Prediction of bond orders using deep neural networks

It has been shown in previous works that machine learning (ML) can be utilized to correctly predict atomic partial charges. This success has paved the way for predicting more advanced molecular properties. The hierarchical interacting particle neural network (HIP-NN) provides a model framework to predict covalent, ionic, and total bond order matricies. Utilizing HIP-NN, we were able to predict, with high efficiency, the coefficient bond order matrix with reference to density functional theory (DFT) calculations. The neural network was trained to a set of different molecular arrangements of hydrogen, oxygen, carbon, and nitrogen atoms. By training HIP-NN to this large set of DFT computed training molecules we were able to reduce the error between the outputted values of the coefficient bond order matrix and the values predicted by DFT to only be on the scale of 1e-3. This error reduction, combined with computational speed, demonstrates that ML could be a potential avenue for computing bond order matrices for molecules.