Molecular dynamics simulations memory requirements

There are presently several groups around the world conducting molecular dynamics simulations of micellization and liquid crystallization of more or less realistic models of water, hydrocarbon, and surfactants. The memory and speed of a supercomputer required to produce reliably equilibrated microstructures constitute a challenge not yet met, in my opinion. By taking advantage of identified or hypothesized elemental structures one can, however, hope to learn a great deal about the dynamics and stability of the various identified microstructures. 

One drawback of this method is that it requires the use of an extra parameter, t, the time constant for the decay of the memory function. In practice, this should not be a problem as long as t is longer than the period of the longest motions which are important for NOE averaging. A limitation of the method is that, in contrast to the two-state model of Scarsdale et al.,47 the procedure can only average over conformations that can interchange during the short time of a molecular dynamics simulation. 

Besides these limitations the most unwelcome feature common to the SSSV and RMFA theories is that they need dynamical information from outside, usually from molecular dynamics simulations. For example, diffusion coefficients of constituent atoms of a molecule are required in the SSSV theory prior to its application, whereas in the RMFA the memory function of a reference dynamical variable has to be determined in advance from molecular dynamics simulations. 

A distinctive feature of our theory is that it offers a self-contained framework for solving dynamical problems without making reference to any dynamical information from outside, i.e., it requires only the knowledge of parameters for potential functions and molecular geometry (such as bond lengths) as in simulation studies. This is in contrast to previous theories, the SSSV and RMFA theories, in which diffusion coefficients or memory functions of some reference dynamical variables have to be determined in advance from molecular dynamics simulations. 

Molecular dynamics simulation is perhaps the most powerful computational technique available for obtaining information on time dependent properties of molecular or atomic motion in zeolite crystals. It is used to obtain thermodynamic quantities and detailed dynamical information on sorption and diffusion processes in zeolite systems. For instance, the extent to which intramolecular vibration and framework motion assist sorption and diffusion of molecules can be simulated. The major limitation is its inability to model diffusion of larger sorbed molecules and electronic polarisability due to the huge amount of computer time and memory requirements. However, with the improvement in supercomputers and improved computing facilities, the full application of M.D. simulation to zeolite studies is becoming feasible. 

With their strength tied to available computer speed, simulations continue to become a more powerful tool. A letter to the Journal of Chemical Physics by B. J. Alder and T. E. Wainwright in 1957 was the first work that reported results from molecular dynamics simulations. The Lawrence Radiation Laboratory scientists studied two different sized systems with 32 and 108 hard spheres. They modeled bulk fluid phases using periodic boundary conditions. In the paper they mention that they counted 7000 and 2000 particle collisions for 32 and 108 particle systems, respectively. This required one hour on a UNIVAC computer. Incidentally, this was the fifth such commercial computer delivered out of the 46 ever produced. The computer cost around 200 000 in 1952 and each of ten memory units held 100 words or bytes. Nowadays, a 300 personal computer with a memory of approximately 500000000 bytes can complete this simulation in less than 1 second. And Moore s empirical law that computer power doubles every 18 months still holds. 

The polarizable point dipole model has also been used in Monte Carlo simulations with single particle moves.When using the iterative method, a whole new set of dipoles must be computed after each molecule is moved. These updates can be made more efficient by storing the distances between all the particles, since most of them are unchanged, but this requires a lot of memory. The many-body nature of polarization makes it more amenable to molecular dynamics techniques, in which all particles move at once, compared to Monte Carlo methods where typically only one particle moves at a time. For nonpolarizable, pairwise-additive models, MC methods can be efficient because only the interactions involving the moved particle need to be recalculated [while the other (N - 1) x (]V - 1) interactions are unchanged]. For polarizable models, all N x N interactions are, in principle, altered when one particle moves. Consequently, exact polarizable MC calculations can be... 

---