Skip to content

The hybrid central/WENO scheme ​

IDMilestoneTypeLegsBackendsFigure
HY#119 · T2.1–T2.7P+Chy1–hy13CPU, FNLyes

Case and scope ​

scheme_space = hybrid-kep-weno (turbulence) uses a 6th-order kinetic-energy- and pressure-equilibrium-preserving central flux at the faces away from shocks and keeps WENO where a Ducros-type sensor flags a cell (T2.1–T2.3). WENO is kept on the skin of 2:1 seams and physical faces by default (hybrid_skin, T2.4). The legs check the central flux, pressure equilibrium (T2.5), the model masked to central cells (T2.6), conservation on every grid path and robustness at shocks (T2.7). The campaign reruns the ILES baseline cases and compares every scheme with the 512³ DNS.

Legs ​

src/tests/flume/verification/turbulence/check.sh, CPU and FNL.

LegPhaseWhat it verifiesMeasured
hy1T2.7inviscid Taylor-Green: dissipated K, hybrid against WENO47 times less on FNL
hy2T2.3one rank against 2 with the hybrid scheme on the random field, where the sensor firesnot recorded
hy3T2.2the central flux (test_kep): order, kinetic energy identity, pressure equilibriumorder 5.96–5.99; 4⋅10−16; dp/dt 9⋅10−15 across 100:1
hy4T2.1contact of density ratio 8 and 100u, p within 9⋅10−14
hy5T2.42:1 box, conservation; passes with any skinmass 10−16
hy6T2.4Sod against WENOL1 8.3⋅10−5
hy7T2.4runtime regridding; a forest with open ends10−16; 5.7⋅10−10, the single realm's drift
hy8T2.5density wave, amplitude 0.2; WENO holds it too, so this does not separate themp, u 10−14
hy9T2.6model only off flagged cells; one rank against 2; conservation (does not prove the mask on ghost cells)εsgs exactly 0 with all flagged; 7⋅10−14; 2.3⋅10−16 with a model on a 2:1 box; whether a flagged cell faces a free one at an interface is not checked
hy10T2.4the skin relaxed on 2:1 seams (the hy5 case, central-seams): what conservation does, the last history rowsa measure with no bound: printed, never fails the script
hy11T2.4relaxing the skin changes the density: seams, open ends (does not catch a shifted band)1.6⋅10−4, 6.4⋅10−5
hy12T2.4restart on a refined grid, every cell0
hy13T2.7against WENO, L1: a vortex through a shock (2-D), Shu-Osher and a shock through 2:1 seams (1-D, no vorticity, so the Ducros ratio flags any varying flow): robustness, no accuracy claim1.0⋅10−4, 1.7⋅10−5, 3.4⋅10−5

Taylor-Green vortex and isotropic turbulence ​

Setup
casesthe ILES baseline inputs with scheme_space = hybrid-kep-weno
sensorthresholds 0.2 and 0.1 and the floor at their defaults (not calibrated)
modelnone, or Vreman acting off flagged cells only (T2.6)
runturbulence/tgv_campaign.sh, 4 ranks on A30

The hybrid scheme on the Taylor-Green vortex

Top: M 0.1, the hybrid against WENO-5; middle: with Vreman; bottom: M0=1.25. Left −dK/dt, right the enstrophy; dashed: the 512³ spectral DNS, on the M 0.1 rows.

Case, M 0.1Peak −dK/dt, rawOver Δt=0.5at tFlicker, 7<t<10Peak ZNumerical share to the peak
WENO-5, 6430.013770.013575.090.3 %1.7177 %
hybrid, 6430.015910.013348.097.2 %3.8250 %
hybrid + Vreman, 6430.013470.011738.213.5 %2.92
WENO-5, 12830.012420.012168.290.5 %3.4058 %
hybrid, 12830.013890.013178.482.0 %6.8618 %
hybrid + Vreman, 12830.012580.012268.580.8 %5.53
WENO-5, 25630.012720.012618.720.7 %6.1827 %

The flicker is the rms of −dK/dt about its running mean, over the smoothed peak. The share with Vreman is left out: it does not separate the model from the scheme. Against the same runs with the model everywhere, Z moves under 1 % at M 0.1, where the sensor rarely fires.

Case, M0=1.25Peak −dK/dtPeak ZNumerical share
WENO-5, 643 / 1283 / 25630.01383 / 0.01081 / 0.011011.81 / 3.34 / 5.9580 % / 46 % / 20 %
hybrid, 643 / 1283 / 25630.01266 / 0.01057 / 0.011122.20 / 4.83 / 7.2173 % / 36 % / 10 %
Isotropic turbulencePeak Z, 643 / 1283K(t/τ=4), 643 / 1283
WENO-534.5 / 51.0 (60.5 at 2563)0.303 / 0.384 (0.413 at 2563)
hybrid41.0 / 57.40.367 / 0.408
hybrid + Vreman37.8 / 54.1

What it shows.

  • At M 0.1 the hybrid at 1283 reaches a higher enstrophy peak than WENO-5 at 2563, and at 643 more than WENO-5 at 1283, with the dissipation peak where the finer runs have it. Z is still below the DNS's 10.3.
  • At M0=1.25 the gain is smaller: the sensor keeps WENO around the shocklets.
  • The raw peak of −dK/dt is not a measure on the hybrid. It flickers, 7 % at 643, 2 % at 1283. The source is not established; the sensor switching on and off is the first candidate. Smoothed, the hybrid is 4–6 % above WENO-5 at 2563.
  • Vreman on the hybrid removes half of the flicker and lowers Z by 20 %. At 643 it lands below the finer runs (0.0117). Against the DNS it makes the hybrid worse at both resolutions (below).

On a 2:1 box, and what the skin costs ​

Setup
gridM 0.1, base 643, a box of half the side refined once
skinhybrid_skin = weno (default) or central-seams

The hybrid scheme on a 2:1 box

M 0.1, a box of half the side refined once on 643. Left −dK/dt, right the enstrophy.

CaseSmoothed peak −dK/dtPeak Z
WENO-5, box0.013041.87
hybrid, box, WENO skin (default)0.012393.69
hybrid, box, central skin0.013194.13
hybrid, uniform 6430.013333.82

What it shows.

  • With the WENO skin the box ends below the uniform run in Z: the skin takes more than the extra cells give.
  • The central skin is stable to t=20 and above the uniform run. Against the DNS it is the closer one (table below), on this smooth flow; the case with a shock has no reference.
  • A shock through the seams runs with the central skin. A vortex after Jiang & Shu (the code's profile, same strength and radius, unit square) through a Mach 1.1 shock crossing the box seams differs from the WENO skin by at most 2.8⋅10−4 in density at t=0.35.

Against the DNS ​

M 0.1 against the 512³ DNS, from turbulence/tgv_dns.py. Peak −dK/dt over a window of 0.5; L2 errors over 0.3<t<19.7 relative to the DNS, −dK/dt over the same window. This table collects every T0, T1 and T2 run at M 0.1.

CasePeak −dK/dtat tPeak Zat tL2 −dK/dtL2 ZL2 K
DNS0.012868.9710.298.97
WENO-5, 6430.013575.091.7096.250.6600.8440.187
WENO-5, 12830.012168.293.4008.110.2780.6430.083
WENO-5, 25630.012608.726.1828.560.0900.3310.020
HLLC, 12830.012358.323.5858.110.2600.6230.070
WENO-7, 12830.012158.113.1718.140.2910.6880.101
WENO-9, 12830.012748.274.4548.030.1770.5390.037
Vreman, 6430.013735.111.5916.230.6710.8540.200
Vreman, 12830.011938.283.1778.150.2950.6620.092
WALE, 12830.011838.263.1648.160.2960.6600.091
Sigma, 12830.011928.293.1848.140.2930.6600.091
WENO-5, 2:1 box0.013046.411.8656.390.6170.8290.178
hybrid, 6430.013338.093.8227.860.2370.6150.057
hybrid, 12830.013178.476.8568.250.0670.2710.014
hybrid + Vreman, 6430.011738.212.9178.090.2830.6780.086
hybrid + Vreman, 12830.012268.585.5298.330.1140.4000.029
hybrid, 2:1 box, WENO skin0.012398.143.6887.880.2690.6180.063
hybrid, 2:1 box, central skin0.013198.114.1347.910.2360.5800.059

What it shows.

  • The hybrid at 1283 is closer than WENO-5 at 2563 on all three errors, with one eighth of the cells. At 643 it is closer than WENO-5 at 1283.
  • Its peak of −dK/dt is 2.4 % above the DNS and comes at t=8.47 rather than 8.97; Z at 1283 reaches two thirds of the DNS peak, so the gradients are not converged.
  • A model makes it worse. Vreman takes the −dK/dt error from 0.067 to 0.114 and the Z error from 0.27 to 0.40 at 1283. All three models also worsen WENO-5. At Re 1600 the subgrid scales are not far from resolved, and the constants were not tuned; on these integrals no model is recommended.
  • WENO-9 at 1283 (0.177) is much closer than WENO-5 (0.278); WENO-7 is further (0.291), see the ILES baseline.
  • The skin. On the 2:1 box the WENO skin (0.269) is further than uniform 643 (0.237); the central skin ties it on −dK/dt (0.236) and beats it on Z (0.580 against 0.615). Smooth flow only.

The flow ​

Vorticity of the Taylor-Green vortex and of the isotropic turbulence

hybrid-kep-weno at 1283 (tgv_cases_vis.txt). Taylor-Green, Re 1600, Mach 0.1: ωx on the periodic face x≈0 at t=4.2 and 9.4 (the peak of the dissipation is at t≈9; the Workshop's figures show the vorticity norm on this face) and ωz on the symmetry plane z=π at t=9.4. Isotropic turbulence, Mt 0.6, Reλ 100, k0=4: ωz on z=π at t=2. Planes from turbulence/field_slices.py --time; make_doc_figures.py --only tgv-fields.

  • At t=4.2 the vorticity is still in sheets; by t=9.4 they have rolled up and broken into the small structures of the transition.
  • The flow's mirror symmetries survive the breakdown. On the face x≈0, ωx is antisymmetric about y=π and about z=π to 8⋅10−13 of its maximum at t=9.4 (0 at t=4.2), as the initial condition is.
  • On z=π the flow is still large-scale at t=9.4: four counter-rotating cells with thin layers at their edges.
  • The isotropic field shows no visible imprint of the blocks (22 per direction). One plane, one time: a description, not a measurement.

Cost ​

200 steps on 1283, two A30, three samples:

CaseWall time
WENO24.1–24.6 s
hybrid, every cell flagged24.2–24.7 s
hybrid, every cell central26.7–26.9 s

The sensor costs under 1 %, the central flux about 10 % on top of WENO, startup included. The scheme runs both flux kernels; a single kernel with a mask was not tried, so what skipping WENO would save is not measured.

Not done ​

  • Calibration of the sensor thresholds and floor.
  • The source of the flicker of −dK/dt.
  • A model masked by face rather than by cell; the mask on ghost cells at an interface.
  • central-bcs on a forest seam (the sensor ghost is computed locally there).
  • A shock case with a reference for the relaxed skin.
  • A single flux kernel with a mask.