Molecular Dynamics (MD) simulations of atomic or molecular microscopic systems are very important in several fields of science. The present project at Center for Atomic--scale Materials Physics of the Technical University of Denmark, in collaboration with Joint Research Center for Atom Technology in Tsukuba deals with the materials--physics properties of ``large'' metallic solid--state systems, consisting of typically 10.000 to 1 million atoms.
The basis of all MD studies is a theoretical model of the interactions between the constituent atoms. In the present work a theoretical framework called the Effective Medium Theory[1] (EMT) for atomic interactions has been used. In this theory, charge density ``tails'' from neighboring atoms are accumulated by every atom, and a non-linear functional is used to derive the interatomic energies and forces. Electrostatic energies are added to the density term, and one-electron energies may be included as well if required. Owing to the physical properties of metals, these interactions have a short range of 4--5 atomic radii. Some forms of the EMT theory resemble semi-empirical methods such as the Embedded Atom Method[2].
Short ranged interactions dictate the types of algorithms chosen for MD simulations: Only well localized interactions need to be taken into account, in contrast to other types of systems requiring global interactions. Hence MD algorithms must be considered which allow for efficient access to atomic data in localized regions of space. A review of current research in parallel short-range MD algorithms is presented in ref. [3].
The parallel algorithm chosen in the present work is based upon a decomposition of the physical system into identical cells laid out on a regular 3--dimensional grid. Thus the algorithm only needs to consider data that reside in nearby cells on the grid. This is naturally reflected in the implementation on distributed--memory parallel computers such as the Connection Machines. This was briefly described in our previous work[4], which also reported performance measurements for CM-200 and CM-5E supercomputers.