Molecular Geometry Impact on Deep Learning Predictions of Inverted Singlet–Triplet Gaps

Abstract

We present a deep learning model able to predict excited singlet−triplet gaps with a mean absolute error (MAE) of ≈20 meV to obtain potential inverted singlet−triplet (IST) candidates. We exploit cutting-edge spherical message passing graph neural networks designed specifically for generating 3D graph representations in molecular learning. In a nutshell, the model takes as input a list of unsaturated heavy atom Cartesian coordinates and atomic numbers, producing singlet−triplet gaps as output. We exploited available large data collections to train the model on ≈40,000 heterogeneous density functional theory (DFT) geometries with available ADC(2)/cc-pVDZ singlet−triplet gaps. We ascertain the predictive power of the model from a quantitative perspective obtaining predictions on a test set of ≈14,000 molecules, whose geometries have been generated at DFT level (the same employed for the geometries in the training set), at GFN2-xTB level, and through Molecular Mechanics. We notice performance degradation upon switching to lower-quality geometries, with GFN2-xTB ones maintaining satisfactory results (MAE ≈ 50 meV on GFN2-xTB geometries, MAE ≈ 180 meV on generalized AMBER force field geometries), hinting at caution when dealing with specific chemical classes. Finally, we verify the performance of the model from the qualitative point of view, obtaining predictions on a different data set of ≈15,000 molecules already used to identify new IST molecules. We obtained predictions using both DFT and experimental X-ray geometries, with results on IST candidates similar to those provided by quantum chemical methods, with clear hints for the path toward improved performance.