Photoredox Chemistry with Organic Catalysts: Role of Computational Methods.

Organic catalysts have the potential to carry out a wide range of otherwise thermally inaccessible reactions via photoredox routes. Early demonstrated successes of organic photoredox catalysts include one-electron CO2 reduction and H2 generation via water splitting. Photoredox systems are challenging to study and design owing to the sheer number and diversity of phenomena involved, including light absorption, emission, intersystem crossing, partial or complete charge transfer, and bond breaking or formation. Designing a viable photoredox route therefore requires consideration of a host of factors such as absorption wavelength, solvent, choice of electron donor or acceptor, and so on. Quantum chemistry methods can play a critical role in demystifying photoredox phenomena. Using one-electron CO2 reduction with phenylene-based chromophores as an illustrative example, this perspective highlights recent developments in quantum chemistry that can advance our understanding of photoredox processes and proposes a way forward for driving the design and discovery of organic catalysts.

Photoredox Chemistry: Metal-Free catalysts, Challenges, and Opportunities

Photoredox catalysts are capable of performing challenging oxidation and reduction reactions by harnessing the power of ultraviolet or visible light.[1−10] When photon absorption promotes an electron to its excited state, the catalyst undergoes reduction or oxidation via electron transfer (ET), either directly by substrates or indirectly by means of a sacrificial electron donor or acceptor followed by substrate attack. The final ET step serves to regenerate the catalyst to its original ground state. These catalytic cycles can also be paired with other catalytic cycles or processes. Figure by Romero and Nicewicz illustrates the steps constituting a photoredox catalytic cycle.[1]

Figure 1

Oxidative and reductive quenching cycles of a photoredox catalyst. Reprinted (adapted) with permission from ref (1) Copyright 2016 American Chemical Society.

As illustrated in Figure , organic photoredox catalysts span a wide range of conjugated molecules such as dyes, quinones, pyryliums, and phenylene derivatives.[1,11−13] Organic catalysts have immense untapped potential owing to the diversity in structures of conjugated molecules as well as high tunability achievable in activity and stability with even small structural variations.[14,15] In addition, they promise greener, metal-free alternatives to more widely studied inorganic photoredox catalysts, consisting mainly of ruthenium or iridium transition-metal centers coordinated to conjugated ligands.[16−20]

Figure 2

Chemical structures of common organic photocatalysts. Reprinted (adapted) with permission from ref (13) Copyright 2018 Wiley.

These catalysts are capable of carrying out transformations that are otherwise inaccessible via thermal activation. One such example is the utilization of anthropogenic CO2 as C1 feedstock, by reducing the molecule to make solar fuels and chemicals. Most homogeneous photoredox pathways for CO2 reduction employ inorganic catalysts with organic or inorganic chromophores that serve as photosensitizers.[21] A photosensitizer such as phenazine transfers an electron to an inorganic electron mediator upon excitation and quenching, which then proceeds to reduce CO2.[22]

By themselves, organic chromophores such as terphenyls become powerful reductants upon photoexcitation (>290 nm) and quenching with a sacrificial electron donor (e.g., triethylamine, TEA). One-electron reduction of CO2 with short-chain oligo(p-phenylene) (OPP), with activity higher than its ortho- and meta-isomers, was demonstrated in the 1990s by Mastuoka and co-workers.[14,23,24] These studies inspired more recent developments in flow reactors that utilize OPP to catalyze amino acid synthesis and styrene hydrocarboxylation via photoredox CO2 reduction.[25,26] Authors of this perspective carried out the first computational examination of ET kinetics from OPP•– to CO2, and quantified the sensitivity of ET rates to variation in electrophilicity of substituents coordinated to the catalyst.[15] The proposed catalytic cycle for CO2 reduction with OPP is shown in Figure . OPP also sensitizes H2 formation via water splitting[27] in the presence of RuCl3 in acetonitrile solvent.[28] Aqueous suspensions of linear phenylene polymers and their nitrogen-containing derivatives have been shown to catalyze photoredox water splitting in the presence of amine donors.[29−31]

Figure 3

Catalytic cycle of one-electron reduction of CO2 with OPP,[24] where n represents the number of interior phenylenes (i.e., n = 1 corresponds to OPP-3) and R represents terminal, para substitutions (e.g., −OH, −CN, −F). Reprinted (adapted) with permission from ref (15) Copyright 2020 American Chemical Society.

Organic photoredox catalysts therefore have the potential to unlock a rich array of chemical transformations via the generation of highly reactive radical species under relatively mild conditions. Despite lower costs and toxicity compared to metal-based photocatalysts, organic photoredox cycles and their electron-transfer (ET) mechanisms are poorly understood. This stems from the fact that photoredox cycles typically involve ground- and excited-state potential energy surfaces, generation of several radical intermediates, non-negligible solvation effects, and varied product distributions. Systematic computational studies driven by electronic structure methods are therefore essential to advance our understanding of photoredox mechanisms and interactions between chromophore, solvent, substrate, and donor/acceptor species.

Beyond the inherent intricacies of photoredox catalytic cycles, there are fundamental challenges to the design and development of practically viable organic catalysts. One of the biggest hurdles facing organic photoredox catalysts, which is also poorly understood, is the limited turnover numbers achievable with these catalysts. While most catalytic materials exhibit reduced activity over time as a result of poisoning or degradation, organic catalysts degrade exceptionally quickly via dearomatization reactions. For example, various isomers of oligophenylenes that catalyze CO2 reduction exhibit single-digit turnover numbers.[14] An experimental study of CO2 reduction with of o-, m-, and p-terphenyl observed very little formation of the product formic acid with o- and m-isomers despite higher reduction potentials compared to the p-isomer. This study attributed the low activity of these two isomers to their preference for carboxylation by CO2. It was proposed that photo-Birch reduction[32] via protonation of the phenylene radical anion is another reason for poor turnover numbers across all three isomers.[14]

Mechanisms of carboxylation and Birch reduction as well as the relative propensities of various chromophores to undergo degradation chemistry have yet to be examined through the lens of the viability of organic photoredox catalysts. However, these reactions have been extensively examined in the context of organic synthesis.[16,33−35] For instance, studies show that relative charge densities at constituent carbon centers determine the site of first protonation or regioselectivity of Birch reduction.[36,37] Therefore, one can leverage domain knowledge in organic synthesis to design stable catalysts by identifying characteristics that render chromophores less susceptible to such degradation pathways.

This perspective highlights the utility of quantum chemistry methods in developing an understanding of key steps constituting the photoredox cycle. For this purpose, one-electron reduction of CO2 with phenylene oligomers (Figure ) is used as the illustrative photoredox cycle. In addition to uncovering mechanisms underlying photoredox processes, computational studies offer means to systematically tune properties of chromophores, solvents, and sacrificial donor/acceptor species to identify features that are critical to the design of viable catalytic systems. The ability of single-reference quantum chemistry methods to reliably predict energetics and interactions between reactive species have led to significant progress in large-scale, computationally driven screening and design of heterogeneous inorganic catalysts at gas–solid and electrocatalytic interfaces for a wide range of reactions.[38,39] Appreciable progress has also been made in descriptor-driven screening of organic chromophores for applications in photovoltaic materials.[40] The authors of this perspective therefore envision a promising future for the design and development of organic photoredox catalysts to perform otherwise thermally challenging chemical transformations. Catalyst design will be driven by a combination of advances in electronic structure methods for the treatment of condensed phases and excited states, mechanistic studies of degradation pathways inspired from organic synthesis, descriptor-driven screening approaches, and machine learning methods to automate the search of chemical space and drive discovery of novel chromophores.

Role of Computational Chemistry: Overview

Quantum chemistry methods have made significant advances in their ability to accurately model condensed phases, excited states, and charge-transfer processes.[41] Yet, surprisingly few applications are to be found in photoredox chemistry and associated catalyst design. Electronic structure methods yield a detailed understanding of ground- and excited-state potential energy surfaces, photophysical characteristics, mechanisms of charge transfer, and interfragment interactions that govern key steps of the catalytic cycle. This perspective aims to serve as a practitioner’s guide and presents state-of-the-art computational approaches and model approximations to examine photoredox cycles. Figure outlines key processes and associated methods that are described in subsequent sections.

Figure 4

Proposed photophysical (blue), charge transfer (green), and degradation via bond-breaking/formation (red) processes in photoredox CO2 reduction with OPP-3 as the chromophore and TEA (Et3N) as the sacrificial electron donor. Methods for characterizing interfragment interactions at various stages of the photoredox cycle are also described (gray box). Reprinted (adapted) with permission from ref (15) Copyright 2020 American Chemical Society.

Owing to ongoing efforts by the authors of this perspective toward uncovering design rules for photoredox CO2 reduction with phenylene-based organic chromophores, this chemistry is utilized to illustrate computational approaches. Emphasis is on single-reference methods despite known limitations in their treatment of charge-transfer states and distorted geometries that possess multireference character. This is because single-reference, semiempirical electronic structure methods, such as density functional theory (DFT) for ground states and time-dependent density functional theory (TDDFT) for excited states, yield the most favorable cost-accuracy trade-off. This makes them both computationally tractable as well as reliable for mechanistic and large-scale catalyst screening and design studies. It must be noted that the methods described in this perspective do not constitute an exhaustive list. The methods are chosen in part because they are either compatible with or commercially available/under development in a single ab initio software package, Q-Chem.[42]

Photophysical Processes

Photophysical processes (shown in blue in Figure ) including absorption, emission, and radiationless events have been widely studied for organic chromophores owing to their potential applications as materials for photovoltaics and light-emitting diodes.[43−49] TDDFT is the most widely used technique for identifying participating electronic excited states, calculating vertical absorption and emission spectra, and geometry optimization and vibrational analysis on excited-state potential energy surfaces.[50,51] TDDFT benchmarking studies show that pure functionals are poor predictors of excited-state properties and recommend the use of either global or range-separated hybrid density functional approximations.[52]

Implicit models for solvation, despite accuracy and parameter sensitivity concerns,[53] are more commonly used than explicit inclusion of solvent molecules owing to favorable computational costs and effort. To capture the solvent response to vertical excitation via polarization of its electron cloud, the dielectric constant corresponding to infinite frequency (ϵ∞) is required in TDDFT calculations. Additional use of nonequilibrium solvation models that separate “fast” (electronic) from “slow” (nuclear) response of the solvent environment may also be necessary to accurately determine solvatochromic shifts.[54−58]

It is possible that multiple excited states participate in the photoredox cycle. For instance, both excited state singlet (S1) and triplet (T1) states are quenched by electron donors to form reducing anion radicals in phenylene-catalyzed H2 evolution as observed in transient absorption spectroscopy studies.[28] In such cases, the rate of dynamical intersystem crossing (ISC) from S1 to T1 determines the dominant state that undergoes quenching and consequently the rate of quenching. ISC is more likely if spin-orbit coupling is large, which can also be calculated using TDDFT. Other important features of excited-state potential energy surfaces include conical intersections, or crossing between potential energy surfaces, which can lead to radiationless decay of the excited state. These are typically identified via trajectory generation using nonadiabatic excited-state molecular dynamics methods,[59−61] which have been employed to probe radiationless events and long-lived triplet states in organic chromophores.[62,63] Photochemistry applications of nonadiabatic excited-state molecular dynamics are rare,[64] mainly due to high computational costs. Recent efforts in this field therefore focus on rendering these dynamics approaches more tractable for practical systems. One such framework, PySurf, includes on-the-fly construction of databases that accelerate the generation of potential energy surfaces.[65] These databases use fitting algorithms to interpolate properties in regions that are already evaluated, and additional quantum chemistry calculations are carried out only when the trajectory leaves the region where these fits are accurate. These advances are expected to significantly reduce the cost barrier to studying excited-state potential energy surfaces and the identification of chromophore features that impact radiationless processes and consequently quantum yields.

Charge-Transfer Kinetics and Exciplexes

As illustrated in Figure , a chromophore in its excited state can undergo a series of charge-transfer steps. In the case of OPP-driven CO2 reduction, the excited-state chromophore is quenched via complete charge-transfer (or electron transfer, ET) by the sacrificial electron donor, TEA. Incomplete charge-transfer can also occur leading to the formation of an excited-state complex or exciplex, described later in this section. Upon quenching via ET, the anion radical OPP•– by virtue of being highly reducing, transfers the electron to CO2 and is regenerated to its original ground state.

Figure depicts key states involved in the ET step from OPP•– to CO2. The calculation of ET rate coefficients (in the absence of bond formation) is carried out either using Marcus theory if ET is nonadiabatic, also known as the limit of weak coupling[66]or the following expression if ET is adiabatic (when states are strongly coupled), in which the pre-exponential term in eq is replaced by ν, a weighted average of vibrational frequencies for modes that contribute to the reaction coordinate (≈1013s–1)[67]In eqs and 2 above, kB is the Boltzmann’s constant, R is its molar equivalent, T represents temperature, ℏ is the reduced Planck’s constant, λ is reorganization energy, J is the charge coupling constant (also known as charge-transfer integral or diabatic coupling), and ΔG*is the effective barrier to ET given byHere, ΔG is the free energy driving force of the reaction or the difference between final state (FS) and initial state (IS) free energies. ΔG, λ, and J are also illustrated in Figure .

Figure 5

Qualitative energy vs reaction coordinate representation for an ET step, shown here for ET between OPP•– and CO2. The pink curve represents the adiabatic limit arising from the strong coupling. The black curve reflects the charge distribution of the initial state (IS, reactant complex) and blue represents that of the final state (FS, product complex). ΔG is the free energy difference between the two states and λ corresponds to the energy of reorganization of the solute and solvent. Reprinted (adapted) with permission from ref (15) Copyright 2020 American Chemical Society.

Initial and final diabatic states can consist of excited states and/or ionic radical species. While TDDFT can be employed to treat the former, ground-state ionic radical species are not well described by DFT. If the IS and FS consist of interacting ionic radicals as is the case in Figure , their charge separation must be described correctly to determine their structures and free energy difference, ΔG. Constrained density functional theory (CDFT) was developed to overcome the inability of conventional DFT to impose user-defined charge and spin constraints on fragments constituting a system.[68−70] As the imposition of constraints can lead to convergence difficulties, only the minimum possible set of constraints must be employed in CDFT calculations. Furthermore, as constraints specify excess charge on fragments rather than absolute charges, the charge distribution resulting from CDFT must always be verified using tools such as Becke population analysis.[71] In a prior study by the authors of this perspective, the IS and FS geometries depicted in Figure were optimized using CDFT to determine ΔG and its sensitivity to substituent variations to OPP.[15]

Figure also depicts the reorganization energy λ, which corresponds to the difference between the vertically excited charge-transfer state (VCT) and the IS. As ET is significantly faster than nuclear relaxation, the VCT state represents a nuclear configuration of the solvent + solute system that is out of equilibrium with electron density. Therefore, the calculation of λ requires the use of nonequilibrium solvation models.[54−57] It must be noted that ΔG and λ are approximated calculated as differences between electronic energies rather than free energies. This is because the calculation of λ involves a nonequilibrium state. As this state is not an energy minimum, it likely possesses one or more imaginary vibrational frequencies, making it difficult to accurately determine vibrational contributions to free energies. The use of an implicit solvation model, however, ensures that enthalpic and entropic contributions of the environment are captured via the free energy of solvation.

The magnitude of the coupling coefficient, J, determines whether ET is in the adiabatic or Marcus regime. Several schemes are available to construct diabatic states and calculate coupling between them, such as the generalized Mulliken–Hush (GMH)[72] or the fragment charge difference (FCD),[73] both of which require vertical excitation calculations using methods such as TDDFT. The study described in Figure employed a CDFT-based diabatization scheme known as CDFT configuration interaction (CDFT-CI) to calculate ground-state couplings between IS and FS.[74] However, diabatic couplings with CDFT are highly sensitive to level of theory, and a recent study demonstrated that diabatic states generated using absolutely localized molecular orbitals (ALMO) instead of CDFT yield accurate couplings and more systematic errors across functional classes.[75,76]

Incomplete charge transfer between an excited- and ground-state species occurs often in photoredox systems, leading to the formation of excited-state charge-transfer complexes or exciplexes.[77] These encounter complexes are known as excimers if formed between interacting dimers, commonly found in arene systems. The donor–acceptor pair in an exciplex shares both charge and electronic excitation. The exciplex is characterized by a broad red-shifted emission peak relative to that of the isolated excited state. These peaks have been observed for, among others, interacting arene and amine systems including the CO2 cycle described in Figure .[78,79]

Despite challenges associated with the identification of exciplex geometries on largely shallow, excited-state potential energy surfaces, only a handful of prior studies establish procedures for TDDFT-driven geometry optimization and characterization.[80,81] In a more recent study of exciplexes formed between excited-state OPP*and TEA, authors of this perspective outlined the detailed procedure for TDDFT optimization of exciplexes.[79] As the two fragments exhibit fractional charges in the exciplex state that cannot be determined a priori or specified as constraints, TDDFT and not CDFT was employed in this study. To minimize errors arising from known limitations of TDDFT in describing excited and charge-transfer states, the study calculated exciplex geometries using both hybrid as well as long range-corrected hybrid density functional approximations.[82] TDDFT optimization typically follows a predefined excited state supplied to the algorithm by the user as a numerical index. The index of the excited state that corresponds to the donor–acceptor charge transfer (or exciplex) is likely to change in the course of optimization. When this occurs, TDDFT optimization continues to follow the same index, which now corresponds to an incorrect mode and consequently yields a geometry that is not an exciplex. To overcome such practical difficulties, the TDDFT study recommends multiple short optimization runs in place of a single longer run along with user intervention to verify that the correct mode is always followed. In addition, very small step sizes and tight tolerances for convergence of gradients are recommended for the TDDFT optimization of exciplexes.[79]

The same study also outlined several methods for verifying whether TDDFT optimization converges correctly to an exciplex state in addition to validation with experimental fluorescence spectra and solvatochromic shifts.[79] A simple test of convergence to an exciplex is to replace the donor atom containing the lone pair with a group that does not contain a lone pair (e.g., N in TEA with a CH group). If the exciplex geometry is correctly identified, this replacement will result in the absence of the red-shifted emission peak. Other tests include visualization of natural transition orbitals[83] and energy decomposition in the excited state (described in a later section).[81]

Exciplexes are not always examined in photoredox studies owing to the fact that they often result in the same end product as full ET, in other words, the charge-separated anion–cation radical pair.[1] However, the proximity of the donor–acceptor pair in an exciplex may provide favorable conditions for proton transfer and subsequent photo-Birch reduction. Therefore, identifying exciplex structures and characterizing donor–acceptor interactions in these complexes constitute steps toward understanding degradation mechanisms of chromophores.

Catalyst Degradation: Bond Breaking and Formation

A critical hurdle that limits practical applications of phenylenes and other organic chromophores is small turnover numbers.[14,23,84] Reasonable guesses regarding degradation pathways emerge from experimental product distributions, as is illustrated using red arrows in Figure . In the case of OPP-driven CO2 reduction, photo-Birch reduction via protonation of the chromophore is proposed.[14] This is because the formation of H2 is observed even in aprotic solvents. The most probable source of protons is amine rather than the aprotic solvent. In addition, protonation of the chromophore itself is likely facilitated by the formation of an exciplex described in the previous section. The kinetics of protonation from the exciplex state, however, have yet to be examined. Degradation via photocarboxylation with CO2 is observed for, in addition to ortho- and meta- isomers of phenylenes, phenanthrene, anthracene, and pyrene chromophores.[14,85]

The computational toolkit routinely employed to examine mechanisms of thermally activated bond breaking or formation reactions is readily applicable here to assess the relative importance of various degradation pathways. The DFT toolkit consists of one of several methods available for single- or double-ended search for transition structures, mode-following methods for transition structure refinement, and verification via vibrational analysis and intrinsic reaction coordinate calculations.[86] Once the transition structures and intermediates constituting a degradation pathway are identified, standard thermochemical corrections such as the rigid-rotor harmonic oscillator approximation can be employed to compute reaction free energy profiles and rate coefficients.[87] Calculation of pathways and their sensitivity to chromophore variations—isomeric form, substitutions, oligomer chain length, and so on—are expected to provide valuable insights into preferred mechanisms and rates of chromophore degradation. These insights can then be leveraged to identify catalyst features that enhance stability by lowering susceptibility to degradation.

Characterization of Chromophore Interactions

Energy decomposition analysis (EDA) is one of the most widely used methods for breaking down electronic structure into physically meaningful descriptions of interactions between the fragments constituting a system.[88,89] In a stepwise, constrained fashion, EDA decomposes the interaction energy between fragments into fragment preparation energies, electrostatics, Pauli repulsions, dispersion, polarization (or intrafragment relaxation), and charge-transfer interactions. To avoid confusion between the last EDA term and charge-transfer phenomena in photoredox chemistry, this perspective refers to the EDA charge-transfer term as “interfragment relaxation”. While many formalisms of EDA are available, each differing in the definitions and procedures for calculation of intermediate steps, this perspective focuses on absolutely localized molecular orbital EDA or ALMO-EDA.[90,91] This is because ALMO-EDA and its variants described below are ideally suited to study interactions underlying photoredox chemistry in the condensed phase. For the cycle described in Figure , for example, these techniques are applicable to examining interactions between OPP and a substituent R, OPP*and Et3N, OPP•– and CO2, and OPP and CO2•–.

ALMO-EDA was originally developed to describe interactions between fragments in the ground state and in the gas phase. ALMO-EDA(solv), developed more recently, calculates all contributions to the total interaction energy like gas-phase EDA but in the presence of a solvent that is described by an implicit solvation model.[92] When contrasted with gas-phase EDA results from a prior study of ET to CO2 from OPP•– that is bound to para-substituents of varying electrophilicity,[15] this work found that both in the presence and absence of a solvent, electrostatic terms vary the strongest with electrophilicity, and are dominant contributors to overall interaction energy.[92] This study also found that the choice of functional strongly affects EDA outcomes and delocalization errors arising from functionals like B3LYP[93] may lead to higher contributions from interfragment relaxation.[94]

The original computational study of ET to CO2 from substituted OPP•– also outlined EDA procedures for CDFT outcomes when VCT states are involved, say, when analyzing contributions to reorganization energy λ.[15] As the final EDA step that calculates interfragment relaxation releases all CDFT constraints, the VCT state becomes identical to the IS (Figure ). To overcome this issue, interfragment relaxation energies can be calculated in an approximate manner using the maximum overlap method (MOM). When applied to ALMO guess orbitals, MOM results in solutions that are closer to charge-separated VCT states.[95,96]

A variational formalism of ALMO-EDA to probe frontier orbitals in conjugated systems[97] inspired a deeper analysis of observed trends in ΔG in the computational study of ET from OPP•– to CO2.[15] The study attempted to understand why the ΔG of ET varies mostly linearly with decreasing electrophilicity (quantified by Hammett parameter σ),[98] but plateaus as σ becomes more negative, or substituents become more electron donating. As ΔG variations mirror those in LUMO levels of ground-state-substituted OPPs, only the latter are shown in Figure .[15] The linear region arises from the fact that electron-withdrawing groups make it easier to accept an electron by lowering LUMO. The flattening in the electron-donating region is similar to observations by Mao and co-workers, who utilized variational EDA to examine interactions between naphthalene and its substituent.[97] They found that mixing between the virtual orbitals of substituent and parent leads to significant lowering of LUMO levels for electron-withdrawing groups. On the other hand, mixing between virtual orbitals is very small for electron-donating groups and therefore has little impact on LUMO levels.

Figure 6

Highest occupied molecular orbital (HOMO, red) and lowest unoccupied molecular orbital (LUMO, blue) energies (eV) vs σ for substituted OPPs. Inset: canonical molecular HOMO and LUMO for unsubstituted OPP. Trendlines, developed using the modified Akiake information criterion,[99] reflect quadratic dependence on σ (RHOMO2 = 0.94, RLUMO2 = 0.90). A qualitative depiction, pending future analysis with variational EDA, of regions of varying extents of orbital mixing (OOM: occupied–occupied mixing; VVM: virtual–virtual mixing) is also shown. Reprinted (adapted) with permission from ref (15) Copyright 2020 American Chemical Society.

An EDA formalism was also developed for examining gas-phase interactions between fragments constituting excited-state complexes.[81] This technique decomposes the change in excitation energy resulting from the interaction between the excited-state and ground-state fragment into frozen (electrostatic + Pauli repulsion), polarization, and charge-transfer (interfragment relaxation) terms. By applying this technique to the TDDFT-optimized exciplex structure between OPP* and TEA, the authors of this perspective successfully demonstrated that exciplex stabilization is dominated by interfragment relaxations arising from incomplete charge transfer from TEA to OPP*.[79] EDA therefore offers a suite of powerful characterization methods to not only deconvolute interactions between chromophores, substrates, and electron donors/acceptors but also generate quantitative insights into the role of solvent in condensed-phase photoredox chemistry.

Catalyst Screening, Design, and Discovery

With enhanced computing speeds and increasing synergies between quantum chemistry and machine learning methods, automated approaches for catalyst screening and design are becoming routine for thermochemical and electrochemical processes.[100−102] It is possible to develop similar approaches in photoredox chemistry with sufficient insights into the underlying processes and mechanisms and the formulation of design rules that enable screening and selection of active as well as stable chromophores, solvents, and donor/acceptor molecules. Thus far, most experimental and computational screening and design efforts for organic chromophores have focused on utilizing photophysical properties to identify viable photovoltaic materials.[103] An elegant illustration of DFT screening of organic chromophores to guide experiment is a study of nitrogen-substituted pentacenes by Hall and co-workers.[103] Using multiple attributes benchmarked to C60—HOMO and LUMO energies, HOMO-LUMO gaps, dipole moments, and reorganization energies—the study screened over a hundred possible chromophores arising from N-substitutions and identified the most viable targets for synthesis and further development of organic photovoltaic materials.

Such screening studies of organic chromophores are rare for applications in photoredox chemistry. Exceptions include a combined experimental and computational study by Sprick and co-workers that explored N-containing poly(phenylene) chromophores for photoredox water splitting to evolve H2.[31] They utilized quantum chemistry methods to calculate thermodynamic driving forces, optical gaps, and solvation energies of several oligomeric model systems and identified a trade-off between driving force for sacrificial donor oxidation (enhanced by electron-poor monomers) and excitation in the visible spectrum (induced by electron-rich monomers).

Such inherent trade-offs between catalyst performance metrics can be identified across nearly all classes of catalytic reactions. One such example is the Sabatier Principle in thermally activated reactions on the surfaces of solid catalysts, which states that both very strong and very weak binding of reaction intermediates are detrimental to catalytic activity.[104] Optimal conditions are located somewhere in between, and this trade-off is routinely leveraged in surface catalysis to identify quantitative (DFT-based) descriptors for catalyst performance and guide screening and selection of viable catalysts. Despite the clear need for identification of these trade-offs in photoredox chemistry demonstrated by the water-splitting study,[31] guiding principles are yet to be fully established. Rigorous characterization of such trade-offs will enable selection of only those chromophores that are viable in practice—in other words, demonstrate acceptable activity and turnover numbers (or stability) and are excited in or close to the visible region of the electromagnetic spectrum.

Based on the design rules or descriptors that characterize catalyst performance, quantum chemistry-driven screening can be combined with machine learning methods to accelerate the search for viable materials. An early example of chromophore screening for photovoltaic applications is the Harvard Clean Energy Project (CEP),[40] with the CEP database utilized more recently to train deep neural networks to reliably predict power conversion efficiencies of organic chromophores.[105] Fewer applications of machine learning-driven search for photoredox catalysts are available. Notable studies that highlight the challenges associated with identifying structure–performance correlations include combined experimental and computational studies of H2 evolution with conjugated organic molecules.[106,107] These studies find that no single quantum chemical descriptor adequately represents catalytic activity and that the underlying descriptor-performance relationships are most likely nonlinear. Furthermore, purely structure-driven machine learning models exhibit surprisingly comparable performance to expensive quantum chemical descriptor-driven models in predicting experimental hydrogen evolution rates.[107]

Armed with the knowledge of key factors that govern catalyst viability, their underlying trade-offs, and a set of candidates that are demonstrated to be active for a given reaction, it is possible to leverage recent advances in organic discovery methods to automate the search for novel chromophores with desired performance characteristics. Genetic algorithms (GAs) are global optimization methods inspired from natural evolution and the survival-of-the-fittest paradigm.[108] They utilize crossover and (low-probability) mutation operations to create off-springs from parents and select the fittest ones to create the subsequent generation. Prior chromophore studies have successfully employed GAs to identify supramolecular assemblies with desired properties for optical electronics applications[109] and simultaneous optimization of emissive, electron transport, and hole transport layers in organic light-emitting devices.[110] GA performance and diversity can also be enhanced with neural networks. For example, a neural network can diversify GA search by recognizing molecules that repeat across multiple generations and assigning lower scores to reduce their fitnesses.[111] For photoredox catalysts, the fitness and scoring functions can incorporate multiple descriptors (with appropriate weighting) such as absorption wavelength, redox potential, and susceptibility to degradation. Key criteria for the selection of descriptors include appropriateness and ease of calculation with routine DFT simulations. For instance, charge densities at carbon centers, obtained inexpensively from Mulliken charge analysis, may serve as descriptors for susceptibility to Birch reduction.[36,37] Alternatively, advanced reinforcement learning methods can be explored, specifically variational autoencoders (VAE),[112] which have been shown to explore chemical space of organic compounds based on characteristic features similar to training data.[113] Unlike traditional deep-learning methods, VAEs offer the advantage of being amenable to smaller data sets owing to their underlying Bayesian framework. The development of structured screening and design strategies in other domains of catalysis and photovoltaics can therefore serve as foundations on which we construct rational screening and selection procedures for organic photoredox catalysts.

Outlook

Organic photoredox catalysts have the potential to unlock light-driven, green pathways for chemical reactions that are otherwise thermally challenging. The design and development of photoredox systems that can catalyze CO2 reduction and subsequent conversion to fuels and chemicals under mild conditions, for example, will constitute a gigantic step forward in combating CO2-induced climate change. Photoredox processes span multiple time and length scales involving ground state, excited state, and ionic and radical species, all of which are strongly influenced by the solvent environment. Quantum chemistry methods can not only deconvolute these complex interactions and address gaps in the fundamental description of photoredox catalytic cycles but also accelerate the process of rational chromophore screening and selection to guide experiments. Although single-reference DFT and TDDFT methods yield the most favorable cost-accuracy compromise in studying these systems, the choice of functional approximation plays a critical role. In general, hybrid and range-separated hybrid functionals yield acceptable accuracies for excited state, charge transfer, and barriers for bond-breaking/making processes. In addition to its routine application to photophysical processes, TDDFT optimization can be employed to identify charge-transfer states (excimers or exciplexes) on excited-state potential energy surfaces. To examine ET between ground states, CDFT is a useful approach that imposes exact spin and integer charge constraints on fragments. EDA is a powerful tool for probing the physical origins of interactions between fragments in various steps of the photoredox cycle as well as the influence of solvent dielectric on these interactions. Insights into photoredox catalytic cycles generated using these state-of-the-art methods will aid in the development of structure–performance relationships and the identification of any underlying trade-offs between performance metrics. Going forward, these design rules are expected to drive rational screening and discovery of novel, viable organic photoredox catalysts, along similar lines to current progress in quantum chemistry- and machine learning-aided design of organic photovoltaic materials and thermal catalysts.