Hard particle Monte Carlo
HPMC performs hard particle Monte Carlo simulations of a variety of classes of shapes.
HPMC implements hard particle Monte Carlo in HOOMD-blue. It supports:
- Dimensions: 2D and 3D
- Box shape: triclinic
- Spheres / disks (
- Union of spheres (
- Convex polygons (
- Convex spheropolygons (
- Simple polygons (
- Ellipsoids / ellipses (
- Convex polyhedra (
- Convex spheropolyhedra (
- Faceted spheres (
- General polyhedra (
- Spheres / disks (
- Canonical hard particle MC on a single CPU core
- Parallel update scheme on a single GPU
- Parallel updates on many CPU cores / GPUs using MPI
- Frenkel-Ladd free energy determination
- File I/O:
- Loose integration with pos_writer
The following quantities are provided by the integrator for use in HOOMD-blue’s
hpmc_sweep- Number of sweeps completed since the start of the MC integrator
hpmc_translate_acceptance- Fraction of translation moves accepted (averaged only over the last time step)
hpmc_rotate_acceptance- Fraction of rotation moves accepted (averaged only over the last time step)
hpmc_d- Maximum move displacement
hpmc_a- Maximum rotation move
hpmc_move_ratio- Probability of making a translation move (1- P(rotate move))
hpmc_overlap_count- Count of the number of particle-particle overlaps in the current system configuration
With non-interacting depletant (implicit=True), the following log quantities are available:
hpmc_fugacity- The current value of the depletant fugacity (in units of density, volume^-1)
hpmc_ntrial- The current number of configurational bias attempts per overlapping depletant
hpmc_insert_count- Number of depletants inserted per colloid
hpmc_reinsert_count- Number of overlapping depletants reinserted per colloid by configurational bias MC
hpmc_free_volume_fraction- Fraction of free volume to total sphere volume after a trial move has been proposed (sampled inside a sphere around the new particle position)
hpmc_overlap_fraction- Fraction of deplatants in excluded volume after trial move to depletants in free volume before move
hpmc_configurational_bias_ratio- Ratio of configurational bias attempts to depletant insertions
compute.free_volume provides the following loggable quantities:
hpmc_free_volume - The free volume estimate in the simulation box obtained by MC sampling (in volume units)
update.boxmc provides the following loggable quantities:
hpmc_boxmc_trial_count- Number of box changes attempted since the start of the boxmc updater
hpmc_boxmc_volume_acceptance- Fraction of volume/length change trials accepted (averaged from the start of the last run)
hpmc_boxmc_shear_acceptance- Fraction of shear trials accepted (averaged from the start of the last run)
hpmc_boxmc_aspect_acceptance- Fraction of aspect trials accepted (averaged from the start of the last run)
hpmc_boxmc_betaPCurrent value of the \(\beta p\) value of the boxmc updater
update.muvt provides the following loggable quantities.
hpmc_muvt_insert_acceptance- Fraction of particle insertions accepted (averaged from start of run)
hpmc_muvt_remove_acceptance- Fraction of particle removals accepted (averaged from start of run)
hpmc_muvt_volume_acceptance- Fraction of particle removals accepted (averaged from start of run)
HOOMD-blue started as an MD code where timestep has a clear meaning. MC simulations are run for timesteps. In exact terms, this means different things on the CPU and GPU and something slightly different when using MPI. The behavior is approximately normalized so that user scripts do not need to drastically change run() lengths when switching from one execution resource to another.
In the GPU implementation, one trial move is applied to a number of randomly chosen particles in each cell during one
timestep. The number of selected particles is
nselect*ceil(avg particles per cell) where nselect is a user-chosen
parameter. The default value of nselect is 4, which achieves optimal performance for a wide variety of benchmarks.
Detailed balance is obeyed at the level of a timestep. In short: One timestep is NOT equal to one sweep,
but is approximately nselect sweeps, which is an overestimation.
In the single-threaded CPU implementation, one trial move is applied nselect times to each of the N particles during one timestep. In parallel MPI runs, one trial moves is applied nselect times to each particle in the active region. There is a small strip of inactive region near the boundaries between MPI ranks in the domain decomposition. The trial moves are performed in a shuffled order so detailed balance is obeyed at the level of a timestep. In short: One timestep is approximately nselect sweeps (N trial moves). In single-threaded runs, the approximation is exact, but it is slightly underestimated in MPI parallel runs.
To approximate a fair comparison of dynamics between CPU and GPU timesteps, log the
quantity to get the number sweeps completed so far at each logged timestep.
See J. A. Anderson et. al. 2016 for design and implementation details.
hoomd.hpmc is stable. When upgrading from version 2.x to 2.y (y > x),
existing job scripts that follow documented interfaces for functions and classes
will not require any modifications. Maintainer: Joshua A. Anderson