Eulerian 3-D dispersion (L2) — derivation and numerical scheme

Eulerian 3-D dispersion: derivation and finite-volume discretisation¶

This note derives the equations solved by plume_simulation.les_fvm and explains how the terms map onto the finitevolX Arakawa C-grid operators. The L2 model sits between the analytic Gaussian puff (L1, plume_simulation.gauss_puff) and a full resolved-flow LES (L3, future port): it evolves a passive methane tracer on a 3-D Cartesian grid under a prescribed wind field with a K-theory (gradient-diffusion) closure. No turbulence is resolved; the model pays for full 3-D transport with spatially-varying wind and physically realistic boundaries (ground, inlet, outlet) that the Gaussian puff cannot represent.

1. Governing equation¶

We solve the conservative advection-diffusion equation for a passive scalar $C(\vec x, t)$ [kg/m³]:

with prescribed 3-D wind $\vec u(\vec x, t) = (u, v, w)$, diagonal eddy diffusivity $\mathbf K = \operatorname{diag}(K_h, K_h, K_z)$ (horizontally isotropic), and source $S$. Equation (2) is Stockie equation (2.2) ^[1] without the A3 / A4 / A5 simplifications of the Gaussian plume — we keep every transport term and impose physically meaningful boundaries instead.

2. Closure assumptions¶

• (E1) Prescribed flow. $\vec u(\vec x, t)$ is an external input (no momentum equation solved). The wind need not be divergence-free — incompressibility is not enforced, since the model treats $\vec u$ as a kinematic input; breaking $\nabla \cdot \vec u = 0$ only affects the tracer continuity slightly and is typically $< 10\%$ for realistic atmospheric inputs.
• (E2) K-theory. Turbulent fluxes are modelled by the gradient closure $\overline{u'C'} = -\mathbf K \nabla C$, with $\mathbf K$ specified directly (as a scalar, anisotropic pair $(K_h, K_z)$, or field) or derived from Pasquill-Gifford σ curves via the Taylor identity $K = \sigma^2 / (2t)$ — see les_fvm.diffusion.pg_eddy_diffusivity.
• (E3) Boundary conditions. Unlike the Gaussian puff (unbounded, ground reflection by method of images), L2 lives in a bounded box with user-chosen BCs per face. The defaults are Dirichlet-clean on the inlet, outflow on the downstream face, periodic laterally, and Neumann (no flux) on the ground and top.
• (E4) Passive scalar. No buoyancy, decay, or chemistry; $S$ is a Gaussian-regularised point source with user-specified constant or time-varying emission rate.

3. Spatial discretisation¶

3.1 Arakawa C-grid¶

Fields live on a Cartesian 3-D Arakawa C-grid (finitevolx.CartesianGrid3D). The horizontal stagger is classical:

• $C, w, K$ at T-points (cell centres);
• $u$ at U-points (east faces);
• $v$ at V-points (north faces).

The vertical is collocated at T-points — finitevolX’s 3-D operators treat $z$ as a batch axis, so there is no vertical staggering inside a field array. Arrays have shape [Nz, Ny, Nx] with one ghost cell per axis for boundary conditions; the interior spans [1:-1, 1:-1, 1:-1].

3.2 Advection term¶

The divergence of the advective flux splits cleanly into horizontal and vertical parts:

The horizontal part — two terms per z-level, always acting in the (y, x) plane — is delegated to finitevolx.Advection3D, which applies a WENO5 (default) or user-chosen reconstruction at U- and V-faces for every z-level. The vertical part is built by les_fvm._vertical_ops.vertical_advection_tendency as a first-order upwind flux at k-faces:

with $w_{k+1/2} = \tfrac12(w_k + w_{k+1})$. The tendency at interior T-point $k$ is then $-(\widehat{F}_{k+1/2} - \widehat{F}_{k-1/2}) / \Delta z$. First-order upwind is monotone, which matters for a non-negative tracer; vertical resolution in a plume simulation is typically too coarse for a higher-order scheme to pay off anyway.

3.3 Diffusion term¶

The K-theory flux $-\mathbf K \nabla C$ similarly splits:

The horizontal Laplacian (first two terms) is computed per z-level by finitevolx.Diffusion3D. The vertical part is assembled by _vertical_ops.vertical_diffusion_tendency with face-averaged diffusivity and central finite differences on a uniform grid:

For uniform $K_z$ this collapses to the standard three-point stencil $(C_{k+1} - 2 C_k + C_{k-1})/\Delta z^2$.

3.4 Source term¶

The regularised point source writes

with $Z$ chosen so that $\int \rho_s\, \mathrm d V = 1$ over the interior (discrete cell-sum $\times \Delta x \Delta y \Delta z$). The default radius is $r_s = 2 \max(\Delta x, \Delta y, \Delta z)$ — wide enough to avoid single-cell aliasing on coarse grids. A time-varying rate $q(t)$ is supplied as a JAX-compatible callable; a scalar rate is treated as a closure that discards $t$. See les_fvm.source.make_gaussian_source.

4. Boundary conditions¶

finitevolX ships 2-D BC atoms (Dirichlet1D, Neumann1D, Outflow1D, Periodic1D, ...) that update a ghost ring on a [Ny, Nx] field. Applying them to a 3-D field needs two composable helpers that live in les_fvm.boundary:

1. HorizontalBC — wraps a finitevolx.BoundaryConditionSet and vectorises it with eqx.filter_vmap over the z-axis, so every z-slice receives the same per-face BC update.
2. VerticalBC — directly updates the [0, :, :] and [-1, :, :] ghost slices with {Dirichlet, Neumann, Outflow, Periodic} flavours.

The default concentration BCs are:

• x-axis: Dirichlet with value 0 upstream (clean inlet) / Outflow downstream (zero-gradient).
• y-axis: Periodic laterally (standard assumption for plume experiments in a homogeneous transect).
• z-axis: zero-Neumann at the ground and top (no flux through solid walls).

These are all overridable at the simulate_eulerian_dispersion call site.

5. Time integration¶

The tendency is assembled in les_fvm.dynamics.EulerianDispersionRHS and wrapped in a diffrax.ODETerm. The default solver is diffrax.Tsit5 (adaptive 5th-order Runge-Kutta) with PID step control; fixed-step alternatives (diffrax.Heun, finitevolx.RK3SSP) are available for benchmarking. Because the RHS is pure, monolithically JIT-compiled, and operates entirely on JAX arrays, the entire simulation — grid construction, wind evaluation, advection-diffusion tendencies, BC application, and source injection — runs inside a single compiled graph.

6. CFL and stability¶

The explicit scheme is CFL-limited by the fastest wave in the system:

with a safety factor around 0.4 for WENO5. Tsit5’s adaptive controller in practice selects $\Delta t$ that stays well inside this bound, but users running fixed-step (Heun / RK3SSP) should set dt0 explicitly.

7. Cross-check against gauss_puff¶

Under spatially-uniform wind + zero vertical velocity + PG-calibrated diffusivity at a reference distance matching the observation point, the L2 solution should agree with the Gaussian puff centreline concentration to within discretisation error. Notebook Puff vs Eulerian runs this cross-check.