Quantum-informed regression: local hybrid functionals enhance machine learning prediction of dye absorption maxima

Abstract Prediction of optical properties of molecules still remains a challenge in the design of dye-sensitized solar cells. By linking quantum-chemical accuracy with the flexibility of machine learning (ML), this work offers a glimpse into how data-driven approaches can reshape the way we design next-generation photovoltaic materials. In this work, we combine density functional theory (DFT) and time-dependent DFT (TDDFT) with ML models to assess the predictive performance of maximum absorption of three classes of exchange-correlation functionals: the global hybrid (B3LYP), the range-separated (CAM-B3LYP) and the local hybrid functional (Lh12ct-SsifPW92). Our dataset consists of 21 well-characterized double donor/acceptor dyes extracted from the literature and was generated through TURBOMOLE calculations. Inputs from the quantum-chemical software were then trained by 14 different regression models, including linear, instance and tree-based models. Validations of predictive results were done by root mean squared error, leave-one-out cross-validation and, at the end, with an external set of five dyes, which were not included in the training set, to ensure independent testing. Results showed that the local hybrid functional offers the most accurate agreement with the TDDFT reference values and, therefore, sets itself as the potential best choice for the small dataset approach.