Skip to content

Subgrid-scale models ​

IDMilestoneTypeLegsBackendsFigure
SG#119 · T1P+Csg1–sg4CPU, FNLyes

Case and scope ​

T1 adds the algebraic eddy-viscosity models Vreman, WALE and Sigma ([turbulence] sgs_model, turbulence) and asks the go/no-go question of #119: does a model improve on the ILES baseline? The legs check the formulas, the energy budget with the model and rank invariance; the campaign reruns the Taylor-Green vortex of T0.3 with each model.

Legs ​

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

LegWhat it verifiesMeasured
sg1the formulas against numpy on 200 tensors (test_sgs); the zeros of each model3.6⋅10−15, 1.2⋅10−15, 1.8⋅10−11
sg2budget with the model, M 0.4: Vreman 36 %, 5.7 %, 0.27 % left at 323, 643, 1283at 1283: 21.6 % without εsgs, 43 % flipped
sg3no viscosity with each modelε=0, mass and energy 3.3⋅10−16
sg4one rank against 2 with a model10−14

Taylor-Green vortex with a model ​

Setup
flow, scheme, runas the ILES baseline: Re 1600, M 0.1, WENO-5, runge-kutta-ssp-54, CFL 0.5, 4 ranks on A30
modelsVreman, WALE, Sigma, default constants (not tuned)
grids643, 1283; Vreman also at M0=1.25, 1283

The models and the 2:1 box

Top: M 0.1 at 643; middle: at 1283; bottom: a 2:1 box on the centre of the 643 grid against the uniform grids.

Case, M 0.1Peak −dK/dtat tPeak ZShare to the peakK(20)
no model, 6430.013775.081.7177 %0.0245
Vreman, 6430.013855.051.5979 %0.0236
WALE, 6430.013745.041.6079 %0.0239
Sigma, 6430.013805.171.5979 %0.0237
no model, 12830.012428.243.4058 %0.0219
Vreman, 12830.012108.293.1861 %0.0213
WALE, 12830.011998.293.1661 %0.0210
Sigma, 12830.012068.303.1860 %0.0213

The share leaves εsgs in: it is the energy lost not accounted for by ε−pd, so it includes the model's own dissipation. At M0 1.25, 1283, Vreman gives peak −dK/dt 0.01085 and peak Z 3.10, against 0.01081 and 3.34 without a model.

Against the 512³ DNS (L2 errors over 0.3<t<19.7; the full table is in Against the DNS):

CaseL2 −dK/dtL2 ZL2 K
WENO-5, 6430.6600.8440.187
Vreman, 6430.6710.8540.200
WENO-5, 12830.2780.6430.083
Vreman, 12830.2950.6620.092
WALE, 12830.2960.6600.091
Sigma, 12830.2930.6600.091

What it shows.

  • No model beats no model. On a dissipative scheme each one lowers the enstrophy peak further from the reference (1.59 against 1.71 at 643, 3.18 against 3.40 at 1283), and all three worsen every L2 error.
  • The three models are nearly the same.
  • The eddy viscosity adds to a numerical dissipation that already dominates. In the laminar phase, to t=2 at M 0.4, it takes 21–53 % of the energy lost.
  • Go/no-go: no-go for a model on WENO. Under the rule of #119 this makes the hybrid central/WENO scheme (T2) necessary before any model is recommended.

The constants come from other discretizations. The same models in decaying isotropic turbulence are in the ILES baseline; on the hybrid scheme in T2.

Not done ​

  • A spectral comparison with filtered DNS, which could rank the models differently.
  • Tuned constants.
  • A model on 2:1 grids beyond conservation (hy9); a model with runtime AMR.