Representation of intermolecular potential functions by neural networks

The use of neural networks to represent intermolecular potential functions was investigated. Feed-forward networks of different sizes were utilized to reproduce the three-body interaction energies in the molecular system H(sub 2)O-Al(super 3+)-H(sub 2)O. The three-body interaction terms were found to be necessary for an accurate description of the molecular system. Experimental results showed that a neural network representation of potential functions has certain advantages that outweigh the disadvantages in lack of interpretability or computing speed.