DescriptionArticial neural network is revolutionizing many areas in science and technology. We applied articial neural network to solve a non-linear optimization problem in computational chemistry, i.e. molecular geometry optimization, which aims to nd an atomic arrangement that corresponds to a stationary point on the potential energy surface.
The implemented ANN can use both function values and derivatives as the reference data for training. The relative importance of function values and derivatives is studied extensively. With the same amount data points, ANN trained with derivatives tend to generalize better. With only derivatives as the reference data, the trained ANN can predict function values accurately if a common offset is allowed.
We trained ANNs that can predict molecule energies and gradients fairly well. Molecular geometry optimization is performed on testing data that are never seen during the training process. About 5% of the testing structures converge with average 0.047-angstrom RMSD compared with equilibrium state, while others diverge. The divergence is ascribed to poor gradient prediction near the equilibrium.