Machine-Learning Predictions of Rate Constants of Internal Conversion Using Electronic and Structural Descriptors

We present a machine learning approach to predict internal conversion rate constants (k_IC) based on molecular electronic and structural descriptors. Our approach addresses the challenge of accurately estimating k_IC between the first excited singlet state (S₁) and the ground state (S₀), which is an important deactivation channel of photophysical processes. Using a cost-efficient method that assumes excitation-energy transfer to the vibrations of the X-H bonds (X = C, N, O), we generated a data set for k_IC of more than 5000 molecules, including porphyrins and [8]circulenes using parameters calculated at the time-dependent density functional theory (TDDFT) level. We have used both the CatBoost (CB) model and a neural network (NN) including transformer-based and graph-based architectures. Using the CB and NN models we obtained high precision with R² values of about 0.99. The models successfully predict k_IC even for new molecules when they are trained with energy-augmented data sets based on electronic and structural descriptors. We also introduce graph-based X-H bond descriptors to predict rate constants for various chemical classes. The calculated rate constants indicate the promising potential of these descriptors to predict k_IC using machine learning to bypass computationally demanding quantum mechanical calculations to determine the photophysical properties.