A Framework for the Vectorization of Molecular Dynamics Kernels


While the Lennard-Jones class of potentials is the most widely used family of interatomic potentials in molecular simulations, there are many additional families of potentials which are also of importance, particularly in the simulation of solids. Among these are the multiple classes of many-body potentials that are frequently used in simulations of solids, and especially of metals, where strong electron sharing between nearby atoms can occur, and where the dense packing of atoms in a solid structure enables significant interactions between sets of three adjacent atoms, rather than the purely pairwise interactions assumed by the Lennard-Jones and Buckingham families of potentials. Two commonly encountered sets of potentials in this respect are the Tersoff and AIREBO potentials,which can be used in the simulation of, for instance, graphene, carbon nanotubes, and hydrocarbons. These potentials add a number of additional two- and three-body terms to the Lennard-Jones potentials to account for the aforementioned effects.

In this project, we propose to optimize several existing classes of potentials in LAMMPSby taking advantage of the Xeon Phis capabilities, and to extend the range of potentials included in the LAMMPS USER-INTEL package. We will focus our efforts on two classes of so-called many-body potentials: the Tersoff and AIREBO classes, which are widely used for simulations of carbon nanotubes and hydrocarbons. Continuing, we will generalize our approach to automatically derive optimized implementations of other many-body potentials, such as Stillinger-Weber, EAM, EIM, BOP or ReaxFF.