Dataset of Singlet and Triplet Energies and Forces for Organic Molecules

This is the training and extensibility set associated with the publication “Predicting Phosphorescence Energies and Inferring Wavefunction Localization with Machine Learning.” As part of the review process, it is necessary to release this dataset so published results can be verified by the reviewers and the broader scientific community. The dataset consists of singlet and triplet molecular energies and energy gradients as computed with the Gaussian software suite for which Los Alamos National Laboratory has a license. The training dataset consists of 355,011 molecular structures optimized in the singlet electronic state and 346,745 molecules optimized the triplet electronic state. The extensibility test set consists of 827 molecules optimized in the singlet state and 912 molecules optimized in the triplet state. This dataset is useful for training neural networks, specifically HIPPYNN which has been released through the Feynman Center for Innovation. The trained neural network can make accurate predictions of single and triplet molecular geometries and energies in both the ground and triplet electronic states much faster than traditional quantum mechanical methods. These neural networks are of relevance to basic research into organic chromophores, where singlet-triplet energy gaps and geometry differences play an important role in optical properties.<br>Representative examples.