(377j) Efficient Neighbor List Calculation for Molecular Simulation of Colloidal Systems Using Graphics Processing Units | AIChE

(377j) Efficient Neighbor List Calculation for Molecular Simulation of Colloidal Systems Using Graphics Processing Units


Howard, M. P. - Presenter, Princeton University
Anderson, J. A., University of Michigan
Nikoubashman, A., Princeton University
Glotzer, S. C., University of Michigan
Panagiotopoulos, A. Z., Princeton University
Soft matter systems are characterized by a significant disparity in relevant length and time scales between constituent components. For example, colloidal particles (nanometers to micrometers in diameter) in solution are separated in size by several orders of magnitude from an atomistic description of the solvent. Molecular dynamics (MD) simulations retaining full atomistic detail of the solvent can become intractable because many solvent atoms must be included to model only a few colloidal particles. Moreover, the time scales associated with the degrees of freedom of the solvent are generally much shorter than the relatively slow motion of the larger colloids. This means that these simulations require very short MD time steps to faithfully capture the dynamics of the solvent, and many such steps are required to observe any appreciable dynamics of the colloids.

Calculation of the forces between particles, and in particular, nonbonded pair interactions, dominates the computational time of the MD algorithm. In order to accelerate these calculations, pair potentials are frequently truncated at a finite cutoff distance, and a neighbor (Verlet) list of particles within each othersâ??s cutoffs is maintained to reduce the number of distance evaluations between particles that are performed. The calculation of the neighbor list itself is accelerated using a structure such as a spatially-binned cell list that reduces the overall computational complexity of the algorithm.

We present an algorithm based on linear bounding volume hierarchies (LBVHs) for computing neighbor (Verlet) lists using graphics processing units (GPUs) for colloidal systems characterized by large size disparities. We compare this in the HOOMD-blue simulation package to a GPU implementation of the current state-of-the-art CPU algorithm based on stenciled cell lists. We report benchmarks for both neighbor list algorithms in a Lennard-Jones binary mixture with synthetic interaction range disparity and a realistic colloid solution. LBVHs outperform the stenciled cell lists for systems with moderate or large size disparity and dilute or semidilute fractions of large particles, conditions typical of colloidal systems. The LBVH neighbor list algorithm has recently been applied to study transport and structure formation in colloidal dispersions under flow.