Molecular dynamics simulations allow for the approximation of particle movements using
numerical methods. Often, 2-body potentials are used for this, but the accuracy achieved is
insufficient in more specific scenarios, which is why 3-body potentials, such as the Axilrod-
Teller potential, can also be calculated.
However, this can only be achieved with significantly higher computational effort, as the
computational cost of 3-body potentials grows cubically with the number of particles, and
molecular dynamics often involves millions of particles. Therefore, it is crucial to use the
most efficient algorithms for force calculation depending on the composition of the simulation
space.
AutoPas is a C++ library that enables auto-tuning of molecular dynamics simulations.
To offer AutoPas more options for calculating 3-body potentials, we implement three new
methods for traversing particle triplets based on the Verlet Cluster List algorithm and
compare them with similar methods from the library that are based on the Linked Cells
and Verlet Lists algorithms.
We find that the new traversal methods generally perform better in most scenarios than
traversals using the Linked Cells algorithm, but they perform significantly worse when
compared to similar methods based on the Verlet Lists algorithm. The force calculation
during traversal seems to be inefficient with the Array-of-Structures data layout, so any
increase in the number of particles per cluster degrades the performance — only for one of
our traversals the optimal cluster size turned out to be 2.
However, due to the generally better parallel efficiency, lower memory requirements, and
the lower increase in runtime both for a higher cutoff and for higher particle density, the
traversal has potential for larger simulations where the cluster structure might be utilized
better.
«
Molecular dynamics simulations allow for the approximation of particle movements using
numerical methods. Often, 2-body potentials are used for this, but the accuracy achieved is
insufficient in more specific scenarios, which is why 3-body potentials, such as the Axilrod-
Teller potential, can also be calculated.
However, this can only be achieved with significantly higher computational effort, as the
computational cost of 3-body potentials grows cubically with the number of particles, and...
»