3. Setting up the system

3.1. Setting global variables in Python

The global variables in Python are controlled via the espressomd.system.System class. Global system variables can be read and set in Python simply by accessing the attribute of the corresponding Python object. Those variables that are already available in the Python interface are listed in the following. Note that for the vectorial properties box_l and periodicity, component-wise manipulation like system.box_l[0] = 1 or in-place operators like += or *= are not allowed and result in an error. This behavior is inherited, so the same applies to a after a = system.box_l. If you want to use a vectorial property for further calculations, you should explicitly make a copy e.g. via a = numpy.copy(system.box_l).

  • box_l

    (float[3]) Simulation box lengths of the cuboid box used by ESPResSo. Note that if you change the box length during the simulation, the folded particle coordinates will remain the same, i.e., the particle stay in the same image box, but at the same relative position in their image box. If you want to scale the positions, use the command change_volume_and_rescale_particles().

  • periodicity

    (int[3]) Specifies periodicity for the three directions. If the feature PARTIAL_PERIODIC is set, ESPResSo can be instructed to treat some dimensions as non-periodic. Per default espresso assumes periodicity in all directions which equals setting this variable to [1,1,1]. A dimension is specified as non-periodic via setting the periodicity variable for this dimension to 0. E.g. Periodicity only in z-direction is obtained by [0,0,1]. Caveat: Be aware of the fact that making a dimension non-periodic does not hinder particles from leaving the box in this direction. In this case for keeping particles in the simulation box a constraint has to be set.

  • time_step

    (float) Time step for MD integration.

  • time

    (float) The simulation time.

  • min_global_cut

    (float) Minimal total cutoff for real space. Effectively, this plus the skin is the minimally possible cell size. Espresso typically determines this value automatically, but some algorithms, virtual sites, require you to specify it manually.

  • max_cut_bonded

    read-only Maximal cutoff of bonded real space interactions.

  • max_cut_nonbonded

    read-only Maximal cutoff of bonded real space interactions.

3.1.1. Accessing module states

Some variables like or are no longer directly available as attributes. In these cases they can be easily derived from the corresponding Python objects like

n_part = len(espressomd.System().part[:].pos)

or by calling the corresponding get_state() methods like:

temperature = espressomd.System().thermostat.get_state()[0]['kT']

gamma = espressomd.System().thermostat.get_state()[0]['gamma']

gamma_rot = espressomd.System().thermostat.get_state()[0]['gamma_rotation']

3.2. Cellsystems

This section deals with the flexible particle data organization of ESPResSo. Due to different needs of different algorithms, ESPResSo is able to change the organization of the particles in the computer memory, according to the needs of the used algorithms. For details on the internal organization, refer to section Internal particle organization.

3.2.1. Global properties

The properties of the cell system can be accessed by espressomd.system.System.cell_system:

(int) Maximal number of cells for the link cell algorithm. Reasonable values are between 125 and 1000, or for some problems \(n_part / nnodes\).

(int) Minimal number of cells for the link cell algorithm. Reasonable values range in \(10^{-6} N^2\) to \(10^{-7} N^2\). In general just make sure that the Verlet lists are not incredibly large. By default the minimum is 0, but for the automatic P3M tuning it may be wise to set larger values for high particle numbers.

(int[3]) 3D node grid for real space domain decomposition (optional, if unset an optimal set is chosen automatically).

(float) Skin for the Verlet list. This value has to be set, otherwise the simulation will not start.

Details about the cell system can be obtained by espressomd.System().cell_system.get_state():

  • cell_grid Dimension of the inner cell grid.
  • cell_size Box-length of a cell.
  • local_box_l Local simulation box length of the nodes.
  • max_cut Maximal cutoff of real space interactions.
  • n_layers Number of layers in cell structure LAYERED
  • n_nodes Number of nodes.
  • type The current type of the cell system.
  • verlet_reuse Average number of integration steps the Verlet list is re-used.

3.2.2. Domain decomposition

Invoking set_domain_decomposition selects the domain decomposition cell scheme, using Verlet lists for the calculation of the interactions. If you specify use_verlet_lists=False, only the domain decomposition is used, but not the Verlet lists.

system = espressomd.System()

system.cell_system.set_domain_decomposition(use_verlet_lists=True)

The domain decomposition cellsystem is the default system and suits most applications with short ranged interactions. The particles are divided up spatially into small compartments, the cells, such that the cell size is larger than the maximal interaction range. In this case interactions only occur between particles in adjacent cells. Since the interaction range should be much smaller than the total system size, leaving out all interactions between non-adjacent cells can mean a tremendous speed-up. Moreover, since for constant interaction range, the number of particles in a cell depends only on the density. The number of interactions is therefore of the order N instead of order \(N^2\) if one has to calculate all pair interactions.

3.2.3. N-squared

Invoking set_n_square selects the very primitive nsquared cellsystem, which calculates the interactions for all particle pairs. Therefore it loops over all particles, giving an unfavorable computation time scaling of \(N^2\). However, algorithms like MMM1D or the plain Coulomb interaction in the cell model require the calculation of all pair interactions.

system = espressomd.System()

system.cell_system.set_n_square()

In a multiple processor environment, the nsquared cellsystem uses a simple particle balancing scheme to have a nearly equal number of particles per CPU, \(n\) nodes have \(m\) particles, and \(p-n\) nodes have \(m+1\) particles, such that \(n*m+(p-n)*(m+1)=N\), the total number of particles. Therefore the computational load should be balanced fairly equal among the nodes, with one exception: This code always uses one CPU for the interaction between two different nodes. For an odd number of nodes, this is fine, because the total number of interactions to calculate is a multiple of the number of nodes, but for an even number of nodes, for each of the \(p-1\) communication rounds, one processor is idle.

E.g. for 2 processors, there are 3 interactions: 0-0, 1-1, 0-1. Naturally, 0-0 and 1-1 are treated by processor 0 and 1, respectively. But the 0-1 interaction is treated by node 1 alone, so the workload for this node is twice as high. For 3 processors, the interactions are 0-0, 1-1, 2-2, 0-1, 1-2, 0-2. Of these interactions, node 0 treats 0-0 and 0-2, node 1 treats 1-1 and 0-1, and node 2 treats 2-2 and 1-2.

Therefore it is highly recommended that you use nsquared only with an odd number of nodes, if with multiple processors at all.

3.2.4. Layered cell system

Invoking set_layered selects the layered cell system, which is specifically designed for the needs of the MMM2D algorithm. Basically it consists of a nsquared algorithm in x and y, but a domain decomposition along z, i.e. the system is cut into equally sized layers along the z axis. The current implementation allows for the CPUs to align only along the z axis, therefore the processor grid has to have the form 1x1xN. However, each processor may be responsible for several layers, which is determined by n_layers, i.e. the system is split into N* layers along the z axis. Since in x and y direction there are no processor boundaries, the implementation is basically just a stripped down version of the domain decomposition cellsystem.

system = espressomd.System()

system.cell_system.set_layered(n_layers=4)

3.3. Thermostats

The thermostat can be controlled by the class espressomd.thermostat.Thermostat

The different available thermostats will be described in the following subsections. Note that for a simulation of the NPT ensemble, you need to use a standard thermostat for the particle velocities (Langevin or DPD), and a thermostat for the box geometry (the isotropic NPT thermostat).

You may combine different thermostats at your own risk by turning them on one by one. Note that there is only one temperature for all thermostats, although for some thermostats like the Langevin thermostat, particles can be assigned individual temperatures.

Since ESPResSo does not enforce a particular unit system, it cannot know about the current value of the Boltzmann constant. Therefore, when specifying the temperature of a thermostat, you actually do not define the temperature, but the value of the thermal energy \(k_B T\) in the current unit system (see the discussion on units, Section On units).

Note that there are three different types of noise which can be used in ESPResSo. The one used typically in simulations is flat noise with the correct variance and it is the default used in ESPResSo, though it can be explicitly specified using the feature FLATNOISE. You can also employ Gaussian noise which is, in some sense, more realistic. Notably Gaussian noise (activated using the feature GAUSSRANDOM) does a superior job of reproducing higher order moments of the Maxwell–Boltzmann distribution. For typical generic coarse-grained polymers using FENE bonds the Gaussian noise tends to break the FENE bonds. We thus offer a third type of noise, activate using the feature GAUSSRANDOMCUT, which produces Gaussian random numbers but takes anything which is two standard deviations (\(2\sigma\)) below or above zero and set it to \(-2\sigma\) or \(2\sigma\) respectively. In all three cases the distribution is made such that the second moment of the distribution is the same and thus results in the same temperature.

3.3.1. Langevin thermostat

In order to activate the Langevin thermostat the member function set_langevin of the thermostat class espressomd.thermostat.Thermostat has to be invoked. Best explained in an example:

import espressomd
system = espressomd.System()
therm = system.Thermostat()

therm.set_langevin(kT=1.0, gamma=1.0)

As explained before the temperature is set as thermal energy \(k_\mathrm{B} T\). The Langevin thermostat consists of a friction and noise term coupled via the fluctuation-dissipation theorem. The friction term is a function of the particle velocities. By specifying the diffusion coefficient for the particle becomes

\[D = \frac{\text{temperature}}{\text{gamma}}.\]

The relaxation time is given by \(\text{gamma}/\text{MASS}\), with MASS the particle’s mass. For a more detailed explanation, refer to [GK86]. An anisotropic diffusion coefficient tensor is available to simulate anisotropic colloids (rods, etc.) properly. It can be enabled by the feature PARTICLE_ANISOTROPY.

If the feature ROTATION is compiled in, the rotational degrees of freedom are also coupled to the thermostat. If only the first two arguments are specified then the diffusion coefficient for the rotation is set to the same value as that for the translation.

A separate rotational diffusion coefficient can be set by inputting gamma_rotate. This also allows one to properly match the translational and rotational diffusion coefficients of a sphere. Feature ROTATIONAL_INERTIA enables an anisotropic rotational diffusion coefficient tensor through corresponding friction coefficients.

Finally, the two options allow one to switch the translational and rotational thermalization on or off separately, maintaining the frictional behavior. This can be useful, for instance, in high Péclet number active matter systems, where one only wants to thermalize the rotational degrees of freedom and translational motion is effected by the self-propulsion.

Using the Langevin thermostat, it is possible to set a temperature and a friction coefficient for every particle individually via the feature LANGEVIN_PER_PARTICLE. Consult the reference of the part command (chapter Setting up particles) for information on how to achieve this.

3.3.2. Dissipative Particle Dynamics (DPD)

To realize a complete DPD fluid model, three parts are needed: The DPD thermostat, which controls the temperate, a dissipative interaction between the particles that make up the fluid, see DPD interaction, and a repulsive conservative force.

The DPD thermostat can be invoked by the function: espressomd.thermostat.Thermostat.set_dpd which takes \(k_\mathrm{B} T\) as the only argument.

The friction coefficients and cutoff are controlled via the DPD interaction on a per type-pair basis. For details see there.

As a conservative force any interaction potential can be used, see Isotropic non-bonded interactions. A common choice is a force ramp which is implemented as Hat interaction.

A complete example of setting up a DPD fluid and running it to sample the equation of state can be found in samples/dpd.py.

DPD adds a velocity dependent dissipative force and a random force to the conservative pair forces.

The friction (dissipative) and noise (random) term are coupled via the fluctuation- dissipation theorem. The friction term is a function of the relative velocity of particle pairs. The DPD thermostat is better for dynamics than the Langevin thermostat, since it mimics hydrodynamics in the system.

When using a Lennard-Jones interaction, \({r_\mathrm{cut}} = 2^{\frac{1}{6}} \sigma\) is a good value to choose, so that the thermostat acts on the relative velocities between nearest neighbor particles. Larger cutoffs including next nearest neighbors or even more are unphysical.

Boundary conditions for DPD can be introduced by adding the boundary as a particle constraint, and setting a velocity and a type on it, see espressomd.constraints.Constraint. Then a DPD interaction with the type can be defined, which acts as a boundary condition.

3.3.3. Isotropic NPT thermostat

This feature allows to simulate an (on average) homogeneous and isotropic system in the NPT ensemble. In order to use this feature, NPT has to be defined in the myconfig.hpp. Activate the NPT thermostat with the command set_npt() and set the following parameters:

  • kT: (float) Thermal energy of the heat bath
  • gamma0: (float) Friction coefficient of the bath
  • gammav: (float) Artificial friction coefficient for the volume fluctuations.

Also, setup the integrator for the NPT ensemble with set_isotropic_npt() and the parameters:

  • ext_pressure: (float) The external pressure as float variable.
  • piston: (float) The mass of the applied piston as float variable.

This thermostat is based on the Anderson thermostat (see [And80][Man05]) and will thermalize the box geometry. It will only do isotropic changes of the box. See this code snippet for the two commands:

import espressomd

system = espressomd.System()
system.thermostat.set_npt(kT=1.0, gamma0=1.0, gammav=1.0)
system.integrator.set_isotropic_npt(ext_pressure=1.0, piston=1.0)

Be aware that this feature is neither properly examined for all systems nor is it maintained regularly. If you use it and notice strange behavior, please contribute to solving the problem.

3.4. CUDA

CudaInitHandle() command can be used to choose the GPU for all subsequent GPU-computations. Note that due to driver limitations, the GPU cannot be changed anymore after the first GPU-using command has been issued, for example lbfluid. If you do not choose the GPU manually before that, CUDA internally chooses one, which is normally the most powerful GPU available, but load-independent.

system = espressomd.System()

dev = system.cuda_init_handle.device
system.cuda_init_handle.device = dev

The first invocation in the sample above returns the id of the set graphics card, the second one sets the device id.

3.4.1. GPU Acceleration with CUDA

Note

Feature CUDA required

ESPResSo is capable of GPU acceleration to speed up simulations. Not every simulation method is parallelizable or profits from GPU acceleration. Refer to Available simulation methods to check whether your desired method can be used on the GPU. In order to use GPU acceleration you need a NVIDIA GPU and it needs to have at least compute capability 2.0.

For more information please check espressomd.cuda_init.CudaInitHandle.

3.4.2. List available CUDA devices

If you want to list available CUDA devices you should access espressomd.cuda_init.CudaInitHandle.device_list, e.g.,

system = espressomd.System()

print(system.cuda_init_handle.device_list)

This attribute is read only and will return a dictionary containing the device id as key and the device name as its’ value.

3.4.3. Selection of CUDA device

When you start pypresso your first GPU should be selected. If you wanted to use the second GPU, this can be done by setting espressomd.cuda_init.CudaInitHandle.device as follows:

system = espressomd.System()

system.cuda_init_handle.device = 1

Setting a device id outside the valid range or a device which does not meet the minimum requirements will raise an exception.