ParticleSystem interface

The ParticleSystem interface facilitates the use of CellListMap for the majority of cases.

The mapped function

The purpose of CellListMap is to compute a pairwise-dependent function for all pairs of particles that are closer to each other than a defined cutoff. This pairwise function must be implemented by the user and adhere to the following interface:

function f(x, y, i, j, d2, output)
    # update output variable
    return output
end

where x and y are the positions of the particles, already wrapped relative to each other according to the periodic boundary conditions (a minimum-image set of positions), i and j are the indexes of the particles in the arrays of coordinates, d2 is the squared distance between the particles, and output is the variable to be computed.

Basic examples

For example, computing the energy, as the sum of the inverse of the distance between particles, can be done with a function like:

function energy(d2,u)
    u += 1 / sqrt(d2)
    return u
end

and the additional parameters required by the interface can be eliminated by the use of an anonymous function, directly on the call to the map_pairwise function:

u = map_pairwise(
    (x,y,i,j,d2,u) -> energy(d2,u), 
    system
)

(what system is will be explained in the examples below). Note that the energy function does not use the x, y, i, and j input parameters, such that the anonymous function managing the interface could also be written as (_, _, _, _, d2, u) -> energy(d2, u), making explicit the dummy character of these variables in the example.

Alternatively, the function might require additional parameters, such as the masses of the particles. In this case, we can use a closure to provide such data:

function energy(i,j,d2,u,masses)
    u += masses[i]*masses[j] / sqrt(d2)
    return u
end
const masses = # ... some masses
u = map_pairwise((x,y,i,j,d2,u) -> energy(d2,u,masses), system)

Here we reinforce the fact that the energy functions defined above compute the contribution to the energy of the interaction of a single pair of particles. This function will be called for every pair of particles within the cutoff, automatically, in the map_pairwise call.

The ParticleSystem constructor

Potential energy example

For example, here we read the coordinates of Argon atoms from a PDB file. The coordinates are given as vector of SVectors. We then compute an "energy", which in this case is simply the sum of 1/d over all pair of particles, within a cutoff.

The ParticleSystem constructor receives the properties of the system and sets up automatically the most commonly used data structures necessary.

julia> using CellListMap, PDBTools

julia> argon_coordinates = coor(readPDB(CellListMap.argon_pdb_file))

julia> system = ParticleSystem(
           xpositions=argon_coordinates,
           unitcell=[21.0,21.0,21.0], 
           cutoff = 8.0, 
           output = 0.0,
           output_name = :energy
       );

Now, let us compute the energy of the particles, assuming a simple formula which depends on the inverse of the distance between pairs:

julia> function energy(x, y, i, j, d2, energy)
           energy += 1 / sqrt(d2)
           return energy
       end

julia> map_pairwise!(energy, system)
207.37593043370865

Note that the first four parameters of energy are not used here but are needed to adhere to the interface. The function input could be written as (_, _, _, _, d2, energy) to make that explicit.

The system.energy field accesses the resulting value of the computation:

julia> system.energy
207.37593043370865

because the output_name field was provided. If it is not provided, you can access the output value from the system.energy field.

Computing forces

Following the example above, let us compute the forces between the particles. We have to define the function that computes the force between a pair of particles and updates the array of forces:

function update_forces!(x,y,i,j,d2,forces)
    d = sqrt(d2)
    df = (1/d2)*(1/d)*(y - x)
    forces[i] += df
    forces[j] -= df
    return forces
end

Importantly, the function must return the forces array to follow the API.

Now, let us setup the system with the new type of output variable, which will be now an array of forces with the same type as the positions:

julia> using CellListMap, PDBTools

julia> argon_coordinates = coor(readPDB(CellListMap.argon_pdb_file))

julia> system = ParticleSystem(
           xpositions=argon_coordinates,
           unitcell=[21.0, 21.0, 21.0], 
           cutoff = 8.0, 
           output = similar(argon_coordinates),
           output_name = :forces
       );

Let us note that the forces where reset upon the construction of the system:

julia> system.forces
1000-element Vector{SVector{3, Float64}}:
 [0.0, 0.0, 0.0]
 [0.0, 0.0, 0.0]
 ⋮
 [0.0, 0.0, 0.0]

A call to map_pairwise! with the appropriate function definition will update the forces:

julia> map_pairwise!((x,y,i,j,d2,forces) -> update_forces!(x,y,i,j,d2,forces), system)
100-element Vector{SVector{3, Float64}}:
 [0.026493833307357332, 0.18454277989323772, -0.012253902366284965]
 [0.07782602581235695, 0.2791082233740261, 0.21926615329195248]
 ⋮
 [0.11307234751448932, 0.006353545239676281, -0.05955687310348302]
 [-0.03101200918307673, 0.03543655648545697, 0.031849121630976335]

Computing both energy and forces

In this example we define a general type of output variable, for which custom copy, reset, and reduction functions must be defined. It can be followed for the computation of other general properties from the particle positions.

using CellListMap, StaticArrays, PDBTools

The computation of energies and forces in a single call is an interesting example for the definition of a custom output type and the required interface functions. Let us first define an output variable containing both quantities:

mutable struct EnergyAndForces
    energy::Float64
    forces::Vector{SVector{3,Float64}}
end

Now we need to define what it means to copy, reset, and reduce this new type of output. We overload the default corresponding functions, for our new output type:

The copy method creates a new instance of the EnergyAndForces type, with copied data:

function CellListMap.copy_output(x::EnergyAndForces) 
    return EnergyAndForces(copy(x.energy), copy(x.forces))
end

The reset method will zero both the energy and all forces:

function CellListMap.reset_output!(output::EnergyAndForces)
    output.energy = 0.0
    for i in eachindex(output.forces)
        output.forces[i] = SVector(0.0, 0.0, 0.0)
    end
    return output
end

The reducer function defines what it means to combine two output variables obtained on independent threads. In this case, we sum the energies and forces. Different reduction functions might be necessary for other custom types (for example if computing minimum distances).

function CellListMap.reducer(x::EnergyAndForces, y::EnergyAndForces)
    e_tot = x.energy + y.energy
    x.forces .+= y.forces
    return EnergyAndForces(e_tot, x.forces)
end

Note that in the above example, we reuse the x.forces array in the return instance of EnergyAndForces. You must always reduce from right to left, and reuse the possible buffers of the first argument of the reducer (in this case, x).

Now we can proceed as before, defining a function that updates the output variable appropriately:

function energy_and_forces!(x,y,i,j,d2,output::EnergyAndForces)
    d = sqrt(d2)
    output.energy += 1/d
    df = (1/d2)*(1/d)*(y - x)
    output.forces[i] += df
    output.forces[j] -= df
    return output
end

To finally define the system and compute the properties:

argon_coordinates = coor(readPDB(CellListMap.argon_pdb_file))

system = ParticleSystem(
    xpositions = argon_coordinates,
    unitcell = [21.0, 21.0, 21.0], 
    cutoff = 8.0, 
    output = EnergyAndForces(0.0, similar(argon_coordinates)),
    output_name = :energy_and_forces
);

map_pairwise((x,y,i,j,d2,output) -> energy_and_forces!(x,y,i,j,d2,output), system);

The output can be seen with the aliases of the system.output variable:

julia> system.energy_and_forces.energy
207.37593043370862

julia> system.energy_and_forces.forces
100-element Vector{SVector{3, Float64}}:
 [0.02649383330735732, 0.18454277989323772, -0.012253902366284958]
 [0.07782602581235692, 0.27910822337402613, 0.21926615329195248]
 ⋮
 [0.11307234751448932, 0.006353545239676281, -0.05955687310348303]
 [-0.031012009183076745, 0.03543655648545698, 0.03184912163097636]

Updating coordinates, unit cell, and cutoff

If the map_pairwise! function will compute energy and/or forces in a iterative procedure (a simulation, for instance), we need to update the coordinates, and perhaps the unit cell and the cutoff.

• Updating coordinates
• Updating the unit cell
• Updating the cutoff

Updating coordinates

The coordinates can be updated (mutated, or the array of coordinates can change in size by pushing or deleting particles), simply by directly accessing the xpositions field of the system. The xpositions array is a Vector of SVector (from StaticArrays), with coordinates copied from the input array provided. Thus, the coordinates in the ParticleSystem structure must be updated independently of updates in the original array of coordinates.

Let us exemplify the interface with the computation of forces:

julia> using CellListMap, StaticArrays

julia> positions = rand(SVector{3,Float64}, 1000);

julia> system = ParticleSystem(
           xpositions = positions,
           unitcell=[1,1,1], 
           cutoff = 0.1, 
           output = similar(positions),
           output_name = :forces
       );

julia> system.xpositions[1]
3-element SVector{3, Float64} with indices SOneTo(3):
 0.6391290709055079
 0.43679325975360894
 0.8231829019768698

julia> system.xpositions[1] = zeros(SVector{3,Float64})
3-element SVector{3, Float64} with indices SOneTo(3):
 0.0
 0.0
 0.0

julia> push!(system.xpositions, SVector(0.5, 0.5, 0.5))
1001-element Vector{SVector{3, Float64}}:
 [0.0, 0.0, 0.0]
 [0.5491373098208292, 0.23899915605319244, 0.49058287555218516]
 ⋮
 [0.4700394061063937, 0.5440026379397457, 0.7411235688716618]
 [0.5, 0.5, 0.5]

If the output array has to be resized, that has to be done with the resize_output! function, which will keep the consistency of the auxiliary multi-threading buffers. This is, for instance, the case in the example of computation of forces, as the forces array must be of the same length as the array of positions:

julia> resize_output!(system, length(system.xpositions));

julia> map_pairwise!((x,y,i,j,d2,forces) -> update_forces!(x,y,i,j,d2,forces), system)
1001-element Vector{SVector{3, Float64}}:
 [756.2076075886971, -335.1637545330828, 541.8627090466914]
 [-173.02442398784672, -178.782819965489, 4.570607952876692]
 ⋮
 [-722.5400961501635, 182.65287417718935, 380.0394926753039]
 [20.27985502389337, -193.77607810950286, -155.28968519541544]

In this case, if the output is not resized, a BoundsError: is be obtained, because updates of forces at unavailable positions will be attempted.

Updating the unit cell

The unit cell can be updated to new dimensions at any moment, with the update_unitcell! function:

julia> update_unitcell!(system, SVector(1.2, 1.2, 1.2))
ParticleSystem1 of dimension 3, composed of:
    Box{OrthorhombicCell, 3}
      unit cell matrix = [ 1.2, 0.0, 0.0; 0.0, 1.2, 0.0; 0.0, 0.0, 1.2 ]
      cutoff = 0.1
      number of computing cells on each dimension = [13, 13, 13]
      computing cell sizes = [0.11, 0.11, 0.11] (lcell: 1)
      Total number of cells = 2197
    CellListMap.CellList{3, Float64}
      1000 real particles.
      623 cells with real particles.
      1719 particles in computing box, including images.
    Parallelization auxiliary data set for: 
      Number of batches for cell list construction: 8
      Number of batches for function mapping: 12
    Type of output variable (forces): Vector{SVector{3, Float64}}

Updating the cutoff

The cutoff can also be updated, using the update_cutoff! function:

julia> update_cutoff!(system, 0.2)
ParticleSystem1 of dimension 3, composed of:
    Box{OrthorhombicCell, 3}
      unit cell matrix = [ 1.0, 0.0, 0.0; 0.0, 1.0, 0.0; 0.0, 0.0, 1.0 ]
      cutoff = 0.2
      number of computing cells on each dimension = [7, 7, 7]
      computing cell sizes = [0.2, 0.2, 0.2] (lcell: 1)
      Total number of cells = 343
    CellListMap.CellList{3, Float64}
      1000 real particles.
      125 cells with real particles.
      2792 particles in computing box, including images.
    Parallelization auxiliary data set for: 
      Number of batches for cell list construction: 8
      Number of batches for function mapping: 8
    Type of output variable (forces): Vector{SVector{3, Float64}}

julia> map_pairwise!((x,y,i,j,d2,forces) -> update_forces!(x,y,i,j,d2,forces), system)
1000-element Vector{SVector{3, Float64}}:
 [306.9612911344924, -618.7375562535321, -607.1449767066479]
 [224.0803003775478, -241.05319348787023, 67.53780411933884]
 ⋮
 [2114.4873184508524, -3186.265279868732, -6777.748445712408]
 [-25.306486853608945, 119.69319481834582, 104.1501577339471]

Computations for two sets of particles

If the computation involves two sets of particle, a similar interface is available. The only difference is that the coordinates of the two sets must be provided to the ParticleSystem constructor as the xpositions and ypositions arrays.

We will illustrate this interface by computing the minimum distance between two sets of particles, which allows us to showcase further the definition of custom type interfaces:

First, we define a variable type that will carry the indexes and the distance of the closest pair of particles:

julia> struct MinimumDistance
           i::Int
           j::Int
           d::Float64
       end

The function that, given two particles, retains the minimum distance, is:

julia> function minimum_distance(i, j, d2, md)
           d = sqrt(d2)
           if d < md.d
               md = MinimumDistance(i, j, d)
           end
           return md
       end
minimum_distance (generic function with 1 method)

We overload copy, reset, and reduce functions, accordingly:

julia> import CellListMap: copy_output, reset_output!, reducer!

julia> copy_output(md::MinimumDistance) = md
copy_output (generic function with 5 methods)

julia> reset_output!(md::MinimumDistance) = MinimumDistance(0, 0, +Inf)
reset_output! (generic function with 5 methods)

julia> reducer!(md1::MinimumDistance, md2::MinimumDistance) = md1.d < md2.d ? md1 : md2
reducer! (generic function with 2 methods)

Note that since MinimumDistance is immutable, copying it is the same as returning the value. Also, resetting the minimum distance consists of setting its d field to +Inf. And, finally, reducing the threaded distances consists of keeping the pair with the shortest distance.

Next, we build the system

julia> xpositions = rand(SVector{3,Float64},1000);

julia> ypositions = rand(SVector{3,Float64},1000);

julia> system = ParticleSystem(
           xpositions = xpositions,
           ypositions = ypositions, 
           unitcell=[1.0,1.0,1.0], 
           cutoff = 0.1, 
           output = MinimumDistance(0,0,+Inf),
           output_name = :minimum_distance,
        )

And finally we can obtain the minimum distance between the sets:

julia> map_pairwise((x,y,i,j,d2,md) -> minimum_distance(i,j,d2,md), system)
MinimumDistance(276, 617, 0.006009804808785543)

Additional options

• Turn parallelization on and off
• Displaying a progress bar
• Fine control of the paralellization
• Avoid cell list updating
• Control CellList cell size
• Coordinates as matrices

Turn parallelization on and off

The use of parallel computations can be tunned on and of by the system.parallel boolean flag. For example, using 6 cores (12 threads) for the calculation of the minimum-distance example:

julia> f(system) = map_pairwise((x,y,i,j,d2,md) -> minimum_distance(i,j,d2,md), system)
f (generic function with 1 method)

julia> Threads.nthreads()
8

julia> system.parallel = true
true

julia> @btime f($system)
  268.265 μs (144 allocations: 16.91 KiB)
MinimumDistance(783, 497, 0.007213710914619913)

julia> system.parallel = false
false

julia> @btime f($system)
  720.304 μs (0 allocations: 0 bytes)
MinimumDistance(783, 497, 0.007213710914619913)

Displaying a progress bar

Displaying a progress bar: for very long runs, the user might want to see the progress of the computation. Use the show_progress keyword parameter of the map_pairwise! function for that.

For example, we execute the computation above, but with much more particles:

julia> xpositions = rand(SVector{3,Float64},10^6);

julia> ypositions = rand(SVector{3,Float64},10^6);

julia> system = ParticleSystem(
                  xpositions = xpositions,
                  ypositions = ypositions, 
                  unitcell=[1.0,1.0,1.0], 
                  cutoff = 0.1, 
                  output = MinimumDistance(0,0,+Inf),
                  output_name = :minimum_distance,
               );

julia> map_pairwise(
           (x,y,i,j,d2,md) -> minimum_distance(i,j,d2,md), system; 
           show_progress = true
       )
Progress:  24%|██████████▏                               |  ETA: 0:00:29

By activating the show_progress flag, a nice progress bar is shown.

Fine control of the paralellization

The number of batches launched in parallel runs can be tunned by the nbatches keyword parameter of the ParticleSystem constructor. By default, the number of batches is defined as heuristic function dependent on the number of particles, and possibly returns optimal values in most cases. For a detailed discussion about this parameter, see Number of batches.

For example, to set the number of batches for cell list calculation to 4 and the number of batches for mapping to 8, we can do:

julia> system = ParticleSystem(
           xpositions = rand(SVector{3,Float64},1000), 
           unitcell=[1,1,1], 
           cutoff = 0.1, 
           output = 0.0,
           output_name = :energy,
           nbatches=(4,8), # use this keyword 
       );

Most times it is expected that the default parameters are optimal. But particularly for inhomogeneous systems increasing the number of batches of the mapping phase (second parameter of the tuple) may improve the performance by reducing the idle time of threads.

Avoid cell list updating

To compute different properties without recomputing cell lists, use update_lists=false in the call of map_pairwise methods, for example,

using CellListMap, StaticArrays
system = ParticleSystem(xpositions=rand(SVector{3,Float64},1000), output=0.0, cutoff=0.1, unitcell=[1,1,1])
# First call, will compute the cell lists
map_pairwise((x,y,i,j,d2,u) -> u += d2, system)
# Second run: do not update the cell lists but compute a different property
map_pairwise((x,y,i,j,d2,u) -> u += sqrt(d2), system; update_lists = false)

in which case we are computing the sum of distances from the same cell lists used to compute the energy in the previous example (requires version 0.8.9). Specifically, this will skip the updating of the cell lists, thus be careful to not use this option if the cutoff, unitcell, or any other property of the system changed.

For systems with two sets of particles, the coordinates of the xpositions set can be updated, preserving the cell lists computed for the ypositions, but this requires setting autoswap=false in the construction of the ParticleSystem:

using CellListMap, StaticArrays
system = ParticleSystem(
    xpositions=rand(SVector{3,Float64},1000), 
    ypositions=rand(SVector{3,Float64},2000),
    output=0.0, cutoff=0.1, unitcell=[1,1,1],
    autoswap=false # Cell lists are constructed for ypositions
)
map_pairwise((x,y,i,j,d2,u) -> u += d2, system)
# Second run: preserve the cell lists but compute a different property
map_pairwise((x,y,i,j,d2,u) -> u += sqrt(d2), system; update_lists = false)

Control CellList cell size

The cell sizes of the construction of the cell lists can be controled with the keyword lcell of the ParticleSystem constructor. For example:

julia> system = ParticleSystem(
           xpositions = rand(SVector{3,Float64},1000), 
           unitcell=[1,1,1], 
           cutoff = 0.1, 
           output = 0.0,
           output_name = :energy,
           lcell=2,
       );

Most times using lcell=1 (default) or lcell=2 will provide the optimal performance. For very dense systems, or systems for which the number of particles within the cutoff is very large, larger values of lcell may improve the performance. To be tested by the user.

Coordinates as matrices

Coordinates can also be provided as matrices of size (D,N) where D is the dimension (2 or 3) and N is the number of particles. For example:

julia> using CellListMap

julia> system = ParticleSystem(
           xpositions=rand(2,100),
           ypositions=rand(2,200),
           cutoff=0.1,
           unitcell=[1,1],
           output=0.0,
       )
ParticleSystem2{output} of dimension 2, composed of:
    Box{OrthorhombicCell, 2}
      unit cell matrix = [ 1.0 0.0; 0.0 1.0 ]
      cutoff = 0.1
      number of computing cells on each dimension = [13, 13]
      computing cell sizes = [0.1, 0.1] (lcell: 1)
      Total number of cells = 169
    CellListMap.CellListPair{Vector{StaticArraysCore.SVector{2, Float64}}, 2, Float64, CellListMap.Swapped}
       200 particles in the reference vector.
       61 cells with real particles of target vector.
    Parallelization auxiliary data set for:
      Number of batches for cell list construction: 1
      Number of batches for function mapping: 1
    Type of output variable (output): Float64

Complete example codes

• Simple energy computation
• Force computation
• Energy and forces
• Two sets of particles
• Particle simulation

Simple energy computation

In this example, a simple potential energy defined as the sum of the inverse of the distance between the particles is computed.

using CellListMap
using StaticArrays
system = ParticleSystem(
    xpositions = rand(SVector{3,Float64},1000), 
    unitcell=[1.0,1.0,1.0], 
    cutoff = 0.1, 
    output = 0.0,
    output_name = :energy
)
map_pairwise!((x,y,i,j,d2,energy) -> energy += 1 / sqrt(d2), system)

Force computation

Here we compute the force vector associated to the potential energy function of the previous example.

using CellListMap
using StaticArrays
positions = rand(SVector{3,Float64},1000) 
system = ParticleSystem(
    xpositions = positions, 
    unitcell=[1.0,1.0,1.0], 
    cutoff = 0.1, 
    output = similar(positions),
    output_name = :forces
)
function update_forces!(x,y,i,j,d2,forces)
    d = sqrt(d2)
    df = (1/d2)*(1/d)*(y - x)
    forces[i] += df
    forces[j] -= df
    return forces
end
map_pairwise!((x,y,i,j,d2,forces) -> update_forces!(x,y,i,j,d2,forces), system)

Energy and forces

In this example, the potential energy and the forces are computed in a single run, and a custom data structure is defined to store both values.

using CellListMap
using StaticArrays
# Define custom type
mutable struct EnergyAndForces
    energy::Float64
    forces::Vector{SVector{3,Float64}}
end
# Custom copy, reset and reducer functions
import CellListMap: copy_output, reset_output!, reducer
copy_output(x::EnergyAndForces) = EnergyAndForces(copy(x.energy), copy(x.forces))
function reset_output!(output::EnergyAndForces)
    output.energy = 0.0
    for i in eachindex(output.forces)
        output.forces[i] = SVector(0.0, 0.0, 0.0)
    end
    return output
end
function reducer(x::EnergyAndForces, y::EnergyAndForces)
    e_tot = x.energy + y.energy
    x.forces .+= y.forces
    return EnergyAndForces(e_tot, x.forces)
end
# Function that updates energy and forces for each pair
function energy_and_forces!(x,y,i,j,d2,output::EnergyAndForces)
    d = sqrt(d2)
    output.energy += 1/d
    df = (1/d2)*(1/d)*(y - x)
    output.forces[i] += df
    output.forces[j] -= df
    return output
end
# Initialize system
positions = rand(SVector{3,Float64},1000);
system = ParticleSystem(
    xpositions = positions,
    unitcell=[1.0,1.0,1.0], 
    cutoff = 0.1, 
    output = EnergyAndForces(0.0, similar(positions)),
    output_name = :energy_and_forces
)
# Compute energy and forces
map_pairwise((x,y,i,j,d2,output) -> energy_and_forces!(x,y,i,j,d2,output), system)

Two sets of particles

In this example we illustrate the interface for the computation of properties of two sets of particles, by computing the minimum distance between the two sets.

using CellListMap
using StaticArrays
# Custom structure to store the minimum distance pair
struct MinimumDistance
    i::Int
    j::Int
    d::Float64
end
# Function that updates the minimum distance found
function minimum_distance(i, j, d2, md)
    d = sqrt(d2)
    if d < md.d
        md = MinimumDistance(i, j, d)
    end
    return md
end
# Define appropriate methods for copy, reset and reduce 
import CellListMap: copy_output, reset_output!, reducer!
copy_output(md::MinimumDistance) = md
reset_output!(md::MinimumDistance) = MinimumDistance(0, 0, +Inf)
reducer!(md1::MinimumDistance, md2::MinimumDistance) = md1.d < md2.d ? md1 : md2
# Build system 
xpositions = rand(SVector{3,Float64},1000);
ypositions = rand(SVector{3,Float64},1000);
system = ParticleSystem(
       xpositions = xpositions,
       ypositions = ypositions, 
       unitcell=[1.0,1.0,1.0], 
       cutoff = 0.1, 
       output = MinimumDistance(0,0,+Inf),
       output_name = :minimum_distance,
)
# Compute the minimum distance
map_pairwise((x,y,i,j,d2,md) -> minimum_distance(i,j,d2,md), system)

Particle simulation

In this example, a complete particle simulation is illustrated, with a simple potential. This example can illustrate how particle positions and forces can be updated. Run this simulation with:

julia> system = init_system(N=200); # number of particles

julia> trajectory = simulate(system);

julia> animate(trajectory)

One important characteristic of this example is that the system is built outside the function that performs the simulation. This is done because the construction of the system is type-unstable (it is dimension, geometry and output-type dependent). Adding a function barrier avoids type-instabilities to propagate to the simulation causing possible performance problems.

using StaticArrays
using CellListMap
import CellListMap.wrap_relative_to
# Function that updates the forces, for potential of the form:
# if d < cutoff k*(d^2-cutoff^2)^2 else 0.0 with k = 10^6
function update_forces!(x, y, i, j, d2, forces, cutoff)
    r = y - x
    dudr = 10^6 * 4 * r * (d2 - cutoff^2)
    forces[i] += dudr
    forces[j] -= dudr
    return forces
end
# Function that initializes the system: it is preferable to initialize
# the system outside the function that performs the simulation, because
# the system (data)type is defined on initialization. Initializing it outside
# the simulation function avoids possible type-instabilities. 
function init_system(;N::Int=200)
    Vec2D = SVector{2,Float64}
    positions = rand(Vec2D, N)
    unitcell = [1.0, 1.0]
    cutoff = 0.1
    system = ParticleSystem(
        positions=positions,
        cutoff=cutoff,
        unitcell=unitcell,
        output=similar(positions),
        output_name=:forces,
    )
    return system
end
function simulate(system=init_system(); nsteps::Int=100, isave=1)
    # initial velocities
    velocities = [ randn(eltype(system.positions)) for _ in 1:length(system.positions) ]
    dt = 1e-3
    trajectory = typeof(system.positions)[]
    for step in 1:nsteps
        # compute forces at this step
        map_pairwise!(
            (x,y,i,j,d2,forces) -> update_forces!(x,y,i,j,d2,forces,system.cutoff),
            system
        )
        # Update positions and velocities
        for i in eachindex(system.positions, system.forces)
            f = system.forces[i]
            x = system.positions[i]
            v = velocities[i]
            x = x + v * dt + (f / 2) * dt^2
            v = v + f * dt
            # wrapping to origin for obtaining a pretty animation
            x = wrap_relative_to(x, SVector(0.0, 0.0), system.unitcell)
            # !!! IMPORTANT: Update arrays of positions and velocities
            system.positions[i] = x
            velocities[i] = v
        end
        # Save step for printing
        if step % isave == 0
            push!(trajectory, copy(system.positions))
        end
    end
    return trajectory
end

using Plots
function animate(trajectory)
    anim = @animate for step in trajectory
        scatter(
            Tuple.(step),
            label=nothing,
            lims=(-0.5, 0.5),
            aspect_ratio=1,
            framestyle=:box,
        )
    end
    gif(anim, "simulation.gif", fps=10)
end

Docstrings

ParticleSystem(;
    xpositions::Union{AbstractVector{<:AbstractVector},AbstractMatrix},
    #or
    xpositions::Union{AbstractVector{<:AbstractVector},AbstractMatrix},
    ypositions::Union{AbstractVector{<:AbstractVector},AbstractMatrix},
    # and
    unitcell::Union{Nothing,AbstractVecOrMat} = nothing,
    cutoff::Number,
    output::Any;
    output_name::Symbol,
    parallel::Bool=true,
    nbatches::Tuple{Int,Int}=(0, 0),
    autoswap::Bool = true
)

julia> using CellListMap

julia> using PDBTools: readPDB, coor

julia> positions = coor(readPDB(CellListMap.argon_pdb_file));

julia> sys = ParticleSystem(
           xpositions = positions, 
           unitcell = [21.0, 21.0, 21.0],
           cutoff = 8.0, 
           output = 0.0, 
        );

julia> map_pairwise!((x,y,i,j,d2,output) -> output += d2, sys)
43774.54367600001

julia> using CellListMap, PDBTools

julia> xpositions = coor(readPDB(CellListMap.argon_pdb_file))[1:50];

julia> ypositions = coor(readPDB(CellListMap.argon_pdb_file))[51:100];

julia> sys = ParticleSystem(
           xpositions = xpositions, 
           ypositions = ypositions, 
           unitcell = [21.0, 21.0, 21.0],
           cutoff = 8.0, 
           output = 0.0, 
           parallel = false, # use true for parallelization
        );

julia> map_pairwise!((x,y,i,j,d2,output) -> output += d2, sys)
21886.196785000004

copy_output(x)

using CellListMap
# Custom data type
struct A x::Int end
# Custom output type (array of A)
output = [ A(0) for _ in 1:100 ]
# How to copy an array of `A`
CellListMap.copy_output(v::Vector{A}) = [ x for x in v ]

# Alternativelly, in this case, one could have defined:
Base.copy(a::A) = a
CellListMap.copy_output(v::Vector{A}) = copy(v)

CellListMap.copy_output(v::Vector{T}) where {T<:CustomType} = deepcopy(v)

map_pairwise!(
    f::Function, system::AbstractParticleSystem; 
    show_progress = true, update_lists = true
)

function f(x,y,i,j,d2,output)
    # operate on particle coordinates, distance and indexes
    # update output
    return output
end

julia> sys = ParticleSystem(
           xpositions = rand(SVector{3,Float64},1000), 
           unitcell=[1,1,1], 
           cutoff = 0.1, 
           output = 0.0
           );

julia> map_pairwise((x,y,i,j,d2,output) -> output += 1 / (1 + sqrt(d2)), sys)
1870.0274887950268

reducer(x,y)
reducer!(x,y)

reducer(x,y) = +(x,y)

struct MinimumDistance d::Float64 end
reducer(x::MinimumDistance, y::MinimumDistance) = MinimumDistance(min(x.d, y.d))

julia> using CellListMap, PDBTools

julia> positions = coor(readPDB(CellListMap.argon_pdb_file));

julia> struct MinimumDistance d::Float64 end # Custom output type

julia> CellListMap.copy_output(d::MinimumDistance) = MinimumDistance(d.d) # Custom copy function for `Out`

julia> CellListMap.reset_output(d::MinimumDistance) = MinimumDistance(+Inf) # How to reset an array with elements of type `MinimumDistance`

julia> CellListMap.reducer(md1::MinimumDistance, md2::MinimumDistance) = MinimumDistance(min(md1.d, md2.d)) # Custom reduction function

julia> # Construct the system
       sys = ParticleSystem(;
           positions = positions,
           unitcell = [21,21,21],
           cutoff = 8.0,
           output = MinimumDistance(+Inf),
       );

julia> # Obtain the minimum distance between atoms:
       map_pairwise!((x,y,i,j,d2,output) -> sqrt(d2) < output.d ? MinimumDistance(sqrt(d2)) : output, sys)
MinimumDistance(2.1991993997816563)

reset_output(x)
reset_output!(x)

julia> using CellListMap

julia> struct MinimumDistance d::Float64 end

julia> CellListMap.reset_output(x::MinimumDistance) = MinimumDistance(+Inf)

julia> x = MinimumDistance(1.0)
MinimumDistance(1.0)

julia> CellListMap.reset_output(x)
MinimumDistance(Inf)

resize_output!(sys::AbstractParticleSystem, n::Int)

unitcelltype(sys::AbstractParticleSystem)

update_cutoff!(system, cutoff)

julia> using CellListMap, PDBTools

julia> x = coor(readPDB(CellListMap.argon_pdb_file));

julia> sys = ParticleSystem(
           xpositions = x, 
           unitcell=[21.0,21.0,21.0], 
           cutoff = 8.0, 
           output = 0.0
       );

julia> update_cutoff!(sys, 10.0)
ParticleSystem1{output} of dimension 3, composed of:
    Box{OrthorhombicCell, 3}
      unit cell matrix = [ 21.0 0.0 0.0; 0.0 21.0 0.0; 0.0 0.0 21.0 ]
      cutoff = 10.0
      number of computing cells on each dimension = [5, 5, 5]
      computing cell sizes = [10.5, 10.5, 10.5] (lcell: 1)
      Total number of cells = 125
    CellList{3, Float64}
      100 real particles.
      8 cells with real particles.
      800 particles in computing box, including images.
    Parallelization auxiliary data set for:
      Number of batches for cell list construction: 8
      Number of batches for function mapping: 8
    Type of output variable (output): Float64

update_unitcell!(system, unitcell)

julia> using CellListMap, StaticArrays, PDBTools

julia> xpositions = coor(readPDB(CellListMap.argon_pdb_file));

julia> sys = ParticleSystem(
           xpositions = xpositions,
           unitcell=[21,21,21], 
           cutoff = 8.0, 
           output = 0.0
       );

julia> update_unitcell!(sys, [30.0, 30.0, 30.0])
ParticleSystem1{output} of dimension 3, composed of:
    Box{OrthorhombicCell, 3}
      unit cell matrix = [ 30.0 0.0 0.0; 0.0 30.0 0.0; 0.0 0.0 30.0 ]
      cutoff = 8.0
      number of computing cells on each dimension = [6, 6, 6]
      computing cell sizes = [10.0, 10.0, 10.0] (lcell: 1)
      Total number of cells = 216
    CellList{3, Float64}
      100 real particles.
      8 cells with real particles.
      800 particles in computing box, including images.
    Parallelization auxiliary data set for:
      Number of batches for cell list construction: 1
      Number of batches for function mapping: 1
    Type of output variable (output): Float64

mutable struct ParticleSystem1{OutputName, V, O, B, C, A} <: CellListMap.AbstractParticleSystem{OutputName}

mutable struct ParticleSystem2{OutputName, V, O, B, C, A} <: CellListMap.AbstractParticleSystem{OutputName}