al. in several solvents19, providing significant insights into the dyes' geometrical and electronic structures, and evincing adequate structural modifications that serve to optimize the properties of the PYR-based DSSCs – a spirit of simulation-led/rationalized 'molecular design', indeed.
Having clearly established the important contribution of both DFT and AIMD towards accurate modelling of DSCs' properties and function, including such important technical matters like explicit solvation and appropriate treatment of dispersion interactions from structural, electronic and vibrational standpoints7,8,9,10,11,12,13,14, now – in the present work – the focus turns towards the pragmatic question of how well empirical-potential approaches can be tailored to address the apposite and reasonable prediction of structural and vibrational properties of such prototypical DSC systems, taking the N719 dye adsorbed on anatase (101) in the [bmim]+[NTf2]– RTIL as a case in point. This is important, not only because of the large corpus of forcefield-based molecular-simulation activities and methodological machinery available to tackle DSC simulation7, and metal-oxide surfaces more widely, but also because of their staggeringly reduced computational cost vis-à-vis DFT-based approaches, together with the possibility of very efficient coupling to biased-sampling approaches to capture more efficiently phase space and structural evolution in highly viscous RTIL solvents, dominated by solid-like physical properties at ambient temperatures. Therefore, motivated by this open question of gauging and optimizing forcefield approaches, informed by both DFT and AIMD as well as experimental data for vibrational spectra14, we turn to the pressing task of assessing empirical-potential performance at vibrational-spectra prediction from MD, using mass-weighted Fourier transforms of the N719 dye's atomic velocity autocorrelation function (VACF). One key concern is how different partial-charge parameterizations of the RTIL may affect vibrational-spectra prediction, and particular attention was given to this point, as well as the wider task of tailoring forcefields for optimal spectral-mode prediction relative to experiment and AIMD20.
1. Performing MD Simulation using DL_POLY
2. Performing forcefield-based MD Simulation using DL_POLY
3. Comparing results of each of the force-fields
NOTE: It is important to assess partial-charge sets for the RTIL for empirical-potential-based MD simulation in step 2, for ready comparison against each other, experiment and ab initio-MD results in explicit RTIL solvent (using the PBE functional with Grimme-D3 dispersion, given its superior performance for prediction of vibrational spectra)20; these were as follows:
Structural Properties of Binding Motifs
Representative binding motifs for the four different partial-charge sets are depicted in Figure 2, after 15 ps of MD. In Figure 2a, for the (above-described) literature-derived charges, it can be seen that there is a prominent hydrogen-bonding interaction with a surface proton. From careful analyses of the trajectory, the hydrogen bonds are mostly surface-proton-bound whilst the other three (AIMD-derived)20 charge sets do not feature such a strong Coulombic interaction with a surface proton. Referring to Figure 1f of ref. 20, which depicts the AIMD-relaxed dye-binding configuration after ~8.5 ps, there is also less evidence of strong hydrogen-bonding with a surface proton, so this is qualitatively consistent in tailoring the present forcefield-based MD with AIMD-based charge sets to achieve qualitatively similar substrate-binding arrangements (Figure 2b-d). Indeed, the smaller magnitude of the partial charges in the literature-derived case vis-à-vis those sampled in various ways from AIMD (Table 1 & Table 2) leads to a lesser extent of charge shielding in comparison to larger-magnitude partial RTIL charges, which serves to emphasize more the electrostatic (hydrogen-bond) interaction with the surface proton evident in Figure 2a. In any event, interestingly, the Mulliken-derived charge set shows a certain sustained 'kinking' of the dye to have a prominent hydrogen bond with a bridging oxygen atom at the anatase surface (Figure 2b), which is redolent of the PBE system of ref. 15 without Grimme dispersion (Figure 2d therein): the generally-recognized inferior quality of Mulliken charges leads to this less physical, persistent kinking, which has been studied in more detail in refs. 12, 13 & 20. Tellingly, the better-quality charge fits (EHT and Hückel) from AIMD20 lead to more realistic N719-binding motifs in Figure 2c,d, which are in accord with PBE-based BOMD featuring Grimmer-D3 dispersion in ref. 20 (cf. Figure 1f therein); comparison of motifs with ref. 20 shows this is somewhat more the case for Hirshfeld charges.
Empirical-Potential Spectra
Having established at gross structural level already the clear influence of RTIL partial-charge sets in parameterizing and tailoring pragmatically the efficacy of forcefield-based MD vis-à-vis the best-quality AIMD data available, we now turn to considering empirical-potential-based MD for replication of N719 vibrational spectra. The mass-weighted VACF spectra for the four different parametrized charge sets are shown in Figure 3; as mentioned previously, although all four MD-generated spectra have the same forcefields for their dyes and surfaces, they differ in the partial charges applied to the RTIL cations and anions.
Now, prior to discussing the MD-predicted vibrational spectra, let us make some brief explanatory comments about their more fundamental nature and higher-level interpretation. The continuous colored lines in Figure 3 denote the (empirical-potential) MD-based spectra in the range from 0 to 2500 cm-1, for all four RTIL partial-charge sets. The dashed vertical gray lines are established experimental modes for the N719 dye and are at the frequencies 1230, 1380, 1450, 1540, 1600, 1720 and 2100 cm-1, respectively36. The two grey spectral insets are experimental ATR-FTIR results from ref. 37, with the topmost one being the spectra for unsolvated N719 powder and the bottom one that for unsolvated N719 powder adsorbed on anatase. These experimental results are intended as a guide only, in that the spectra are slightly different in other studies and the two experimental insets are themselves somewhat different due to adsorption to anatase. A solvent's presence would be expected to alter the spectra, whether it be an RTIL or the more traditional acetonitrile. Also, it must be noted that the experimental spectra have a reduced frequency window available, and are in effect denoting a mixture of dynamical properties of multiple dyes in various geometries; in contrast, it ought to be borne in mind that the results are for a single N719 molecule adsorbed to the anatase substrate, therefore leading to an inevitably more sharp signal.
Now, the modes themselves are discussed more in refs. 14 & 18; the present discussion focusses more on fidelity of each technique in reproducing these, as opposed to their underlying nature.
0-500 cm-1: The previous AIMD results (i.e., PBE-based BOMD with Grimme-D3 dispersion and explicit RTIL solvation)20 show a cluster of spectroscopic peaks in the 300-400 cm-1 region. In terms of closeness to the ab initio spectra, the classical charge sets were ranked in order from closest to furthest away as EHT, Hirshfeld, literature-based and Mulliken.
500-1000 cm-1: The ab initio MD spectra20 display prominent vibrational peaks at 625, 750 and 825 cm-1; the main peaks present in the classical spectra are at 600 and 800 cm-1 for the literature-derived charges, 525 and 800 cm-1 for the Mulliken charges, 675, 810, and 900 cm-1 for the EHT charges and 650,800 and 900 cm-1 for the Hirshfeld charge set. Although the literature and Mulliken charge sets reproduce grosso modo some of the features of the ab initio spectra, both the EHT and Hirshfeld-derived charge sets generate spectra which have their main peaks only 25-75 cm-1 above those of the ab initio ones. Given the much closer structural resemblance of the binding motif in Figure 2c,d relative to Figure 1f of ref. 20, this excellent, semi-quantitative agreement is encouraging for tailoring and optimising pragmatically forcefields using high-quality AIMD.
1000-1500 cm-1: The BOMD spectra evince strong peaks at 1000, 1300, and 1400 cm-1 20, whereas the main peaks present in the forcefield-based spectra are at 1075 and 1200 cm-1 for the literature-derived charges, 1080, 1350 and 1450 cm-1 for Mulliken charges, 1075 and 1200 cm-1 for the EHT charges and 1075 and 1250 cm-1 for the Hirshfeld charge set. Although the forcefield-based RTIL-charge sets produce results close to those of AIMD simulation20, a key difference between these and BOMD lies in the important fact that the present empirical-potential simulations allow only for physical adsorption rather than also for the possibility of chemical adsorption. Clearly, this will have a particularly strong effect in this vibrational-bending spectral region and serves to explain a good deal of the altering of vibrational modes related to surface binding.
1500+ cm-1: In the stretch-dominated frequency range above 1500 cm-1, the solvated ab initio spectra20 exhibit modes at 1525, 1575, 1600, 1700 and 2075 cm-1. All four forcefield-based spectra have modes in the region of 1525 cm-1, whilst none capture the thiocyano mode around 2075 cm-1. The literature-derived charge set produces spectra which shows vibrational modes at 1625 and 1700 cm-1, whereas spectra derived from the use of Mulliken charges result in modes at 1600, 1675 and 1775 cm-1. The EHT-generated spectra have a mode at 1700 cm-1, and the Hirshfeld-generated spectra does a rather excellent job of reproducing DFT-based MD results with modes at 1575, 1600 and 1700 cm-1. It is clear that the greater sophistication of the Hirshfeld charge fitting does pay important dividends with respect to basic qualitative reproduction of salient spectral features, showcasing further the important effect of RTILs and suitable treatment of their electrostatics on capturing the essential details of N719 vibrational properties.
Figure 1: Cation and anion images of the considered system. Carbon is shown in cyan, nitrogen in blue, oxygen in red, hydrogen in white, sulphur in yellow, and fluorine in pink. Please click here to view a larger version of this figure.
Figure 2: Frontal view showing the relaxed geometries of the systems under consideration, after 15 ps of MD. Carbon is shown in grey, nitrogen in blue, oxygen in red, hydrogen in white, titanium in silver, Sulphur in yellow and ruthenium in light green. The explicitly solvated systems are shown without RTIL ions for ease of viewing. Please click here to view a larger version of this figure.
Figure 3: Vibrational Spectra for adsorbed-N719 from MD; the forcefield-based MD simulations differ from each other only in their partial-charge parametrization of the RTIL. Within each plot the grey (lower) / (upper) inset corresponds to the experimental ATR-FTIR signal from León37, for (dry-N719 adsorbed onto anatase) / (dry-N719 powder). The dashed lines indicate established vibrational modes30. (a) Literature-derived RTIL Charges, (b) Mulliken RTIL Charges, (c) Extended-Huckel-Theory Charges, and (d) Hirshfeld RTIL Charges. Please click here to view a larger version of this figure.
Atom | Anion | |||
Literature | Mulliken | EHT | Hirshfeld | |
N | 0.15 | 0.005 | -0.76 | -0.1 |
C | -0.11 | -0.06 | 0.65 | 0.14 |
N | 0.15 | 0.005 | -0.76 | -0.1 |
C | -0.13 | -0.06 | 0.2 | 0.14 |
C | -0.13 | -0.06 | 0.2 | 0.14 |
C | -0.17 | -0.34 | 0.274 | 0 |
H | 0.21 | 0.2 | 0.15 | 0.11 |
C | -0.17 | -0.34 | 0.354 | 0 |
H | 0.21 | 0.18 | 0.15 | 0.09 |
H | 0.21 | 0.18 | 0.15 | 0.09 |
H | 0.13 | 0.18 | 0.08 | 0.08 |
H | 0.13 | 0.18 | 0.08 | 0.08 |
H | 0.13 | 0.18 | 0.08 | 0.08 |
C | 0 | -0.25 | -0.16 | 0 |
H | 0.13 | 0.18 | 0.08 | 0.08 |
H | 0.13 | 0.18 | 0.08 | 0.08 |
C | 0 | -0.25 | -0.16 | 0 |
H | 0.045 | 0.151 | 0.08 | 0.04 |
H | 0.045 | 0.151 | 0.08 | 0.04 |
C | -0.17 | -0.34 | -0.24 | 0 |
H | 0.045 | 0.151 | 0.08 | 0.04 |
H | 0.045 | 0.151 | 0.08 | 0.04 |
H | 0.045 | 0.151 | 0.08 | 0.04 |
H | 0.045 | 0.151 | 0.08 | 0.04 |
H | 0.045 | 0.151 | 0.08 | 0.04 |
mean modification | 0.0062 | 0.0159 | — | 0.0195 |
Table 1: Different parametrized charge sets for the atomic partial charges of the cation. Mean modification is the per atom alteration of charge necessary to achieve overall charge neutrality.
Atom | Anion | |||
Literature | Mulliken | EHT | Hirshfeld | |
N | -0.368 | -0.44 | -0.368 | -0.62 |
S | 1.311 | 0.5 | 1.311 | 1.41 |
O | -0.717 | -0.3 | -0.717 | -0.64 |
C | 1.09 | 0.25 | 1.09 | 0.8 |
mean modification | — | 0.0062 | — | 0.0045 |
Table 2: The different parameterized charge sets for the atomic partial charges of the anion. Mean modification is the per atom alteration of charge necessary to achieve overall charge neutrality.
Ab initio simulation techniques are expensive to perform and hence to perform simulation on much longer timescales would require the use of empirical forcefields for at least some of the DSC system. Towards this end, an equivalent atomistic model was created of the [bmim]+[NTf2]– solvated interface, using an empirical, classical-simulation forcefield for MD. The anatase was modeled using the Matsui-Akaogi (MA) forcefield, whilst the dye structure was handled using OPLS parameters. For the RTIL, four different forcefields were employed to determine the importance of the charge of the RTIL on the dyes's vibrational dynamics. The force-field parameters were the same for each, and only the charge parameterization was varied, with the forcefield parameters generated from an automatic-fitting procedure utilizing extended Huckel theory (EHT), as implemented in the MOE software package and fitted against the ab initio trajectories.
The ab initio results show a cluster of spectroscopic peaks in the region of 300-400, 625, 750, 825, 1000, 1300, 1400, 1500, 1525, 1575, 1600, 1700 and 2075 cm-1. In terms of closeness to the ab initio spectra, the classical charge sets were ranked, in order from closest to farther away, as EHT, Hirshfeld, Literature, and Mulliken. The main peaks present in the classical spectra are at 600, 800, 1075, 1200, 1525 cm-1 for literature-derived charges. The peaks, using Mulliken charges, arise at 525, 800, 1080, 1350, 1450, 1525, 1625, 1700 cm-1. The EHT charges lead to peaks in the region of 675, 810, 900, 1075, 1200, 1525 cm-1. Finally, the Hirshfeld charge evinces peaks at 650, 800, 900, 1075, 1250 and 1525 cm-1.
Lest the reasonably-superior Hirshfeld-charge performance (both structurally and vibrationally), lead one to suspect those forcefield methods, when tailored pragmatically for charges informed by good-quality DFT and AIMD, may invariably serve as a reasonable substitute for AIMD in prototypical DSC design, this rather optimistic picture does not always hold. Indeed, where the empirical-potential models do fail in the present N719 scenario – somewhat spectacularly, it must be admitted – is in the spectral region between 1800 and 2000 cm-1. Both the PBE-Grimme-BOMD20 and experimental results37 show little activity in this region, whereas all classical-charge parametrizations show, erroneously and misleadingly, strong modes. This, and the already-mentioned lack of the thiocyano mode is most likely due to the forcefield model of the dye (OPLS, in the present case) being unable to capture properly this dynamical behavior. The charge parametrization, in this particular frequency range, would then be of secondary importance compared with the bonding and angular terms in the forcefield. This serves as a 'cautionary tale' that although partial-charge fitting is an important first step, the reparameterization of certain key stretch and bend interactions in the empirical models is likely to be important also.
In closing, vibrational spectra for the photo-active interface in a DSC were reported by means of Fourier transforming mass-weighted VACFs generated from the trajectories of forcefield-based molecular dynamics, using various RTIL-charge parameterizations of increasing complexity and sophistication. In addition, gross structural binding modes have also been considered (Figure 2), together with the sensitivity thereof to partial-charge sets. It has been found that the EHT, and Hirshfeld charge sets, performed reasonably well with respect to reproduction of gross binding motifs vis-à-vis high-quality AIMD, and also reproduction of salient ATR-FTIR vibrational-spectra features (as well as relative to AIMD spectra). However, a salutary and important lesson is that the lack of prediction of thiocyano mode as well as artificial modes in the 1800 – 2000 cm-1 range shows the limits of simplistic forcefields to reproduce good-accuracy AIMD and experiment, and signals the need to reparametrize selected bonded-interaction aspects of subsets of the forcefields. In this respect, force-matching to AIMD is an effective and promising strategy, whilst the use of tight-binding MD with good training sets for parameterization is also likely to be a very important future direction in simulation-enabled DSC-prototype design. Still, the present study has still underlined the effectiveness and importance of pragmatic usage of good-quality AIMD and DFT in guiding the tailoring of effective partial-charge sets for forcefield-based MD, as well as showing the limits of fitting partial charges alone. Nevertheless, using increasingly popular charge-scaling approaches, which yield charges of around ± 0.8 on the anions and cations, is a good direction for future potential work by the community in terms of the development of appropriate partial-charge sets for use with empirical force-fields.
The authors have nothing to disclose.
The authors thank Prof. David Coker for useful discussions and Science Foundation Ireland (SFI) for the provision of High-Performance Computing resources. This research has been supported by the SFI-NSFC bilateral funding scheme (grant number SFI/17/NSFC/5229), as well as the Programme for Research in Third Level Institutions (PRTLI) Cycle 5, co-funded by the European Regional Development Fund.
This was a molecular simulation, so no experimental equipment was used. | |||
The name of the software was DL-POLY (the 'Classic' version of which is available under GnuPublic Licence, via sourceforge) |