Quantification of Epistemic Uncertainties

in the k − ε Model Coefficients

By:
Saeed Salehi
saeed.salehi@chalmers.se

DF

Gothenburg Region OpenFOAM User Group Meeting,

November 20th , 2019
Chalmers University of Technology
Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Introduction

DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 0/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Uncertainty

“Uncertainty is the only certainty there is, and knowing how to live
with insecurity is the only security.”
— John Allen Paulos
Uncertainties are present in most engineering and practical applications.

Sources of Uncertainty:
X physical properties
X initial conditions
X boundary conditions
X geometry
X model parameters
X etc.

DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 1/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Uncertainty

“Uncertainty is the only certainty there is, and knowing how to live
with insecurity is the only security.”
— John Allen Paulos
Uncertainties are present in most engineering and practical applications.

Sources of Uncertainty:
X physical properties
X initial conditions
X boundary conditions
X geometry
X model parameters
X etc.

Despite these uncertainties can predictions be trusted? DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 1/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Uncertainty in Turbomachines

Turbomachines can be intensely sensitive to the

uncertainties.
Aleatory uncertainties:
X Operating conditions
X Geometry

Operational: Volumetric flow rate, turbulence properties,

rotational speed.
Geometrical:
X Manufacturing tolerances.
X Turbomachines are designed to run for years: in-service
erosion and corrosion.

DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 2/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Uncertainty Quantification Using Polynomial Chaos

Expansion

DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 2/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Uncertainty Quantification
How do input uncertainties affect objective functions are affected?
Non-deterministic CFD is a growing field. (e.g. NUMECA)
First step: identification of the sources of uncertainties and their PDFs.

DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 3/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Uncertainty Quantification
How do input uncertainties affect objective functions are affected?
Non-deterministic CFD is a growing field. (e.g. NUMECA)
First step: identification of the sources of uncertainties and their PDFs.
Uncertainty Quantification
Quantitative characterization and reduction of uncertainties in applications.

DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 3/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Uncertainty Quantification
How do input uncertainties affect objective functions are affected?
Non-deterministic CFD is a growing field. (e.g. NUMECA)
First step: identification of the sources of uncertainties and their PDFs.
Uncertainty Quantification
Quantitative characterization and reduction of uncertainties in applications.
Determinstic system
ξ1

.. Computational Model y
. y = U(ξ1 , ξ2 , · · · , ξd )

ξd

DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 3/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Uncertainty Quantification
How do input uncertainties affect objective functions are affected?
Non-deterministic CFD is a growing field. (e.g. NUMECA)
First step: identification of the sources of uncertainties and their PDFs.
Uncertainty Quantification
Quantitative characterization and reduction of uncertainties in applications.
Determinstic system
ξ1

.. Computational Model y
. y = U(ξ1 , ξ2 , · · · , ξd )

ξd

Stochastic system
f ξ1

ξ1
fy
.. Computational Model
. y = U(ξ1 , ξ2 , · · · , ξd )
y
f ξd
DF

ξd

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 3/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Uncertainty Quantification
How do input uncertainties affect objective functions are affected?
Non-deterministic CFD is a growing field. (e.g. NUMECA)
First step: identification of the sources of uncertainties and their PDFs.
Uncertainty Quantification
Quantitative characterization and reduction of uncertainties in applications.
Determinstic system
ξ1 Statistics:
Z +∞
.. Computational Model y
. y = U(ξ1 , ξ2 , · · · , ξd ) hyi = zfy (z) dz
−∞
ξd Z +∞
2 2
σ = (z − E[y]) fy (z) dz
−∞
Z +∞
3
µ3 = (z − E[y]) fy (z) dz
Stochastic system −∞
f ξ1 Z +∞
4
µ4 = (z − E[y]) fy (z) dz
−∞
ξ1
fy
.. Computational Model
. y = U(ξ1 , ξ2 , · · · , ξd )
y
f ξd
DF

ξd

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 3/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Uncertainty Quantification
How do input uncertainties affect objective functions are affected?
Non-deterministic CFD is a growing field. (e.g. NUMECA)
First step: identification of the sources of uncertainties and their PDFs.
Uncertainty Quantification
Quantitative characterization and reduction of uncertainties in applications.
Determinstic system
ξ1 Statistics:
Z +∞
.. Computational Model y
. y = U(ξ1 , ξ2 , · · · , ξd ) hyi = zfy (z) dz
−∞
ξd Z +∞
2 2
σ = (z − E[y]) fy (z) dz
−∞
Z +∞
3
µ3 = (z − E[y]) fy (z) dz
Stochastic system −∞
f ξ1 Z +∞
4
µ4 = (z − E[y]) fy (z) dz
−∞
ξ1
fy
.. Computational Model
. y = U(ξ1 , ξ2 , · · · , ξd )
f ξd
y UQ Methods
Introduce mathematical approaches DF

ξd to solve above integrals.

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 3/ 19
Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Polynomial Chaos Expansion

Uncertainty Quantification approaches

X Sampling methods (non-intrusive)
X Quadrature methods (non-intrusive)
X Spectral methods (intrusive): Polynomial Chaos Expansion
The stochastic field U (x; ξ) is decomposed

PC Expansion
X
U (x; ξ) = uα (x)ψ α (ξ)
06|α|6p

 
p + ns
The number of unknown coefficients (ui ’s) is: P + 1 =
p
Curse of dimensionality
Functions ψi (ξ)’s are the orthogonal polynomials with respect to input PDFs.
Regression approach is employed to calculate the PCE coefficients.
Sobol’ sampling scheme
DF

Variance based sensitivity analysis using Sobol’ indices.

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 4/ 19
Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Sparse Reconstruction of Polynomial Chaos Expansion

Using Compressed Sensing

DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 4/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Introduction

Industrial problems: large number of random

variables. U (x; ξ) =
X
uα (x)ψ α (ξ)
06|α|6p
Curse of dimensionality: Computational cost
of PCE grows exponentially with the number
variables. 100

10−2
Remedy: efficient methods

PCE coefficients
10−4

Efficient methods 10−6

X Adaptive methods 10−8

X Reduced basis methods 10−10 0

10 101 102 103
Index of PC basis

X Multifidelity methods
Figure: PCE coefficient of drag coefficient of
X Sparse methods the RAE2822 airfoil with stochastic
geometry and operating condition

Sparse reconstruction approaches:

X Hyperbolic sparse
X Compressed sensing
DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 5/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Compressed Sensing

Definition
Recover a sparse signal from a set of incomplete observations

Optimization problem (`0 -minimization):

û = argminu kuk0 subject to Ψu = Y

`0 -minimization is NP-hard! Hence, `1 -minimization

û = argminu kuk1 subject to Ψu = Y Original

Noisy signal:
û = argminu kuk1 subject to kΨu − Yk2 6 

p=1 p = 1.5 p=2 p=3 p=∞

û û û û û
A A A A A

u u u u u

DF

Reduced to 10% of the wavelet

coefficients

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 6/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Limitation of Classic PCE

10
PCE basis should be orthogonal with respect to the PDF
Original
PDF of the uncertain input parameters. 8

6
The Wiener-Askey polynomial: exponential convergence
for a limited number of PDFs. 4

What about arbitrary PDFs? 2

0
Gram-Schmidt orthogonalization method. 1.7 1.75 1.8 1.85 1.9
Cε2
1.95 2 2.05 2.1

PDF of Cε2 in k − ε model.

Exponential convergence for arbitrary distributions. 0.05

0.04
The GSPCE is revisited and for the first time used with
the regression method. 0.03

Distribution Polynomials Support 0.02

Gaussian Hermite (−∞, ∞)
Uniform Legendre [−1, 1] 0.01
Gamma Laguerre [0, ∞)
Beta Jacobi [−1, 1] 0
−20 −10 0 10 20 30 40
Wind speed (m/s)
PDF of wind speed from measured
X
U (x; ξ) = uα (x)ψ α (ξ) meteorological data (www.umass.edu).
DF

06|α|6p

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 7/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Gram-Schmidt Orthogonalization Method

Legendre (Uniform)
One-dimensional monic orthogonal polynomials: 1 1
Legendre Polynomials Legendre Polynomials
Gram-Schmidt Gram-Schmidt
j−1
X
0.5 0.5

P3 (x)

P4 (x)
ψj (ξ) = ej (ξ) − cjk ψj (ξ), j = 1, 2, · · · , p, 0 0

k=0 −0.5 −0.5

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
with x x

hej (ξ), ψk (ξ)i 1

Legendre Polynomials
1
Legendre Polynomials

ψ0 = 1 cjk = , 0.5
Gram-Schmidt
0.5
Gram-Schmidt

hψk (ξ), ψk (ξ)i

P5 (x)

P6 (x)
0 0

where the polynomials ej (ξ) are polynomials of −0.5 −0.5

exact degree j. −1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
x x

1 1
Legendre Polynomials Legendre Polynomials
Gram-Schmidt Gram-Schmidt

The polynomials are normalized as: 0.5 0.5

P7 (x)

P8 (x)
0 0

ψj (ξ) −0.5 −0.5

ψj (ξ) =
hψj (ξ), ψj (ξ)i −1
−1 −0.5 0 0.5 1
−1
−1 −0.5 0 0.5 1
x x

DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 8/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Gram-Schmidt Orthogonalization Method

Hermite (Gaussian)
One-dimensional monic orthogonal polynomials: 50
20 Hermite Polynomials Hermite Polynomials
Gram-Schmidt Gram-Schmidt
j−1
X 10

He3 (x)

He4 (x)
ψj (ξ) = ej (ξ) − cjk ψj (ξ), j = 1, 2, · · · , p, 0 0

−10
k=0
−20
−50
−5 0 5 −5 0 5
with x x

hej (ξ), ψk (ξ)i 100

Hermite Polynomials
200
Hermite Polynomials

ψ0 = 1 cjk = , 50
Gram-Schmidt
100
Gram-Schmidt

hψk (ξ), ψk (ξ)i

He5 (x)

He6 (x)
0 0

where the polynomials ej (ξ) are polynomials of −50 −100

exact degree j. −100 −200

−5 0 5 −5 0 5
x x

1000 5000
Hermite Polynomials Hermite Polynomials
Gram-Schmidt Gram-Schmidt

The polynomials are normalized as: 500

He7 (x)

He8 (x)
0 0

ψj (ξ) −500
ψj (ξ) =
hψj (ξ), ψj (ξ)i −1000
−5 0 5
−5000
−5 0 5
x x

DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 8/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Numerical Example: Quantification of epistemic

uncertainties in the k − ε model coefficients in
OpenFOAM channel flow

DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 8/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Epistemic Uncertainties in the k − ε Model Coefficients

Sources of uncertainties in RANS models:

Model formulation
Coefficients

The Launder-Shrama k − ε model:

X Boussinesq (eddy-viscosity) hypothesis:

∂Ui ∂Uj 2
 
ui uj = −νt + + kδij
∂xj ∂xi 3
X Turbulent viscosity:
k2
νt = Cµ fµ
ε
X To obtain νt , transport equations are solved:
√ !2
∂ ∂ νt ∂k ∂ k
  
(Uj k) = ν+ + Pk − ε̃ − 2ν
∂xj ∂xj σk ∂xj ∂xj

∂ ∂ νt ∂ ε̃ ε̃ ε̃2
  
(Uj ε̃) = ν+ + Cε1 Pk − Cε2 f2 +E (1)
∂xj ∂xj σε ∂xj σε k
DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 9/ 19

Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Launder-Sharma k − ε model coefficients

The empirical coefficients of low-Re Launder-Sharma k − ε model

Cµ σk σ Cε1 Cε2
0.09 1.0 1.3 1.44 1.92

The coefficients are related to some basic physical quantities (Durbin and Reif, 2011):
X The decay exponent in decaying homogeneous, isotropic turbulence, n
X The production to dissipation in homogeneous shear flow, P/ε
X The Von-Karman constant, κ
X The dimensionless turbulent kinetic energy in the logarithmic layer, klog /u2τ

 −2
klog Cε2 − 1
Cµ = , Cε1 = + 1,
u2τ P/ε
1−n κ2
Cε2 = , σε = p .
n Cµ (Cε2 − Cε1 )

The reported data for these quantities in the literature are collected and presented as
PDFs (Margheri et al., 2014) DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 10/ 19
Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

PDFs of Basic Physical Quantities

1.4 1.4

1.2 1.2

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
−1.3 −1.2 −1.1 −1 1.4 1.5 1.6 1.7 1.8 1.9
n P/ǫ
1.4 1.4

1.2
1.2

1
1

0.8
0.8

0.6
0.6

0.4
0.4

0.2
0.2
0
0 2.8 3 3.2 3.4 3.6 3.8 4
0.38 0.39 0.4 0.41 klog/u2τ
κ

−2
κ2

klog Cε2 − 1 1−n
Cµ = , Cε1 = + 1, Cε2 = , σε = p .
u2τ P/ε n Cµ (Cε2 − Cε1 ) DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 11/ 19
Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

PDFs of k − ε Coefficients
8
20
PDF
PDF
7 Original
Original

15 6

10 4

5 2

0 0
0.04 0.06 0.08 0.1 0.12 0.14 0.16 1.3 1.4 1.5 1.6 1.7 1.8
Cµ Cε1
10 1.4
PDF PDF
Original Original
1.2
8
1
6
0.8

4 0.6

0.4
2
0.2

0 0
1.7 1.75 1.8 1.85 1.9 1.95 2 2.05 2.1 0.5 1 1.5 2 2.5 3 3.5 4
Cε2 σǫ

The PDFs of Launder-Sharma Coefficients are computed using a Monte-Carlo simulation

Reported standard values lie within its PDF range
DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 12/ 19
Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Test Case: Channel Flow

Fully-developed turbulent channel flow Turbulence model: LaunderSharmaKE

Friction Reynolds number Discretization:
uτ H X gradSchemes: linear
Reτ = = 950
ν X divSchemes: linear
OpenFOAM Solver: boundaryFoam X divSchemes: linear
Steady-state solver for incompressible, 1D Linear solvers:
turbulent flow
X U: PCG, DIC
grad p̄ = divτ̄
X k: smoothSolver, symGaussSeidel
Number of stochastic parameters ns = 4 X epsilon: smoothSolver, symGaussSeidel
UQ analyses preformed with p = 7 25 3
Launder Sharma k − ε Launder Sharma k − ε
DNS 2.5 DNS
Reference solution: 8th order full PC 20

2
15

U+

u′+
1.5
10
1

Wall
Periodic

Periodic

5
0.5

0 −1 0 1 2 3
0 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10
y+ y+
1.5 0.9
Flow Launder Sharma k − ε
0.8
Launder Sharma k − ε
DNS DNS
0.7

1 0.6

H 0.5

u′ v ′+
v′+

0.4

0.5 0.3
Wall 0.2

0.1 DF

0 −1 0 1 2 3
0 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10
y+ y+

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 13/ 19
Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Convergence of Friction Velocity (uτ ) Statistics

Mean Variance PDF

2 6 100
10 10
Full PC, N=990
1 OMP, N=10
10 4 80
10 OMP, N=20
OMP, N=50
0
Rel. Error (%)

Rel. Error (%)

10 OMP, N=100
2 60
10
−1
10
0 40
10
−2
10

Full PC −2 Full PC 20
−3 10
10 Sparse PC (OMP) Sparse PC (OMP)
Monte Carlo Monte Carlo
−4 −4 0
10 1 2 3
10 1 2 3 0.02 0.03 0.04 0.05 0.06 0.07
10 10 10 10 10 10
Number of samples Number of samples uτ /Ub

10
0 Sparse method converged much
Full PC faster.
−2 Sparse PC, OMP
10
PCE coefficients

−4
10
OMP with N = 100 sample method
−6
has successfully explored the
10
important basis in the PCE.
−8
10

−10
10
0 1 2 3
10 10 10 10
Index of PCE basis DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 14/ 19
Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

2D PDFs of Velocity Field

Normalized 2D PDFs
(PDF/max(PDF)).

Increasing number of
samples improves
reconstructed 2D PDFs.

Using N = 100 samples the

(a) OMP, N = 10 (b) OMP, N = 50 constructed 2D PDF is very
similar to the full PC.

The sparse method is 9

times faster!

(c) OMP, N = 100 (d) Full PC, N = 990 DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 15/ 19
Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

2D PDFs of Turbulence Field

Normalized 2D PDFs
(PDF/max(PDF)).

Increasing number of
samples improves
reconstructed 2D PDFs.

Using N = 100 samples the

(a) OMP, N = 10 (b) OMP, N = 50 constructed 2D PDF is very
similar to the full PC.

The sparse method is 9

times faster!

(c) OMP, N = 100 (d) Full PC, N = 990 DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 16/ 19
Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

2D PDFs of Turbulence Field

Normalized 2D PDFs
(PDF/max(PDF)).

Increasing number of
samples improves
reconstructed 2D PDFs.

Using N = 100 samples the

(a) OMP, N = 10 (b) OMP, N = 50 constructed 2D PDF is very
similar to the full PC.

The sparse method is 9

times faster!

(c) OMP, N = 100 (d) Full PC, N = 990 DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 17/ 19
Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

2D PDFs of Turbulence Field

Normalized 2D PDFs
(PDF/max(PDF)).

Increasing number of
samples improves
reconstructed 2D PDFs.

Using N = 100 samples the

(a) OMP, N = 10 (b) OMP, N = 50 constructed 2D PDF is very
similar to the full PC.

The sparse method is 9

times faster!

(c) OMP, N = 100 (d) Full PC, N = 990 DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 18/ 19
Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

Comparing with DNS Data

DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 19/ 19
Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 19/ 19
Introduction Uncertainty Quantification Efficient UQ Method Numerical Example

DF

Saeed Salehi Quantification of epistemic uncertainties in the k − ε model coefficients OFGBG19 19/ 19