Lagrangian particle tracking

Models can keep track of the location and properties of neutrally buoyant particles. Particles are advected with the flow field using forward Euler time-stepping at every model iteration.

Simple particles

If you just need to keep of particle locations $(x, y, z)$ then you can construct some Lagrangian particles using the regular LagrangianParticles constructor

grid = RectilinearGrid(size=(10, 10, 10), extent=(1, 1, 1))

Nparticles = 10

x₀ = zeros(Nparticles)

y₀ = rand(Nparticles)

z₀ = -0.5 * ones(Nparticles)

lagrangian_particles = LagrangianParticles(x=x₀, y=y₀, z=z₀)

# output
10 LagrangianParticles with eltype Particle:
├── 3 properties: (:x, :y, :z)
├── particle-wall restitution coefficient: 1.0
├── 0 tracked fields: ()
└── dynamics: no_dynamics

then pass it to a model constructor

model = NonhydrostaticModel(grid=grid, particles=lagrangian_particles)

# output
NonhydrostaticModel{CPU, RectilinearGrid}(time = 0 seconds, iteration = 0)
├── grid: 10×10×10 RectilinearGrid{Float64, Periodic, Periodic, Bounded} on CPU with 3×3×3 halo
├── timestepper: RungeKutta3TimeStepper
├── advection scheme: Centered(order=2)
├── tracers: ()
├── closure: Nothing
├── buoyancy: Nothing
├── coriolis: Nothing
└── particles: 10 LagrangianParticles with eltype Particle and properties (:x, :y, :z)
Lagrangian particles on GPUs

Remember to use CuArray instead of regular Array when storing particle locations and properties on the GPU.

Custom particles

If you want to keep track of custom properties, such as the species or DNA of a Lagrangian particle representing a microbe in an agent-based model, then you can create your own custom particle type and pass a StructArray to the LagrangianParticles constructor.

using Oceananigans
using StructArrays

struct LagrangianMicrobe{T, S, D}
    x :: T
    y :: T
    z :: T
    species :: S
    dna :: D
end

Nparticles = 3

x₀ = zeros(Nparticles)

y₀ = rand(Nparticles)

z₀ = -0.5 * ones(Nparticles)

species = [:rock, :paper, :scissors]

dna = ["TATACCCC", "CCTAGGAC", "CGATTTAA"]

particles = StructArray{LagrangianMicrobe}((x₀, y₀, z₀, species, dna));

lagrangian_particles = LagrangianParticles(particles)

# output
3 LagrangianParticles with eltype LagrangianMicrobe:
├── 5 properties: (:x, :y, :z, :species, :dna)
├── particle-wall restitution coefficient: 1.0
├── 0 tracked fields: ()
└── dynamics: no_dynamics
Custom properties on GPUs

Not all data types can be passed to GPU kernels. If you intend to advect particles on the GPU make sure particle properties consist of only simple data types. The symbols and strings in this example won't work on the GPU.

Writing particle properties to disk

Particle properties can be written to disk using JLD2 or NetCDF.

When writing to JLD2 you can pass model.particles as part of the named tuple of outputs.

JLD2OutputWriter(model, (; particles=model.particles), filename="particles", schedule=TimeInterval(15))
JLD2OutputWriter scheduled on TimeInterval(15 seconds):
├── filepath: build/model_setup/particles.jld2
├── 1 outputs: particles
├── array type: Array{Float64}
├── including: [:grid, :coriolis, :buoyancy, :closure]
├── file_splitting: NoFileSplitting
└── file size: 19.4 KiB

When writing to NetCDF you should write particles to a separate file as the NetCDF dimensions differ for particle trajectories. You can just pass model.particles straight to NetCDFOutputWriter:

NetCDFOutputWriter(model, model.particles, filename="particles.nc", schedule=TimeInterval(15))
NetCDFOutputWriter scheduled on TimeInterval(15 seconds):
├── filepath: build/model_setup/particles.nc
├── dimensions: particle_id(3), time(0)
├── 1 outputs: particles
└── array type: Array{Float64}
├── file_splitting: NoFileSplitting
└── file size: 10.2 KiB
Outputting custom particle properties to NetCDF

NetCDF does not support arbitrary data types. If you need to write custom particle properties to disk that are not supported by NetCDF then you should use JLD2 (which should support almost any Julia data type).