The homolytic cleavage of covalent bonds underpins much of the chemistry involved in combustion, metabolism and catalysis, among many other areas. The ability to routinely and rapidly predict bond dissociation energy (BDE) values is therefore of great value to a broad cross-section of chemists. However, predicting BDE values for all breakable bonds in molecules using conventional applications of quantum chemistry, while accurate, can be computationally expensive and tedious. Herein, we summarize our progress in building a deep learning model and BDE database, which serve as training set of the network. This approach is applied to thousands of molecules in the PubChem dataset, where it is demonstrated to provide comparable accuracy to DFT calculations at significantly lower costs. Model training could be carried out in several hours on GPUs. We developed an automated workflow built on the Atomic Simulation Environment Data Base to obtain DFT reference BDE values for all breakable bonds in a large collection of organic molecules. We modified a continuous-filter convolutional neural network (Schnet), which refines atom representations considering interactions with neighboring atoms, to predict BDE for molecules.

Predicting bond dissociation energies through deep learning

2019 Spring National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.