Prediction of bond dissociation energies using neural network, statistical, and quantum mechanical approaches

Abstract

Bond dissociation energies (BDEs) are extremely important in chemistry. However, they are notoriously difficult to calculate accurately using quantum mechanical (QM) techniques. Therefore, an alternative that gives similar accuracy to correlated QM techniques, but requires reduced computer time and resources, would be very useful for describing diverse chemical systems. Oxos provide a large library of experimental BDE data with which to assess the potential of different computational methods. Neural network, multiple linear regression, and multiconfiguration self-consistent field calculations are compared. A neural network outperforms MCSCF quantum calculations by a factor of 2.5 in prediction of BDEs for 52 element-oxos, incorporating main group, transition, lanthanide and actinide metal elements. The average absolute difference versus experiment is approximately 12.5 kcal mol −1 for neural network methods versus 28.4 kcal mol −1 for MCSCF calculations.