reacting flows and vortices
TRANSCRIPT
Outline Statement Approaches Test-cases Results Conclusions
Reacting flows and vortices
Sanaz Arabzade, (University of Toronto)Anne Bourlioux(Universite de Montreal)
Alexandre Desfosses Foucault (Universite de Montreal)Pierre Gauthier (Rolls-Royce Canada)
Apala Majumdar (Oxford)Catherine Mavriplis (University of Ottawa)
Mario Morfin Ramirez (University of Toronto)Stephen Peppin(Oxford)
Mary Pugh (University of Toronto)Louis-Xavier Proulx (Universite de Montreal)
Tim Reis (Oxford)Nasim Shahbazian (University of Toronto)
August 21, 2009
Outline Statement Approaches Test-cases Results Conclusions
1 Statement
2 Approaches
3 Test-cases
4 Results
5 Conclusions
Outline Statement Approaches Test-cases Results Conclusions
Problem Statement
Rolls Royce interest: chemistry in gas turbine combustor
Challenge: reliable prediction of CO, NO and other pollutants
Huge computational challenge: direct simulation with detailedchemistry of reacting turbulent flow.
Currently feasible: large eddy simulations with simplifiedchemistry
Still needed: chemicals above, only available with detailedchemistry
Strategy: some form of hybrid CFD/detailed chemistrycomputation
Outline Statement Approaches Test-cases Results Conclusions
Hybrid approach considered by Rolls Royce
1 Use CFD package for large eddy simulation of the reactiveNavier-Stokes model with simplified chemistry,
2 Time-average CFD output to create spatially-dependent buttime–independent values for the velocity (u, v ,w),temperature T , pressure p, carbon monoxide CO, nitric oxideNO, carbon dioxide CO2, water H20, and oxygen O2,
3 Post–proces time-averaged data to identify so called“vortices” (regions of space in which a fluid particle wouldspend a good chunk of time before leaving the region),
4 Compute in each region a “residence time”, the spatialaverages of CO, NO, CO2, H20, O2, the temperature, and thepressure, and the fluxes of these quantities between regions.
5 Pass these quantities, along with the geometry of thedecomposition (which region abuts which) to a chemistrysoftware Chemkin (detailed chemistry).
Outline Statement Approaches Test-cases Results Conclusions
Hybrid approach considered by Rolls Royce - summary
In the procedure just described, a stiff PDE is handled in threestages:
1 the computation of a less-stiff PDE via CFD software (step 1),
2 the post-processing of the CFD solution (steps 2-4),
3 the computation of a detailed chemical network model viaChemkin (step 5).
Outline Statement Approaches Test-cases Results Conclusions
Specific objective of working group
The working group has been asked to help with step 3: thesegmenting of the fluid domain into subdomains that, in some way,correspond to largely-separate reaction chambers.
Outline Statement Approaches Test-cases Results Conclusions
Approaches
A quick literature survey indicates a variety of existing methods forvarious applications. The group focussed on understanding andtesting some of those existing classical or not-so-classicalapproaches.
1 Greek methods: Q-criterion, δ,λ2, eigen helicity
2 critical points method
3 symbolic dynamic method
4 geometric methods
5 Haller’s method (reference suggested by industrial partner)
Outline Statement Approaches Test-cases Results Conclusions
Greek methods - general framework
Local vortex-identification criteriaGiven the vector field ~u := (u(x , y), v(x , y)) the deformationmatrix
∇~u := ∇(
uv
)=
(ux uy
vx vy
)is computed. This matrix is written in terms of its symmetriccomponent and its antisymmetric component
∇~u = S + Ω =1
2
(∇~u + (∇~u)T
)+
1
2
(∇~u − (∇~u)T
)(1)
Outline Statement Approaches Test-cases Results Conclusions
Q criterion
The Q criterion segments the domain by identifying the points atwhich Ω is “bigger” than S as being points within a vortex and thepoints at which Ω is “smaller” than S as being points outside avortex. Hunt et al. use the matrix norm
‖Ω‖2 := Tr(
Ω ΩT)
= Ω211 + Ω2
12 + Ω221 + Ω2
22.
That is, ∇~u is viewed as a vector in R4 with the usual dot product.In this context, S and Ω are orthogonal.
S ∈ span
1001
,
100−1
,
0110
Ω ∈ span
01−1
0
The Q criterion is then
‖S(x , y)‖2 < ‖Ω(x , y)‖2 =⇒ (x , y) is in a vortex
Outline Statement Approaches Test-cases Results Conclusions
This can be written in terms of the matrix invariants, Tr(∇~u) andDet(∇~u):
Tr(∇~u)2 − 2Det(∇~u) < 0 =⇒ (x , y) is in a vortex (2)
and
incompressible flow and Det(∇~u) > 0 =⇒ (x , y) is in a vortex(3)
Outline Statement Approaches Test-cases Results Conclusions
∆ criterion
The ∆ criterion identifies a vortex as a region where ∇~u hascomplex eigenvalues. For the 2-d flow, the eigenvalues of ∇~u are
1
2Tr(∇~u)± 1
2
√Tr(∇~u)2 − 4Det(∇~u)
If we compare the projection approach (now to be called theQ-criterion) to the eigenvalue approach (the ∆ criterion)
Tr(∇~u)2 − 4Det(∇~u) < 0 =⇒ (x , y) is in a vortex (4)
and
incompressible flow and Det(∇~u) > 0 =⇒ (x , y) is in a vortex(5)
Outline Statement Approaches Test-cases Results Conclusions
λ2 criterion
Jeong and Hussain proposed seeking minima of the pressure as amethod for identifying vortices. Taking the gradient of theNavier–Stokes equations and writing it in terms of the matrices Sand Ω yields
DSij
Dt− νSij ,kk + Ωik Ωkj + Sik Skj = −1
ρp ,ij . (6)
A local pressure minimum requires that the Hessian of p have twopositive eigenvalues. If the flow is steady and the viscosity is zerothen the Hessian of the pressure is simply Ω2 + S2 and
Ω2+S2 has two negative eigenvalues =⇒ (x , y) is in a vortex(7)
Outline Statement Approaches Test-cases Results Conclusions
Special cases: 2d incompressible flows
If the flow is incompressible then the eigenvalues of Ω2 + S2 are−Det(∇~u) with multiplicity two and so
incompressible flow and Det(∇~u) > 0 =⇒ (x , y) is in a vortex(8)
Hence the Q method, the ∆ criterion, and the λ2 method areidentical for 2-d incompressible flows but could have differentresults for compressible flows.
Outline Statement Approaches Test-cases Results Conclusions
Eigen helicity method
The Q, ∆, and λ2 methods relied fully on the vorticity tensor Ωand the strain rate tensor S . However, eigenvectors of the strainrate tensor S do not necessarily align with the vortex tubes. Thiscan lead to spurious predictions of the location of vortexboundaries.Levy et al. introduced the Helicity method which uses thenormalized helicity Hn to extract vortex core lines. Hn is the cosineof the angle between v and ω = ∇× u
Hn =v · ω|v||ω|
. (9)
The sign of Hn indicates the direction of swirl (clockwise oranticlockwise) and the sign changes whenever there is a transitionbetween primary and secondary vortices.
Outline Statement Approaches Test-cases Results Conclusions
Zhang et al. propose a new scheme that presents a differentalignment of the vorticity vector with the full eigen system of thevelocity gradient tensor.The quantity He is defined as follows.
He =nswirl · ω|nswirl ||ω|
(10)
where nswirl = −(e1 × e2)i/2, and e1 and e2 are the twoeigenvectors corresponding to the complex conjugate eigenvaluesof ∇u. By analogy with the Helicity method, the vortex corecorresponds to regions where the magnitude of He is maximum.
Outline Statement Approaches Test-cases Results Conclusions
Critical points methods
critical point: vortex centre or saddle point?
classification using the properties of critical point’s Jacobian,∇U which is defined as: ∂u
∂x∂u∂y
∂v∂x
∂v∂x
Calculating the negative trace q and determinant, r of this matrix:
q = −(∂u
∂x+∂v
∂y) (11)
r =∂u
∂x
∂v
∂y− ∂u
∂y
∂v
∂x(12)
If r < 0 the point is a saddle point and if r > q2
4 it is a vortexcentre.
Outline Statement Approaches Test-cases Results Conclusions
Symbolic dynamics method
Given u(x , y), v(x , y) and a grid decomposition of thedomain, does some mass initially in one region end up inanother region?
With this information construct the connectivity matrix:M(i , j) = 1 if some of the gas in region i went into region j ,and zero otherwise.
jth− column the entries with value 1 will tell us which regionssend material into the jth region. (directed graph)
Mp(i , j) contains the number of paths of length p from regioni into region j .
Regions that receive more mass than others have largercolumn sums. Mass cycles will be detected as recurrences
Outline Statement Approaches Test-cases Results Conclusions
For example:
M =
0 1 1 10 0 1 00 0 1 01 0 1 0
, M2 =
1 0 3 00 0 1 00 0 1 00 1 2 1
, M3 =
0 1 4 10 0 1 00 0 1 01 0 4 0
We can see a cycle in the entries M(1, 1) and M(1, 4) This cycleindicate the presence of a vortex.The Algorithm:First, locate the regions that are receiving more mass. We do thisdetection, we add up all the columns of the powers of the matrix.This will give us a function on the domain: for each element of thepartition, the number of ways there are to send mass to thatelement in the number of iterations that we pick for ourexperiment. Call this function F
Outline Statement Approaches Test-cases Results Conclusions
Second, having located these regions, we can build level sets in thefollowing way:
1 Take the domain and color the local maximums of F with adark shade.
2 Using M, look at the elements around that gave mass to thiselement of the partition and color them with a lighter shade.
3 Do the same for M2 with a lighter shade.
4 Repeat this process for regions that were colored already.
5 If we find an element of the partition that is darker than theone we want to color, this means that this element gives moremass to another local maximum. So we have found a borderbetween two vortexes.
Outline Statement Approaches Test-cases Results Conclusions
Geometric methods
Instead of physical quantities use geometric properties ofcurves such as streamlines and pathlines.
Sadarjoen and Post have used the curvature centre methodfor vortex detection.
Calculate curvature centre for different points on thestreamlines.
Vortex regions can be identified on this grid of curv. pts:curvature centre points accumulate in the center of vortices.
This method would detect easily circular vortices.
Limitations since centre points might spread out such as inthe case of elliptic vortices.
Outline Statement Approaches Test-cases Results Conclusions
Geometric methods - Advantages
The algorithm is parameter-free.
Arbitrary shaped vortex regions can be extracted.
Vector fields with and without divergence can be handledalike.
The method is grid-independent.
inputs: streamline tracing and critical points. Vortex regions aredefined in the case of a divergence-free vector field as a union ofclosed streamlines winding around a common center.
Outline Statement Approaches Test-cases Results Conclusions
The method applies a rotation matrix on the vector field. It thenuses a loop which is a positively oriented closed streamline in therotated vector field. Such a loop intersects the original vector fieldat a constant angle since rotating the vectors around that angleyields a vector field tangential to the loop.Vortex region candidates which are extracted are bounded by loopsthat start and end at saddle points. Due to continuity, at least onesaddle is included in the closure of a bounded region Ri . Thevortex region candidates Ri are disjunct but may be nested.
Outline Statement Approaches Test-cases Results Conclusions
Haller’s methods
Provides an “objective” definition of a vortex: one that doesnot depends on the frame of reference. A region that islabelled a “vortex” will still be labelled as such even if there’sa change of coordinates: x = Q(t)x + b(t)
Although based on ”Normal Hyperbolicity”, it’s not hard tocode up.
The method is defined for 3-d flows and reduces to 2-d flowsin a natural manner.
Unfortunately, the method does not generalise to compressibleflows.
Outline Statement Approaches Test-cases Results Conclusions
Test-cases
We focussed on 2-d (planar or axisymmetric) flows.
1 test-case 1: ABC flow: 2-D planar analytical solution of 2dEuler equations
2 test-case 2: Immersed droplet flow (test-case 1 in Haller’spaper): analytical axisymmetric solution for Stokes equations
3 test-case 3: Driven cavity flow: direct numerical solution for2d Navier-Stokes equations (downloaded fromwww.cavityflow.com)
4 test-case 4: Flow around a cylinder: Lattice-Boltzmannsolution to 2d Navier-Stokes equations, computed by TimReis.
Outline Statement Approaches Test-cases Results Conclusions
Test-case 1 : ABC flow
0 10 20 30 40 50 60 70 80 90 1000
10
20
30
40
50
60
70
80
90
100
Outline Statement Approaches Test-cases Results Conclusions
Test-case 2 : Immersed droplet flow
Outline Statement Approaches Test-cases Results Conclusions
Test-case 3 : Driven cavity flow, a. Reynoldsnumber=1000
Outline Statement Approaches Test-cases Results Conclusions
Test-case 3 : Driven cavity flow, b. Reynoldsnumber=10000
Outline Statement Approaches Test-cases Results Conclusions
Test-case 3 : Driven cavity flow, c. Reynoldsnumber=20000
Outline Statement Approaches Test-cases Results Conclusions
Test-case 4 : Flow around a cylinder
Outline Statement Approaches Test-cases Results Conclusions
Results - Greek- Q criterion
Note: For 2d flows: identical results from other classical greekapproaches.
10 20 30 40 50 60 70 80 90
10
20
30
40
50
60
70
80
90
Figure: Q-criterion, ABC flow.
Outline Statement Approaches Test-cases Results Conclusions
10 20 30 40 50 60 70 80 90
10
20
30
40
50
60
70
80
90
Figure: Q-criterion, Immersed drop.
Outline Statement Approaches Test-cases Results Conclusions
100 200 300 400 500
50
100
150
200
250
300
350
400
450
500
550
Figure: Q-criterion, driven cavity, Re=1000.
Outline Statement Approaches Test-cases Results Conclusions
100 200 300 400 500
50
100
150
200
250
300
350
400
450
500
550
Figure: Q-criterion, driven cavity, Re=10000
Outline Statement Approaches Test-cases Results Conclusions
100 200 300 400 500
50
100
150
200
250
300
350
400
450
500
550
Figure: Q-criterion, driven cavity, Re=20000.
Outline Statement Approaches Test-cases Results Conclusions
50 100 150 200
20
40
60
80
100
120
140
160
180
200
Figure: Q-criterion, Rolls-Q.
Outline Statement Approaches Test-cases Results Conclusions
Results: comparison Q-criterion and eigen-helicity
(b) Q method
(c) He method
Cylinder
Cylinder
Cylinder
(a) Streamlines
Outline Statement Approaches Test-cases Results Conclusions
We see that the Q-method and the He method qualitativelyidentify the vortex regions. They also omit parts of the vortexregions and also identify certain vortex-free regions as beingvortices in some cases. It is disappointing to find that we do notget better results with the He-method. One possible improvementis to use a normalized version of the Eigen Helicity as shown below-
H∗e =
|He |2
max|He |2
(13)
which takes values in the range [0, 1]. The vortices correspond tothe points H∗
e = 1 and one could use the threshold valueH∗
e = 0.95 to trace the boundaries of the vortex regions
Outline Statement Approaches Test-cases Results Conclusions
Results - critical points
The method is applied to a 2D cavity driven flow. Thezero-contour lines of u and v intersect within the potential cellshowing the critical point is located inside the cell (Figure 1).
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.57 0.575 0.58 0.585 0.59 0.595 0.6
y
x
u=0v=0
Figure: Critical point’s location in the potential cell.
Outline Statement Approaches Test-cases Results Conclusions
Applying the method, the location of the vortex centre is obtainedinside the domain as shown in figure 2.
0 0.2 0.4 0.6 0.8 1 1.2 1.40
0.2
0.4
0.6
0.8
1
1.2
1.4
x vortex =0.58783y vortex =0.75064
Figure: Centre of the vortex in the flow vector field.
Outline Statement Approaches Test-cases Results Conclusions
Results - geometric
(Note: Published results, not implemented by working group.)Here are two figures using the boundary loop and the λ2 methodsto detect vortrices for time-dependent simulation of the Karmanvortex behind a cylinder.
Figure: Boundary Loop method
Figure: λ2 method
From the pictures, it is interesting to see how the extracted vortexregions with the boundary loop method coincide better with theintuitive idea of a vortex region than the results obtained with theλ2 method. This might be explained by the fact that the Boundaryloop method try to extract the maximum region of swirling flowwhile the λ2 method is better at defining the core region of thevortex region.Note that this method can be apply to 2D slices through a 3D field.
Outline Statement Approaches Test-cases Results Conclusions
Conclusions
cheap and easy classical methods can get gross map ofrecirculating regions
these have been tested on simple 2d synthetic and classicalflows
we foresee problems though with more arbitrary cases
the predicted zones are not well delineated however
thresholding did not improve the results
perhaps Haller’s method should be programmed
recommend combining this method with other diagnostics ormin flux identification or with avg Temp and mixture fractionas was done previously
given the other approximations of this process this may beadequate
Most of these methods are extendable to 3D andcompressibility
Outline Statement Approaches Test-cases Results Conclusions
Further work and recommendations
either implement the Haller method (if easily tractable)
or combine these methods with a technique to separate therecirculation regions based on:
a crude expansion of the identified regions perhaps using somediagnostic based on local velocitiesa technique to find the ”dividing” streamline between theregions by - identifying either a no or low normal flux line(streamline) - or identifying a line along which the curl ofneighbouring velocities is zero (parallel) and the scalar productis negative (opposite directions) [This would only cover casesof abutting recirculation zones of opposite direction - whichhappens often but not exclusively].some existing method (that we could not find but think mustexist) to find a trough on a topological map. Basically thisoperation would be to calculate the path of lowest flux (orequivalently the locus of local minima).
look at how meteorologists divide streamlines?
Outline Statement Approaches Test-cases Results Conclusions
Further work and recommendations- 2
fluxes between the regions would be easy with the CFDsoftware if it is indeed flux-based
Residence times are also assumed to be calculated as in theprevious Rolls Royce work.
it would be interesting to check the impact of switching theorder of the zone-extraction post-processing and thetime-averaging of the CFD data - due to reactionnonlinearities, output from Chemkin might be very sensitive tothe ordering.