Ecosystem integration
Agents.jl
Agents.jl provides a comprehensive framework for simulation, analysis and visualization of agent-based systems. CellListMap
can be used to accelerate these simulations, and the integration of the packages is rather simple, particularly using the ParticleSystem
interface. A complete integration example can be obtained in the Agents
documentation (currently at the development branch).
The example will produce the following animation:
Unitful and units
The functions of CellListMap.jl support the propagation of generic (isbits) types, and thus units and thus automatic differentiation and the use of Unitful
. A set of working examples can be found in the generic_types.jl file.
We start illustrating the support for unit propagation. We need to define all involved quantities in the same units:
Using the ParticleSystem interface
The only requirement is to attach proper units to all quantities (positions, cutoff, unitcell, and output variables). Here we compute the square of the distances of the particles within the cutoff, where the particle coordinates are in Angstroms, while the box size and cutoff are defined in nanometers:
julia> using CellListMap, Unitful, PDBTools
julia> positions = coor(readPDB(CellListMap.argon_pdb_file))u"Å";
julia> system = ParticleSystem(
positions = positions,
cutoff = 0.8u"nm",
unitcell = [2.1,2.1,2.1]u"nm",
output = 0.0u"nm^2",
output_name = :sum_sqr
);
julia> map_pairwise((x,y,i,j,d2,out) -> out += d2, system)
437.74543675999995 nm^2
Units in neighbor lists
CellListMap.neighborlist
also propagates units correctly:
julia> using CellListMap, Unitful, PDBTools
julia> positions = coor(readPDB(CellListMap.argon_pdb_file))u"Å";
julia> cutoff = 0.8u"nm";
julia> neighborlist(positions, cutoff)
857-element Vector{Tuple{Int64, Int64, Quantity{Float64, 𝐋, Unitful.FreeUnits{(nm,), 𝐋, nothing}}}}:
(1, 20, 0.3163779543520692 nm)
(1, 61, 0.408865185605231 nm)
(1, 67, 0.5939772807102979 nm)
(1, 80, 0.24572289270639777 nm)
(1, 94, 0.5394713986857874 nm)
(13, 15, 0.2678764267344179 nm)
(13, 41, 0.4408015539900013 nm)
(13, 44, 0.6960112211739119 nm)
(13, 61, 0.5939197673086826 nm)
(13, 64, 0.45607558584076857 nm)
⋮
(46, 18, 0.6114385414741209 nm)
(46, 51, 0.799947279512844 nm)
(51, 68, 0.22003574709578452 nm)
(51, 22, 0.663802094000915 nm)
(54, 45, 0.44233083772217385 nm)
(73, 78, 0.2853611220891873 nm)
(73, 88, 0.6078711047582371 nm)
(78, 88, 0.7006116541993859 nm)
(88, 54, 0.7933654076149276 nm)
Automatic differentiation
Allowing automatic differentiation follows the same principles, meaning that we only need to allow the propagation of dual types through the computation by proper initialization of the input data. However, it is easier to work with the low level interface, which accepts matrices as the input for positions and a more fine control of the types of the variables. Matrices are easier input types for auto diff packages.
The variables are each component of each vector, thus the easiest way to represent the points to interface with differentiation packages is providing the coordinates as a matrix:
julia> x = rand(3,1000)
3×1000 Matrix{Float64}:
0.186744 0.328719 0.874102 0.503535 … 0.328161 0.0895699 0.917338
0.176157 0.972954 0.80729 0.624724 0.655268 0.470754 0.327578
0.648482 0.537362 0.599624 0.0688776 0.92333 0.497984 0.208924
The key here is allow all the types of the parameters to follow the type propagation of the elements of x
inside the differentiation routine. The function we define to compute the derivative is, then:
julia> function sum_sqr(x, sides, cutoff)
sys = ParticleSystem(
positions=x,
unitcell=eltype(x).(sides),
cutoff=eltype(x).(cutoff),
output=zero(eltype(x))
)
return map_pairwise((_, _, _, _, d2, sum_sqr) -> sum_sqr += d2, sys)
end
Note that we convert cutoff
and sides
to the same type of the input x
of the function, and set the type of the output
variable accordingly. For a simple call to the function this is inconsequential:
julia> cutoff = 0.1; sides = [1,1,1];
julia> sum_sqr(x,sides,cutoff)
12.897650398753228
but the conversion is required to allow the differentiation to take place:
julia> ForwardDiff.gradient(x -> sum_sqr(x,sides,cutoff),x)
3×1000 Matrix{Float64}:
-0.132567 0.029865 -0.101301 … 0.249267 0.0486424 -0.0400487
0.122421 0.207495 -0.184366 -0.201648 -0.105031 0.218342
0.0856502 0.288924 0.122445 -0.0147022 -0.103314 -0.0862264
Measurements
Propagating uncertainties through the Measurements
and other similar packages requires a different strategy, because within CellListMap
only isbits
types can be used, which is not the case of the type Measurement
type.
In cases like this, it is better to bypass all the internals of CellListMap
and provide the data to the function that computes pairwise properties directly as a closure. For example:
A vector of particles with uncertainties in their coordinates can be created with:
julia> using StaticArrays
julia> x_input = [ SVector{3}(measurement(rand(),0.01*rand()) for i in 1:3) for j in 1:1000 ]
1000-element Vector{SVector{3, Measurement{Float64}}}:
[0.1658 ± 0.003, 0.9951 ± 0.0054, 0.5067 ± 0.0035]
[0.2295 ± 0.0074, 0.2987 ± 0.0021, 0.42828 ± 0.00099]
⋮
[0.1362 ± 0.0034, 0.2219 ± 0.0048, 0.2119 ± 0.0072]
[0.2521 ± 0.0038, 0.4988 ± 0.00013, 0.856046 ± 4.3e-5]
The variables within the CellListMap
functions will be stripped from the uncertainties. We do:
julia> unitcell = [1,1,1]
julia> cutoff = 0.1; box = Box(unitcell,cutoff);
julia> x_strip = [ getproperty.(v,:val) for v in x_input ]
1000-element Vector{SVector{3, Float64}}:
[0.08441931492362276, 0.9911530546181084, 0.07408559584648788]
[0.12084764467339837, 0.8284551316333133, 0.9021906852432111]
⋮
[0.2418752113326077, 0.4429225751775432, 0.13576355747772784]
[0.24440380524702654, 0.07148275176890073, 0.26722687487212315]
The cell list is built with the stripped values:
julia> cl = CellList(x_strip,box)
CellList{3, Float64}
1000 real particles.
637 cells with real particles.
1695 particles in computing box, including images.
The result is initialized with the proper type,
julia> result = measurement(0.,0.)
0.0 ± 0.0
and the mapping is performed with the stripped coordinates, but passing the values with uncertainties to the mapped function, which will perform the computation on the pairs with those values:
julia> using LinearAlgebra: norm_sqr
julia> result = map_pairwise!(
(xᵢ,xⱼ,i,j,d2,sum_sqr) -> begin
x1 = x_input[i]
x2 = CellListMap.wrap_relative_to(x_input[j],x1,unitcell)
sum_sqr += norm_sqr(x2-x1)
return sum_sqr
end,
result, box, cl
)
13.14 ± 0.061
In the function above, the xᵢ
and xⱼ
coordinates, which correspond to the coordinates in x_input[i]
and x_input[j]
, but already wrapped relative to each other, are ignored, because they don't carry the uncertainties. We use only the indexes i
and j
to recompute the relative position of the particles according to the periodic boundary conditions (using the CellListMap.wrap_relative_to
function) and their (squared) distance. Since the x_input
array carries the uncertainties, the computation of sum_sqr
will propagate them.
All these computations should be performed inside the scope of a function for optimal performance. The examples here can be followed by copying and pasting the code into the REPL, but this is not the recommended practice for critical code. The strategy of bypassing the internal computations of CellListMap
may be useful for improving performance even if the previous and simpler method is possible.