Discovery of blue singlet exciton fission molecules via a high-throughput virtual screening and experimental approach

Singlet exciton fission is a mechanism that could potentially enable solar cells to surpass the Shockley-Queisser efficiency limit by converting single high-energy photons into two lower-energy triplet excitons with minimal thermalization loss. The ability to make use of singlet exciton fission to enhance solar cell efficiencies has been limited, however, by the sparsity of singlet fission materials with triplet energies above the bandgaps of common semiconductors such as Si and GaAs. Here, we employ a high-throughput virtual screening procedure to discover new organic singlet exciton fission candidate materials with high-energy ( > 1.4 eV) triplet excitons. After exploring a search space of 4482 molecules and screening them using time-dependent density functional theory, we identify 88 novel singlet exciton fission candidate materials based on anthracene derivatives. Subsequent purification and characterization of several of these candidates yield two new singlet exciton fission materials: 9,10-dicyanoanthracene (DCA) and 9,10-dichlorooctafluoroanthracene (DCOFA), with triplet energies of 1.54 eV and 1.51 eV, respectively. These materials are readily available and low-cost, making them interesting candidates for exothermic singlet exciton fission sensitization of solar cells. However, formation of triplet excitons in DCA and DCOFA is found to occur via hot singlet exciton fission with excitation energies above ∼ 3.64 eV, and prominent excimer formation in the solid state will need to be overcome in order to make DCA and DCOFA viable candidates for use in a practical device.


INTRODUCTION Introduction to singlet exciton fission
Singlet exciton fission is a down-conversion process in organic semiconductors that spontaneously converts one spinsinglet electron-hole pair (exciton) into two spin-triplet excitons. 1 Each triplet exciton carries approximately half the energy of the initial singlet exciton. Conventional single-junction solar cells are limited in efficiency to about 34% (the Shockley-Queisser limit), largely due to loss from unabsorbed below-bandgap photons and thermalization of high-energy excitons. 2 When combined with a lower-bandgap semiconductor, singlet exciton fission materials raise the theoretical efficiency limit of a single-junction solar cell by reducing thermalization of excitons generated by high-energy photons. It has been calculated that the maximum power conversion efficiency of a single-junction photovoltaic device incorporating a layer of materials that can undergo singlet exciton fission is 44.4%. 3 On its own, singlet exciton fission yields no advantage to the power efficiency of solar cells because the potential increase in photocurrent is matched by a decrease in the open circuit voltage. 4 A benefit can be realized, however, if a singlet fission material is matched with a second material that absorbs low-energy photons. For example, in combination with silicon, a singlet exciton fission material ideally absorbs all photons with energies greater than twice the silicon bandgap. 5 The resulting excitons are split into two excitons at or just above the silicon bandgap and transferred to silicon, where they supplement silicon photocurrent generated from direct absorption of photons with energies between the silicon bandgap and twice the silicon bandgap. 6 Singlet exciton fission requires that the energy of the singlet exciton is approximately twice the energy of the triplet exciton. The exchange energy splitting between singlet and triplet excitons scales with the degree of overlap between the highest occupied molecular orbital and the lowest unoccupied molecular orbital. 7 As a result, most polyacene molecules exhibit singlet-triplet exchange energies of approximately 1.0-1.3 eV. 8 The ideal singlet exciton energy of a singlet fission material is therefore approximately 2.0-2.6 eV; above this singlet energy, fission is typically exothermic, while below this singlet energy, fission is typically endothermic. Indeed, the bestknown blue fission material, anthracene, is only capable of hot fission, an inefficient process where the dissociation of singlet excitons competes with internal conversion and vibrational relaxation. 9,10 Modifying anthracene by incorporating chemical side groups can perturb its singlet and triplet energies to make singlet exciton fission thermodynamically feasible directly from the S 1 singlet state. To date, however, such anthracene derivatives have reported triplet energies around 1.1-1.2 eV, providing minimal enthalpic driving force for sensitization of silicon solar cells. 11,12 The bluest, efficient class of singlet exciton fission materials is based on a core of tetracene, with a singlet exciton energy of approximately 2.4 eV. Fission of these excitons yields triplets with energies almost identical in energy to the silicon bandgap. 1,13 At these energies in the short wavelength infrared, molecules are typically weakly luminescent. Thus, coupling dark triplet excitons to photons for sensitization of photovoltaics requires the use of inorganic materials such as PbS nanocrystals, but with minimal energetic allowance for their Stokes shift or the spectral width of their photoluminescence. 2,[14][15][16] To improve the prospects of radiative coupling of triplet excitons to silicon, here we employ high-throughput virtual screening (HTVS) to search for blue exciton fission materials based on an anthracene core. We seek to find viable singlet exciton fission candidates with triplet exciton energies in excess of 1.4 eV. The triplet energy threshold of 1.4 eV is chosen because it should allow for exothermic transfer of the triplet excitons to silicon (with a bandgap of about 1.1 eV), even after accounting for uncertainty in calculated triplet energies.

Introduction to high-throughput virtual screening
High-throughput virtual screening (HTVS) combines quantum chemical calculations and cheminformatics methods to reduce a large molecular space to a set of promising leads that experimental chemists can then synthesize and characterize. [17][18][19][20][21][22][23][24][25][26] Functional organic molecules are particularly well-suited for high-throughput virtual screening, since the important properties of the materials can be approximated by studying individual molecules in the material, often with relatively computationally inexpensive methods, such as density functional theory (DFT). 18,19

Library generation
In order to limit our parameter space, we focus on known molecules with an anthracene core and a smaller fraction of combinatorically generated anthracene derivatives, in contrast to the large combinatorial fragment-based libraries that are often employed in high-throughput virtual screening projects for organic materials. 11,[27][28][29] The chemical space of this study consists of known and commercially available molecules. The libraries are obtained by searching eMolecules and Reaxys databases for molecules with an anthracene substructure. A total of 4482 candidates are examined.

Calculation method and benchmark
Before running full-scale calculations, we benchmark several methods to predict S 1 and T 1 values. In order to test the calculations, we use a dataset of 26 published molecules with experimentally determined S 1 and T 1 energies. Since a large contribution from a multireference character in the ground state is not expected, standard hybrid DFT methods are employed. The overall pipeline for calculations is outlined in Fig. 1(a). At the first stage, molecules encoded as Simplified Molecular-Input Line-Entry System (SMILES) strings from the generated library are fed into a conformer generator. The conformer generator samples conformers using a random distance matrix method, as implemented in RDKit. 30 The samples generated are optimized using the MMFF94 force field 31 and duplicates are eliminated. Next, the conformers are optimized using the DFT-B3 method, 32 which gives improved ground state geometries for pi-conjugated systems compared to the MMFF94 force field, and duplicates are again eliminated. The DFT calculations are performed using the OChem 4.0 package. 33 The DFT-B3 conformers are optimized at the B3LYP/6-31G(d) level of theory for both singlet and triplet ground state electron configurations. Finally, excited state calculations are performed on the S 0 geometry of the optimized conformers with three commonly used functionals: hybrid B3LYP, 34 range-separated ωB97X-D, 35 and range-separated ωLC-PBE0, 36,37 using the 6-31G(d) basis set.
For the calibration dataset, B3LYP/6-31G(d) outperforms both of the range-separated functionals with respect to predicting S 1 and T 1 energies (Figs. S1 and S2). We compare two approaches of calculating T 1 energies: vertical time-dependent DFT and a method where triplet energies are estimated by calculating the energy difference between the optimized structures at the ground state and the triplet excited state (calculated using openshell DFT). The latter performs better in predicting T 1 energies (Fig. S2). With these considerations in mind, we select the pipeline depicted in Fig. 1(a) for calculations of the full candidate material library.

Computational screening results
In this section, we describe our successive criteria for narrowing the library of 4482 molecules to the most promising blue-absorbing singlet exciton fission candidates for exothermic coupling to silicon solar cells. The first two criteria [depicted in Fig. 1(b)] are related to maximizing the energy level compatibility of the S 1 and T 1 states for singlet exciton fission. In Fig. S3(a), the distributions of calculated S 1 and T 1 energies are jointly plotted. The shaded region indicates the region where the differences between the S 1 energy and twice the T 1 energy are within 0.2 eV of each other, which we use as the cutoff for compatibility, given the observation that singlet exciton fission in tetracene is about 0.2 eV uphill. 38 About 20% of the library satisfies this criterion (929 molecules). The second criterion, which evaluates the difference between the T 2 and T 1 energies, is motivated by the desire to avoid molecules for which two T 1 excitons could upconvert into a T 2 exciton. The shaded region in Fig. 2(b) shows the 541 molecules for which T 2 > 2T 1 and for which the S 1 /T 1 criterion is satisfied.
Next, we screen for materials with T 1 energies above 1.4 eV, reducing the number of eligible candidates to 157 [ Fig. 1(c)]. Of the 384 molecules that are removed at this stage, the vast majority contain tetracene or anthraquinone substructures.
To maximize the probability that the molecules will perform well in an aggregate structure, we seek to minimize the variance in the conformer excited state energies. Screening for molecules where the range of S 1 and T 1 adiabatic excitation energies is less than 0.05 eV further reduces the number of eligible candidate materials to 116 (Figs. S4 and S5). As expected from chemical principles, most of the molecules excluded at this stage contain multiple rotatable bonds, often to another aromatic ring.
We next screen for molecules with S 1 energies predominantly in the blue but not in the UV, since such molecules are better suited to the solar spectrum and may be less susceptible to photoinduced chemical degradation. This corresponds to an S 0 → S 1 transition between 2.64 eV and 3.26 eV. Of the 116 molecules remaining before this criterion, all absorb in this range (Fig. S6). However, this criterion would exclude some molecules if a lower T 1 cutoff energy (such as 1.2 eV) is employed for compatibility to Si, and it could also be a useful screening criterion for larger libraries.
Finally, to narrow the focus of our study to true anthracene derivatives, we filter out the remaining molecules containing more of Chemical Physics than five fused rings, leaving us with 88 singlet exciton fission candidate materials, approximately 2.0% of our original library.

Experimental characterization of singlet exciton fission candidates
From the pool of 88 singlet exciton fission candidate materials, two are chosen for experimental characterization: 9,10dicyanoanthracene (DCA) and 9,10-dichlorooctafluoroanthracene (DCOFA). DCA and DCOFA are selected for this study because their calculated triplet energies are high (1.47 eV and 1.45 eV, respectively), making them promising candidates for exothermic triplet transfer and sensitization of silicon. Additionally, both DCA and DCOFA have known crystal structures reported in the Cambridge Crystallographic Data Centre and are readily available from commercial suppliers, having been used as synthetic precursors in prior experimental studies. [39][40][41] Furthermore, both materials are reported to have a single conformer, and therefore, issues related . Both materials exhibit absorbance spectra with a clear vibrational structure. From the onset of their bluest emission features, we estimate the S 1 energies of DCA and DCOFA in solution to be 2.99 eV and 3.10 eV, respectively, compared to the calculated values of 2.98 eV and 2.97 eV. Note that these singlet energies are likely to vary slightly depending on the specific solvent chosen. 42 To prepare for measuring the triplet energy of DCA, the material is dropcast from solution at 2 wt. % in poly(4-bromostyrene) (4BrPS), at a total concentration of ∼20 mg/ml in methoxybenzene. At 2 wt. %, it is not expected that DCA undergoes efficient singlet exciton fission. That said, the presence of bromine in the host material causes enhanced intersystem singlet-to-triplet crossing via spin-orbit coupling and the heavy-atom effect. 43 Moreover, as methoxybenzene dries, 4BrPS forms a rigid polymer matrix, which is expected to reduce nonradiative recombination from coupling to molecular vibronic modes. 44 The film is encapsulated in a nitrogen glovebox before being transferred to a cryostat and pumped down to ∼1 × 10 −7 Torr at a temperature of 77 K. To experimentally measure T 1 phosphorescence from DCA, we employ an optical gating method using out-of-phase choppers to mechanically filter out prompt emission and capture only delayed emission, which is expected to result from longer-lived triplet state emission [as well as delayed emission following subsequent triplettriplet (TT) fusion back to the singlet state]. The resulting spectrum (after subtraction of singlet emission, Fig. S7) is shown in Fig. 2(e).
To measure DCOFA phosphorescence, we use a room temperature phosphorescence method described by Reineke and Baldo. 44 DCOFA is doped in a film of PMMA and benzophenone (BP), a well-known triplet sensitizer with an intersystem crossing efficiency close to 100% at room temperature 45 and a triplet energy of 2.96 eV. 44 The film is excited at 270 nm, where BP absorbs strongly, and the resulting emission spectrum is compared to the spectrum when excited at 385 nm, where DCOFA is the dominant absorber. Subtracting the neat DCOFA singlet exciton spectrum from the spectrum of DCOFA doped in BP yields the phosphorescence spectrum shown in Fig. 2(f).
The triplet energies of DCA and DCOFA are estimated from the onset of their emission to be 1.54 eV and 1.51 eV, respectively (compared to the calculated triplet energies of 1.47 eV and 1.45 eV). The experimental triplet energies of DCA and DCOFA are in close agreement with the values calculated during material screening, as summarized in Table I.
To demonstrate that DCA and DCOFA undergo singlet exciton fission, we measure the magnetic field effect (MFE) on photoluminescence intensity. The theory of MFEs in organic molecular crystals was developed by Merrifield in 1968. 46 When a molecule undergoes singlet exciton fission, it forms a triplet-triplet pair state with an overall spin-singlet character. Singlet excitons can only effectively couple to the subpopulation of the nine possible triplet-triplet (TT) pair states with a singlet character. In the absence of a magnetic field, only three of the nine TT states have partial singlet character. As the field increases, the small splitting in triplet energies results in more of the TT states developing the singlet character, thus increasing the rate of singlet exciton fission (and resulting in an initially negative MFE on the singlet PL intensity). As the magnetic field is increased further, the number of TT states with singlet character decreases to two, consequently reducing the rate of singlet exciton fission (and resulting in a positive MFE on the singlet PL intensity at higher magnetic fields). The typical MFE signature of singlet exciton fission is therefore a negative MFE at low fields, switching to a positive MFE at higher fields, which plateaus as the magnetic field is increased and the number of TT states with singlet character approaches two. Figure 3 shows the MFE on singlet photoluminescence from purified DCA and DCOFA crystals under 340 nm light-emitting diode (LED) excitation. Both DCA and DCOFA exhibit a nonmonotomic, positive MFE with zero-MFE crossings at ∼0.1 T, characteristic of singlet exciton fission. Curiously, however, the MFE on photoluminescence in DCA and DCOFA is inverted when using excitation wavelengths longer than 340 nm and 365 nm, respectively (Fig. S8). To make sense of these differences, we suggest that the observed MFE shapes reflect changes in the origin of delayed fluorescence for varying pump wavelengths. A positive MFE, as is observed under 340 nm excitation, suggests that delayed fluorescence is due to the annihilation of correlated triplet excitons originally formed by singlet exciton fission, as described above. Meanwhile, the inverted and negative MFE observed at longer excitation wavelengths is indicative of delayed fluorescence due to the annihilation of initially uncorrelated triplet excitons (possibly formed through weak intersystem crossing). We speculate therefore that, as in neat anthracene, 9,10 singlet exciton fission in DCA and DCOFA is mediated by excited states above S 1 , while at longer excitation wavelengths, singlet exciton fission is unfavorable.
Indeed, the highest-energy emission peaks from DCA and DCOFA are red-shifted by around 100 nm in crystalline powder as compared to those in solution (Fig. S9). Based on this red-shift and their large spectral width, the prompt emission peaks in Fig. S9 likely contain contributions from emissive excimer states of DCA and DCOFA. From the rising edge of the crystalline DCA and DCOFA prompt emission peaks, we estimate the energies of these states to be 2.59 eV and 2.71 eV, respectively, making singlet exciton fission uphill by more than 300 meV in each case. The substantial endothermicity associated with splitting such states into two triplets further suggests that singlet exciton fission in crystalline DCA and DCOFA is mediated by excited states above S 1 . of Chemical Physics Hot singlet exciton fission has previously been reported in anthracene, tetracene, and several other organic crystalline semiconductors. 8,9,47 Hot fission in these materials has generally been attributed to coupling of the TT pair state to vibrationally excited states, 48,49 with recent studies emphasizing the importance of excited states with charge-transfer (CT) character. 50,51 Given the excitation wavelength dependence that we observe in the MFE of DCA and DCOFA, it is possible that singlet exciton fission in these materials is mediated by an excited CT state. Rational design of DCA and DCOFA dimers could be used to reduce excimer formation and make singlet exciton fission from the ground excited state more favorable. 52

CONCLUSION
In this study, we employ high-throughput virtual screening to reduce an initial set of 4482 singlet exciton fission candidate materials to a list of 88 candidates with calculated triplet energies in excess of 1.4 eV. The effectiveness of the virtual screening procedure is demonstrated via the discovery of two new singlet exciton fission materials: DCA and DCOFA. These materials are readily available, low-cost, and have high triplet energies (∼1.5 eV), making them promising candidates for exothermic singlet exciton fission sensitization of solar cells. Practical use of DCA and DCOFA as singlet fission materials is limited, however, by substantial excimer formation in the solid-state, requiring excitation wavelengths of ∼340 nm or lower in order to observe singlet exciton fission. Further tailoring of molecular coupling in DCA and DCOFA, for example through dimerization, may reduce excimer formation and make these materials more practical for use in a device. Integrating intermolecular coupling into virtual screening criteria is expected to result in higher hit rates for screened materials. This study shows that computational screening can lead to molecules that have necessary but not sufficient properties for function, highlighting one of the most important challenges and opportunities in the field of virtual high-throughput virtual screening.

Materials
The materials used have acronyms as follows. DCA: 9,10dicyanoanthracene, DCOFA: 9,10-dichlorooctafluoroanthracene, 4BrPS: poly(4-bromostyrene), MB: methoxybenzene, BP: benzophenone, and PMMA: poly(methyl 2-methylpropenoate). All materials except DCA and DCOFA were used as received without further purification. DCA (>99.7% sublimed grade) and DCOFA (>99% sublimed grade) were further purified via sublimation in a tube furnace. The experimental work in this manuscript relied on the material deposited near the center of the tube furnace, since this region had the lowest temperature gradient and was expected to yield the purest material. Further details on material purification are included in the supplementary material. Vendor information: DCA, DCOFA, 4BrPS, MB, chloroform, and cyclohexane: Sigma Aldrich; PMMA: Alfa Aesar.

Steady-state absorbance
Absorbance spectra in solution were collected using a UV-Vis absorbance spectrometer (Cary 5000, Agilent). Samples were prepared by dissolving purified DCA and DCOFA in cyclohexane and chloroform, respectively. All solution absorbance measurements were made using 1 cm path length quartz cuvettes.
Solid-state absorbance spectra of purified, crystalline powders of DCA and DCOFA were obtained by diffuse reflection (Cary 5000, Agilent). KBr was used as a reference, and the powders were diluted at 1 wt. % in KBr. The measured reflectance was converted to the Kubelka-Munk parameter, which is proportional to the absorption coefficient.

Singlet state photoluminescence
Photoluminescence spectra were collected using a spectrofluorometer (Fluoromax-3, Horiba). To minimize spectral distortion due to reabsorption, samples were diluted so that the optical density at their peak absorbance wavelengths is below 0.03 cm −1 . All spectra were collected using an excitation wavelength of 375 nm with entrance and exit slits set at a bandpass of 5 nm and 1 nm, respectively.
Triplet state phosphorescence DCA phosphorescence was measured using a thermoelectrically cooled silicon camera (PRO-EM-HS:512BX3, Princeton Instruments). Samples were prepared using 5 mm × 5 mm quartz substrates (MTI Corp). The substrates were cleaned via sequential sonication in detergent solution (Micro-90), deionized water, and acetone. They were then immersed in boiling isopropanol, dried with a nitrogen spray gun, and transferred to a nitrogen glovebox. Solutions of 4BrPS (20 mg/ml in MB) and DCA (1 mg/ml in MB) were prepared and subsequently combined to form a solution of DCA (2 wt. % in 4BrPS). Films of DCA/4BrPS were formed via dropcasting onto the quartz substrates at 70 ○ C and maintaining at 70 ○ C until dry. Samples were encapsulated in the glovebox using a UV curable epoxy (OG159-2, Epoxy Technology) and a second, 2-side polished quartz substrate (MTI Corp).
Samples were then loaded into a helium closed-cycle cryostat (Montana), pumped under vacuum to 1 × 10 −7 Torr, and cooled to 77 K. Phosphorescence spectra were obtained using two chopper wheels operated out-of-phase using phase-locked chopper controllers (MC2000B, Thorlabs) to time-gate sample excitation and emission collection, so as to block prompt emission and collect only the delayed portion of the photoluminescence. A 340 nm LED (M340L4, Thorlabs) was used for excitation, chopped at 270 Hz, defining approximately 2 ms optical gates. Sample emission was collimated, refocused into a monochromator (SP-2300, Princeton Instruments), and subsequently imaged on a thermoelectrically cooled silicon camera (PRO-EM-HS:512BX3, Princeton Instruments). Both singlet photoluminescence and triplet phosphorescence contributed to the recorded spectra. To remove the contribution from singlet emission, the blue portion of the emission spectrum (from 450 to 700 nm) was fit to the 77 K ungated singlet emission spectrum, and the singlet PL contribution was subtracted from the measured spectrum, yielding the phosphorescence spectrum shown in Fig. 2(c).
DCOFA phosphorescence was measured at room temperature according to a method adapted from Reineke et al. 44 A 20 mg/ml solution of DCOFA doped at 2 wt. % in [PMMA:BP] 3:1 was prepared and dropcast from methoxybenzene onto a cleaned 10 mm × 10 mm quartz substrate at 70 ○ C in a nitrogen glovebox. Once dry, the sample was encapsulated using a second quartz substrate and a UV curable epoxy. Phosphorescence was detected on a spectrofluorometer (Fluoromax-3, Horiba) by exciting the sample at 270 nm (where BP absorbs) and comparing against spectra obtained at 385 nm excitation (where only DCOFA absorbs). To remove the contribution from singlet emission, the blue portion of the emission spectra (from 450 to 700 nm) was fit and the singlet PL contribution was subtracted from the measured spectrum, yielding the phosphorescence spectrum shown in Fig. 2(d).

Magnetic field effect
Measurements of the magnetic field effect on DCA and DCOFA photoluminescence were performed according to the procedure described by Congreve et al. 4 A monochromatic 340 nm lightemitting diode (M340L4, Thorlabs) was used to excite the samples. Light from the diode was cleaned with a bandpass filter and mechanically chopped. While the sample was under illumination, an electromagnet was switched between positive and zero magnetic fields at a frequency of 33 mHz and a duty cycle of 50%. A 400 nm longpass filter was used to filter the scatter from the LED and ensure that only sample emission was detected, while a 750 nm shortpass filter was used to ensure that any phosphorescence that might contribute to the sample emission at room temperature was excluded. Photoluminescence from the sample was detected using a silicon photodetector (818-UV, Newport) connected to a lock-in amplifier (SR830, Stanford Research Systems). The magnetic field was monitored using a transverse gaussmeter probe (HMMT-6J04-VF, Lakeshore). The emission intensity and magnetic field were recorded at a frequency of 1 Hz.
For each data point in the plot of the magnetic field effect, the change in photoluminescence was calculated using the following steps: First, the photoluminescence was averaged over the full period in which the positive magnetic field was applied. Second, the PL was averaged over the full period in which the magnetic field was zero. The amplitude of the magnetic field effect is then given by the relative change in signal percentage, MFE = 100% * (PL B − PL 0 )/PL 0 , where PL B and PL 0 are the averaged emission intensities with and without applied magnetic field.

Powder x-ray diffraction
Powder x-ray diffraction PXRD patterns were measured on purified, crystalline powders of DCA and DCOFA using a diffractometer (Advance II, Bruker) equipped with θ/2θ Bragg-Brentano geometry and Ni-filtered Cu Kα radiation (Kα 1 = 1.5406 Å, Kα 2 = 1.5444 Å, Kα 1 /Kα 2 = 0.5). The tube voltage and current were set to 40 kV and 40 mA, respectively. Samples were prepared as a thin layer of powder on a zero-background silicon crystal plate. The angle was scanned from 2θ = 3 ○ -50 ○ in increments of 0.02 ○ using a slit size of 1.0 mm and a scan rate of 1 s/step and resulting PXRD patterns compared to calculated patterns from the data reported in the Cambridge Crystallographic Data Centre.

SUPPLEMENTARY MATERIAL
See the supplementary material for additional figures related to the materials screening procedure and experimental characterization. A full list of singlet exciton fission calibration data and candidate materials considered for this study is included in the associated spreadsheet files uploaded with this manuscript.