A new approach to potential fitting using neural networks

Abstract A methodology is presented for developing transferable empirical potential functions without following the usual procedure of postulating a functional form. Instead, a neural network (NN) is employed to learn the functional relationships of potential energy surfaces from the local geometric arrangement of atoms. The methodology is illustrated by training the NN model on tens of thousands of individual data points derived from the tight-binding (TB) method for a wide range of silicon systems including both small clusters and bulk structures. Comparisons of the potential’s properties with experimental data, quantum methods and other Si potentials have been made. The NN model successfully fitted energy variations of the different test cases as a function of bond distances, bond angles, lattice constants and elastic properties for both equilibrium and non-equilibrium small cluster and bulk structures. This indicates a robust and consistent methodology for fitting empirical potentials which can be applied to a wide range of materials independent of the type of bonding or their crystal structure.