Artificial neural network potential for pure zinc

Abstract

Zinc (Zn), because of its aberrant c/a ratio, has proven difficult to model using classical interatomic potentials such as the Modified Embedded Atom Method (MEAM). The limitations of existing formalisms in modeling Zn have been overcome here using a machine learned interatomic potential. This potential is trained using a database of density functional theory(DFT) calculations generated using the generalized gradient approximation. The resulting feedforward perceptron has a minimal architecture with a single hidden layer of 20 neurons. Validation of the network generated potential demonstrates that the potential correctly reproduced the training database while also predicting the experimentally observed c/a ratio as a function of temperature. The network is able to simultaneously predict the correct c/a ratio while also finding the hexagonally close packed structure as the ground state, which has not been previously demonstrated with semi-emprical potentials. This potential shows various results which are in good agreement with DFT and experimental calculation and will be a useful tool in the simulation of Zn at the molecular dynamics scale.