Skip to content

advection-diffusion2d

English | Japanese

« Home · « Advection-diffusion, reaction and radiation

Advection-diffusion of a passive scalar (2D Gaussian pulse) -- exact-solution verification benchmark. A single D2Q5 advection-diffusion LBM for the scalar C, advected by a PRESCRIBED UNIFORM velocity u = (U, 0) on a doubly-periodic square. The scalar is passive: there is no momentum LBM, no buoyancy coupling -- u is imposed everywhere, so this isolates the D2Q5 collision/streaming as a pure advection-diffusion operator and checks it against the textbook Gaussian Green's function.

Geometry

advection-diffusion2d geometry

Animation

Fields

The fields this computation wrote at its last output step, plotted as they are. 3D cases show the mid-plane.

advection-diffusion2d C

Results

A pure transport benchmark that validates a D2Q5 scalar LBM against an exact solution (the 2D Gaussian kernel). Default arguments: N = 128, lattice advection velocity U = 0.05, D_phys = 0.0064, T = 0.5, σ0 = 0.08. The mapping is D_lat = 0.0064 × 0.05 × 128 = 0.040960, τ = 3 × 0.04096 + 1/2 = 0.62288, steps = 0.5/(0.05/128) = 1280, and the Pe based on the pulse width is Pe = U σ0N/D_lat = 0.05 × 10.24/0.04096 = 12.5 — every number in the 1st line of the log closes with this arithmetic. Wall-clock time 1.11 seconds on 72 Grace CPU cores.

  • L2 error first decreases from 1.7820e-03 (step 25) (minimum 1.3132e-03, around step 325) and ends at 1.8139e-03. Over the whole run it stays in the 1.3–1.8e-3 band.
  • Growth of the variance: final σ² = 209.041 vs exact 209.715. The exact value is σ0² + 2 D_lat t = 104.8576 + 2 × 0.04096 × 1280 = 209.7152 — the run stops exactly at the time when the initial variance has doubled. The diffusion slope of the linear fit is 8.162253e-02, against the theoretical 2D = 2 × 0.04096 = 0.08192, i.e. −0.36%. Intercept 104.5727 vs σ0² = 104.8576.
  • Advection velocity: centroid tracking gives 5.000004e-02 vs U = 0.05 (0.00%). Translation on the lattice is reproduced almost exactly by D2Q5.
  • Conservation: M/M0 = 1.000000 at every step. Evidence that both collision and streaming conserve the total scalar to machine precision.

Governing equations

The advection–diffusion equation for a passive scalar C and its Green's function:

\[
\partial_t C + U\,\partial_x C = D\,\nabla^2 C,\qquad
C=\frac{\sigma_0^2}{\sigma^2}\exp\!\Big(\!-\frac{(x-x_0-Ut)^2+(y-y_0)^2}{2\sigma^2}\Big),\;
\sigma^2=\sigma_0^2+2Dt
\]

This is solved with BGK on D2Q5 (rest + 4 axis directions, w0 = 1/3, w1..4 = 1/6, c_s² = 1/3). The equilibrium distribution and the recovered diffusion coefficient:

\[
g_i^{\rm eq}=w_i\,C\Big(1+\frac{\boldsymbol{c}_i\cdot\boldsymbol{u}}{c_s^2}\Big),
\qquad D=c_s^2\Big(\tau-\tfrac12\Big)\;\Rightarrow\;\tau=3D+\tfrac12
\]

The velocity field u = (U, 0) is prescribed (no momentum LBM). In other words, this case isolates the D2Q5 collision and streaming as a pure advection–diffusion operator and checks it against the textbook Green's function. The exact solution is evaluated as a lattice sum over periodic images, so the reference side does not limit the L2 measurement. Fixing U per grid and raising N gives dx = 1/N and dt = U dx, so a 2nd-order-convergence grid-refinement test is well posed (see the header).

Code walkthrough

Source (363 lines).

%%{init: {'themeVariables': {'fontSize': '11px'}}}%%
flowchart LR
    subgraph DOM["Doubly periodic square lattice 128×128"]
      P(("Gaussian pulse σ0=10.24 cells<br/>translated by prescribed velocity u=(U,0)<br/>self-similar diffusion σ²=σ0²+2Dt"))
    end
    DOM -->|periodic wrap x,y| DOM
  1. Mapping — derive (D_lat, τ, steps) from (N, U, D_phys, T, σ0) and print them.
  2. collide/stream — D2Q5 BGK. The only boundary treatment is periodic array indexing.
  3. Diagnostics — every 25 steps, output the L2 (against the exact solution summed over periodic images), the second moment σ², and the total M/M0.
  4. Summary — slope of the straight-line fit of σ²(t) vs 2D, centroid velocity vs U, RESULT line.

Build & run

cmake -S native -B native/build && cmake --build native/build -j
./native/build/orheo-advection-diffusion2d

advection_diffusion2d [N] [U] [D_phys] [T] [sigma0] [out_prefix]

Cost of the computation that made the figures on this page (Miyabi-G Grace CPU (72 cores, -stdpar=multicore)): 1.1 s wall.

Things to try

  • Grid refinement: advection_diffusion2d 256 → 512. Since U is fixed, L2 should fall as dx². Measure the order of decay yourself.
  • The danger zone τ → 1/2: reduce D_phys to 1/10 and look at how the error behaves around τ = 0.512 (stable, but the error structure changes).
  • Vary Ma: U = 0.05 → 0.2. Check whether contamination of 2nd order in the lattice Mach number appears in the advection-velocity error (0.00% in this run).
  • Make σ0 smaller: narrowing the pulse to a few cells makes the numerical dispersion of D2Q5 visible. How much does the initial L2 value jump?

Validation

Tier T0 — order of convergence verified
QoI temperature-L2-error(abs)
Reference analytic Gaussian Green's function of the AD equation
Gate 🧩 module gate (advection_diffusion)
Measured 0.001813924 (rel. err +0.18%) — harvest N=128

Source & run details

View source

Item Value
Target orheo-advection-diffusion2d
Source native/benchmarks/advection_diffusion2d.cpp
Lattice D2Q5
GPU twin no
Gallery run 2026-07-27 @ 7628c0cf
Args ``
Wall time 1.11 s
CPU time 14.07 s
Device Miyabi-G Grace CPU (72 cores, -stdpar=multicore)

Computation log

The actual output of this computation (head and tail).

advection_diffusion2d N=128 U=0.0500 D_phys=0.0064 T=0.500 sigma0=0.080 | D_lat=0.040960 tau=0.62288 steps=1280  Pe=12.5
  step 25/1280: L2=1.7820e-03 sigma2=106.623 (exact 106.906) M/M0=1.000000
  step 50/1280: L2=1.6671e-03 sigma2=108.664 (exact 108.954) M/M0=1.000000
  step 75/1280: L2=1.5866e-03 sigma2=110.704 (exact 111.002) M/M0=1.000000
  step 100/1280: L2=1.5245e-03 sigma2=112.744 (exact 113.050) M/M0=1.000000
  step 125/1280: L2=1.4748e-03 sigma2=114.785 (exact 115.098) M/M0=1.000000
  step 150/1280: L2=1.4344e-03 sigma2=116.825 (exact 117.146) M/M0=1.000000
  step 175/1280: L2=1.4016e-03 sigma2=118.865 (exact 119.194) M/M0=1.000000
  step 200/1280: L2=1.3751e-03 sigma2=120.906 (exact 121.242) M/M0=1.000000
  step 225/1280: L2=1.3542e-03 sigma2=122.946 (exact 123.290) M/M0=1.000000
  step 250/1280: L2=1.3380e-03 sigma2=124.986 (exact 125.338) M/M0=1.000000
  step 275/1280: L2=1.3261e-03 sigma2=127.026 (exact 127.386) M/M0=1.000000
  step 300/1280: L2=1.3180e-03 sigma2=129.067 (exact 129.434) M/M0=1.000000
  step 325/1280: L2=1.3132e-03 sigma2=131.107 (exact 131.482) M/M0=1.000000
...

  step 1250/1280: L2=1.7975e-03 sigma2=206.594 (exact 207.258) M/M0=1.000000
  step 1275/1280: L2=1.8112e-03 sigma2=208.633 (exact 209.306) M/M0=1.000000
  step 1280/1280: L2=1.8139e-03 sigma2=209.041 (exact 209.715) M/M0=1.000000
--- N=128 summary ---
  L2(final)          = 1.8139e-03
  diffusion slope    = 8.162253e-02  vs 2D = 8.192000e-02  (-0.36%)
  sigma^2 intercept  = 104.5727  vs sigma0^2 = 104.8576
  advection speed    = 5.000004e-02  vs U = 5.000000e-02  (0.00%)
RESULT N=128 L2=1.813924e-03 slope=8.162253e-02 twoD=8.192000e-02 speed=5.000004e-02 U=5.000000e-02 sig2fit0=104.572694 sig0sq=104.857600

References

  • R. G. M. van der Sman & M. H. Ernst, J. Comput. Phys. 160 (2000) 766 — D2Q5 diffusion lattice.
  • X. He, S. Chen & G. D. Doolen, J. Comput. Phys. 146 (1998) 282 — advection–diffusion equilibrium distribution.