Prediction of Redox Power for Photocatalysts: Synergistic Combination of DFT and Machine Learning

The accurate prediction of excited state properties isa key elementof rational photocatalyst design. This involves the prediction ofground and excited state redox potentials, for which an accurate descriptionof electronic structures is needed. Even with highly sophisticatedcomputational approaches, however, a number of difficulties arisefrom the complexity of excited state redox potentials, as they requirethe calculation of the corresponding ground state redox potentialsand the estimation of the 0-0 transition energies (E (0,0)). In this study, we have systematicallyevaluated the performance of DFT methods for these quantities on aset of 37 organic photocatalysts representing 9 different chromophorescaffolds. We have found that the ground state redox potentials canbe predicted with reasonable accuracy that can be further improvedby rationally minimizing the systematic underestimations. The challengingpart is to obtain E (0,0), as calculatingit directly is highly demanding and its accuracy depends stronglyon the DFT functional employed. We have found that approximating E (0,0) with appropriately scaled vertical absorptionenergies offers the best compromise between accuracy and computationaleffort. An even more accurate and cost-effective approach, however,is to predict E (0,0) with machine learningand avoid the use of DFT for excited state calculations. Indeed, thebest excited state redox potential predictions are achieved with thecombination of M062X for ground state redox potentials and machinelearning (ML) for E (0,0). With this protocol,the excited state redox potential windows of the photocatalyst frameworkscould be adequately predicted. This shows the potential of combiningDFT with ML in the computational design of photocatalysts with preferredphotochemical properties.