You are on page 1of 31

May 16 - 20, 2016

On the Relation
Between
Numerical and Parameter
Uncertainties

L. Eça (IST-Lisbon)
G.Vaz (MARIN)

May 16 - 20, 2016


May 16 - 20, 2016

Contents
1. Motivation
2. Parameter and Numerical Uncertainties
3. Example for the transitional flow over a flat
plate
4. Final Remarks

May 16 - 20, 2016


May 16 - 20, 2016

1. Motivation
• The use of CFD simulations for practical
applications requires the assessment of the
modelling error (Validation) to establish their
credibility
• The three contributions to the Validation
Uncertainty defined in the ASME V&V 20
procedure are:
- Experimental Uncertainty
- Numerical Uncertainty
- Parameter Uncertainty

May 16 - 20, 2016


May 16 - 20, 2016

1. Motivation
• The determination of the Parameter Uncertainty
is handled by Uncertainty Quantification (UQ)
that uses techniques including a large number of
simulations
• The large number of simulations required by UQ
makes the process time consuming and so it is
not unusual to use coarse grids and less strict
iterative convergence criteria to reduce the
computational effort

May 16 - 20, 2016


May 16 - 20, 2016

1. Motivation
• However, numerical errors (discretization and
iterative) may pollute the outcome of the UQ
exercises
• In this presentation we investigate the
dependence of the parameter uncertainty on the
discretization error of the simulations
• To this end we have chosen the simple local
sensitivity method

May 16 - 20, 2016


May 16 - 20, 2016

1. Motivation
• Furthermore, the sensitivity coefficients are
determined with finite-difference approximations
for the inlet turbulence quantities of the selected
test case
• Sensitivity coefficients are determined for a wide
range of grid densities to investigate the
influence of the discretization error on the
estimated parameter uncertainty
• All calculations are performed with
(www.refresco.org)
May 16 - 20, 2016
May 16 - 20, 2016

2. Parameter Uncertainty
• For n uncertain input variables X i, the parameter
uncertaintyU input for a quantity of interest S can be
determined from
2
n
 ∂S 
U 2
input = ∑  u X i 
i =1  ∂X i 
• Sensitivity coefficients are approximated by
∂S S ( X i + ∆X i ) − S ( X i − ∆X i )

∂X i 2 ∆X i

May 16 - 20, 2016


May 16 - 20, 2016

2. Numerical Uncertainty
• Contributions of round-off and iterative errors
are reduced to negligible levels when compared
to the discretization error
• Discretization error is estimated from grid
refinement studies performed in geometrically
similar grids with power series expansions
• Solution extrapolated to cell size zero to obtain
“exact solution”

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


• Reynolds-Averaged Navier-Stokes (RANS)
equations (time-averaging)
• Two-equation k-ω eddy-viscosity model with
“Low-Reynolds” damping functions for transition
prediction of Wilcox (1998)
• Quantity of interest is the skin friction coefficient
Cf at the locations measured in the experiments
reported in the ERCOFTAC Classic Database
(T3A)

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


• Reynolds number based on plate length L and
incoming velocity U ∞ equal to Re=107
C p =0
y ∂φ
1.5L =0
∂x
U x =1
U y =0 0 .25L
k = ki
ε = εi x
0 . 25L L 0 .25L

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


• 13 geometrically similar grids ranging from
3073×385 to 385×49 (1≤hi/h1≤8)
0.2
y/L
L

0.1

0 0 0.5 1
x/L
• Calculations converged to maximum normalized
residual of all transport equations below 10-8

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


• Turbulence kinetic energy k and dissipation rate
of turbulence kinetic energy ε are the uncertain
input quantities −3 2
ki = 1.35 × 10 U ∞
3
ε
• Reference conditions: i = 1 .64 L / U ∞

ν ti = ν
∂C f
• calculated with ∆ki = 0.9 × 10 −4 U ∞2 , ∆ki = 2 × 10 −4 U ∞2
∂ki
∂C f
• calculated with ∆ε i = 0.1L / U ∞3 , ∆ε i = 0.2 L / U ∞3
∂ε i

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


8 T3A, I=3.3% ki=1.35×10-33 U2∞
εi=1.64L/U∞
7 νti=ν
6
5 hi/h1=8
3
10
Cf×1

4
3
hi/h1=1
2
1
0 4 5 6 7
10 10 10 10
Rex

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


8 T3A, I=3.3% ki=1.26×103-3U2∞
εi=1.64L/U∞
7 νti=0.87ν
6
5 hi/h1=8
3
10
Cf×1

4
3
hi/h1=1
2
1
0 4 5 6 7
10 10 10 10
Rex

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


8 T3A, I=3.3% ki=1.44×103-3U2∞
εi=1.64L/U∞
7 νti=1.14ν
6
5 hi/h1=8
3
10
Cf×1

4
3
hi/h1=1
2
1
0 4 5 6 7
10 10 10 10
Rex

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


8 T3A, I=3.3% ki=1.15×103-3U2∞
εi=1.64L/U∞
7 νti=0.73ν
6
5 hi/h1=8
3
10
Cf×1

4
3
hi/h1=1
2
1
0 4 5 6 7
10 10 10 10
Rex

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


8 T3A, I=3.3% ki=1.55×103-3U2∞
εi=1.64L/U∞
7 νti=1.32ν
6
5 hi/h1=8
3
10
Cf×1

4
3
hi/h1=1
2
1
0 4 5 6 7
10 10 10 10
Rex

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


8 T3A, I=3.3% ki=1.35×103-3U2∞
εi=1.54L/U∞
7 νti=1.06ν
6
5 hi/h1=8
3
10
Cf×1

4
3
hi/h1=1
2
1
0 4 5 6 7
10 10 10 10
Rex

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


8 T3A, I=3.3% ki=1.35×103-3U2∞
εi=1.74L/U∞
7 νti=0.94ν
6
5 hi/h1=8
3
10
Cf×1

4
3
hi/h1=1
2
1
0 4 5 6 7
10 10 10 10
Rex

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


8 T3A, I=3.3% ki=1.35×103-3U2∞
εi=1.44L/U∞
7 νti=1.14ν
6
5 hi/h1=8
3
10
Cf×1

4
3
hi/h1=1
2
1
0 4 5 6 7
10 10 10 10
Rex

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


8 T3A, I=3.3% ki=1.35×103-3U2∞
εi=1.84L/U∞
7 νti=0.89ν
6
5 hi/h1=8
3
10
Cf×1

4
3
hi/h1=1
2
1
0 4 5 6 7
10 10 10 10
Rex

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate

4
sc
Edi

0
8
h /6
i h 4
i 400000
2 200000
0 Re x

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


3000
∆k=1.8×10-4
-4
∆k=4×10
2000
∂Cf/∂k

1000

0
8
h/ 6 400000
i h 4
i 200000
2
0
Re x

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


1.5

∆ε=0.2
1 ∆ε=0.4
∂Cf/∂ε

0.5

0
8
h/ 6 400000
i h 4
i 200000
2
0
Re x

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


101 Discretization error of Cf
0
10

-1
10
hi / h1
Ediscc

-2
10

-3
10

-4
10

-5
10
100000 150000 200000 250000 300000
Rex

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


104
Solid ∆k=1.8×10-4-4
Dashed ∆k=4×10
3
10

2
10
∂Cf/∂∂k

1
10

0
10

-1
10

-2
10
100000 150000 200000 250000 300000
Rex

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


101
Solid ∆ε=0.2
Dashed ∆ε=0.4
0
10

-1
10
∂Cf/∂∂ε

-2
10

-3
10

-4
10

-5
10
100000 150000 200000 250000 300000
Rex

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


102 Discretization error of Cf
4
Rex=3.2×10
4
10
1 Rex=6.7×10
5
Rex=2×10
5
0 Rex=2.7×10
10 Rex=4.2×10
5

5
Rex=4.9×10
Ediscc

-1
10

-2
10

-3
10

-4
10
1 2 3 4 5 6 7 8
hi/h1

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


104 -4
∆k=1.8×10 4
∆k=4×10-4
Rex=3.2×10
4
Rex=6.7×10
3 5
10 Rex=2×10
5
Rex=2.7×10
5
2 Rex=4.2×10
10
∂Cf/∂∂k

5
Rex=4.9×10

1
10

0
10

-1
10
1 2 3 4 5 6 7 8
hi/h1

May 16 - 20, 2016


May 16 - 20, 2016

3. Transitional Flow over a Flat Plate


101 ∆ε=0.2 ∆ε=0.4
4
Rex=3.2×10
4
10
0 Rex=6.7×10
5
Rex=2×10
5
-1
Rex=2.7×10
10 Rex=4.2×10
5
∂Cf/∂∂ε

5
Rex=4.9×10
-2
10

-3
10

-4
10

-5
10
1 2 3 4 5 6 7 8
hi/h1

May 16 - 20, 2016


May 16 - 20, 2016

4. Final Remarks
• Sensitivity coefficients determined from finite-
difference approximations are dependent on the
discretization error of the simulations

• Checks performed with too coarse grids may


lead to the misleading conclusion that sensitivity
coefficients are independent of the numerical
error of the simulations

May 16 - 20, 2016

You might also like