Neural Networks has employed to correct the systematic deviations of the first principle quantum mechanics calculated results. The approach was termed as the DFT-NEURON method. This combined first-principles calculation and Neural-Networks correction approach is implemented to calculate the Gibbs energy of formation, ionization energy, electron affinity and absorption energy of small to medium sized molecules, and leads to much improved values as compared to the conventional first-principles calculation results.
Neural networks are first trained and determined against accurate experimental data, and employed subsequently to improve the raw calculated properties of interests. This combined first-principle calculation and Neural-Networks-corrections approach is potentially a powerful tool in computational science, and may open the possibility for first-principles methods to be employed routinely as predictive tools in materials research and development.
FIGURE Calculated DG° versus experimental DG° for all 180 compounds. (a) and (b) are for raw B3LYP/6-311+G(d,p) and B3LYP/6-311+G(3df,2p) results, respectively. (c) and (d) are for neural-network corrected B3LYP/6-311+G(d,p) and B3LYP/6-311+G(3df,2p) DG°s, respectively. Triangles are for the training set and crosses for the testing set. The dashed lines are where the calculation and experiments have perfect fits. The correlation coefficient of the linear fits are 0.996, 0.995 in (c) and (d), respectively. All values are in the units of kcal·mol-1.