A combined first-principles calculation and neural networks correction approach for evaluating Gibbs energy of formation

Abstract

Despite of their successes, the results of first-principles quantum mechanical calculations contain inherent numerical errors that are caused by inadequate treatment of electron correlation, incompleteness of basis sets, relativistic effects or approximated exchange-correlation functionals. In this work, we develop a combined density-functional theory and neural-network correction (DFT-NEURON) approach to reduce drastically these errors, and apply the resulting approach to determine the standard Gibbs energy of formation ΔG° at 298 K for small- and medium-sized organic molecules. The root mean square deviation of the calculated ΔG° for 180 molecules is reduced from 22.3kcal-mol-1 to 3.0 kcal-mol-1 for B3LYP/6-311 + G(d,p). We examine further the selection of physical descriptors for the neural network.