View
32
Download
0
Category
Preview:
DESCRIPTION
Paul Fischer Mathematics and Computer Science Division Argonne National Laboratory. Hisham Bassiouny, Henry Tufo University of Chicago Francis Loth, Seung Lee University of Illinois, Chicago Jerry Kruse Juniata College - PowerPoint PPT Presentation
Citation preview
Spectral Element Methods for Transitional Flows in Complex Geometries
Paul Fischer
Mathematics and Computer Science Division
Argonne National Laboratory
fischer@mcs.anl.govwww.mcs.anl.gov/~fischer
Hisham Bassiouny, Henry Tufo University of Chicago
Francis Loth, Seung Lee University of Illinois, ChicagoJerry Kruse Juniata College
Julie Mullen Worcester Polytechnic Inst.
Mathematics and Computer Science Division, Argonne National Laboratory
Accurate simulation of long-time dynamics pattern formation in nonequilibrium flows convection dynamics in deep atmospheres transitional boundary layers heat transfer enhancement vascular flows
Motivation: Spectral Element Simulations
Mathematics and Computer Science Division, Argonne National Laboratory
Outline
Temporal/Spatial Discretization Spectral element method Filter-based stabilization
Computational complexity - pressure solve Pressure projection Overlapping Schwarz
Transition in vascular flows
Mathematics and Computer Science Division, Argonne National Laboratory
Navier-Stokes Equations
Reynolds number Re ~ 1000 - 2000 small amount of diffusion highly nonlinear (small scale structures result)
Time advancement: Nonlinear term: explicit (3rd-order AB or characteristics) Stokes problem: pressure/viscous decoupling:
3 Helmholtz solves for velocity (consistent) Poisson equation for pressure
Filter velocity field at end of each time step for stability (FM 2001)
Mathematics and Computer Science Division, Argonne National Laboratory
Characteristics-Based Convection Treatment(OIFS Scheme - Maday, Patera, Ronquist 90, Characteristics - Pironneau 82)
Mathematics and Computer Science Division, Argonne National Laboratory
Unsteady Stokes Problem at Each Step
linear (allows superposition) implicit (large CFL, typ. 2-5) symmetric positive definite operators (conjugate gradient iteration)
Mathematics and Computer Science Division, Argonne National Laboratory
Spatial Discretization: Spectral Element Method (Patera 84, Maday & Patera 89) Variational method, GL quadrature.
Domain partitioned into K high-order quadrilateral (or hexahedral) elements (decomposition may be nonconforming - localized refinement)
Trial and test functions represented as Nth-order tensor-product polynomials within each element. (N ~ 4 -- 15, typ.)
KN 3 gridpoints in 3D, KN 2 gridpoints in 2D.
Converges exponentially fast with N for smooth problems.
3D nonconforming mesh for arterio-venous graft simulations:K = 6168 elements, N = 7
Mathematics and Computer Science Division, Argonne National Laboratory
Examples of Local Spectral Element Basis Functions in 2D
N th-order Lagrangian interpolants based on the Gauss-Lobatto-Legendre quadrature points in [-1,1]2
N = 4
N = 10
Mathematics and Computer Science Division, Argonne National Laboratory
PN - PN-2 Spectral Element Method for Navier-Stokes (MP 89)
Gauss-Lobatto Legendre points(velocity)
Gauss Legendre points(pressure)
Velocity, u in PN , continuous Pressure, p in PN-2 , discontinuous
Mathematics and Computer Science Division, Argonne National Laboratory
Spectral Convergence - Analytic Navier-Stokes Solution (Kovasznay, 1948)
Mathematics and Computer Science Division, Argonne National Laboratory
Numerical dispersion as a function of polynomial degree
Convection of non-smooth data on a 32x32 grid (K1 x K1 spectral elements of order N). (cf. Gottlieb & Orszag 77)
Mathematics and Computer Science Division, Argonne National Laboratory
Filter-Based Stabilization
Stabilization essential for high Reynolds number flows
At end of each time step, Interpolate u onto GLL points for PN-1
Interpolate back to GLL points for PN
Results are smoother with linear combination: (F. & Mullen 01)
(Gottlieb et al., Don et al., Vandeven, Boyd, ...)
F(u) = (1-) u + IN-1 u ( ~ 0.05 - 0.2)
Preserves interelement continuity and spectral accuracy
Equivalent to multiplying by (1-) the N th coefficient in the expansion
u(x) = uk k (x) (Boyd 98)
k (x) := Lk(x) - Lk-2(x)
F1(u) = IN-1 u
Mathematics and Computer Science Division, Argonne National Laboratory
Numerical Stability Test: Shear Layer Roll-Up (Bell et al. JCP 89, Brown & Minion, JCP 95, F. &
Mullen, CRAS 2001)
2562
2562
1282 2562
25621282
Mathematics and Computer Science Division, Argonne National Laboratory
Spatial and Temporal Convergence (FM, 2001)
Base velocity profile and perturbation streamlines
Error in Predicted Growth Rate for Orr-Sommerfeld Problem at Re=7500
Mathematics and Computer Science Division, Argonne National Laboratory
1D Steady-State Convection-Diffusion Equation
For steady convection-diffusion equation:
ux = uxx + f , x [0,1], f =1
Galerkin formulation is unstable when the number of degrees-of-
freedom is odd and 0 [Canuto 88].
N=16N=15
Solid - unfilteredDashed - filtered
Mathematics and Computer Science Division, Argonne National Laboratory
Filter-Based Stabilization
For unsteady convection-diffusion equation:
ut + ux = uxx + f , x [0,1], f =1
can show that steady state discrete solution for the filtered problem satisfies
( A + C + H (F-1- I) ) u = B f
where
A = SEM Laplacian
C = (skew symmetric) convection operator
H = discrete Helmholtz operator ( A + B/t )
B = SEM mass matrix
The term H (F-1- I) controls the singular mode when 0
and the number of degrees of freedom is odd (N even).
Mathematics and Computer Science Division, Argonne National Laboratory
Filter-Based Stabilization
Note that filtering only the (nonlinear) convective term is insufficient to provide stability.
For example, for the unsteady convection-diffusion equation one has:
Hun+1 = (t -1 B - F C ) un +B f
which solves the steady state problem
( A + F C) u = B f
In this case, there is no stability when 0 since F C is singular
when the number of degrees of freedom is odd (N even).
Mathematics and Computer Science Division, Argonne National Laboratory
Filtering permits Re99 > 700 for transitional boundary layer calculations
blow up
Re = 700
Re = 1000
Re = 3500
Mathematics and Computer Science Division, Argonne National Laboratory
Consistent Splitting for Unsteady Stokes(MPR 90, Perot 93, Couzy 95)
E - consistent Poisson operator for pressure, SPD boundary conditions applied in velocity space preconditioned with overlapping Schwarz method (Dryja & Widlund 87, PF 97, FMT 00)
Mathematics and Computer Science Division, Argonne National Laboratory
Computational Complexity
Pressure solve, Epn = gn, most expensive step.
Remedies:
Compute best fit, p*, in space of previous solutions:
|| pn - p*||E || pn - q||E for all q in span{pn-1, . . . , pn-l}
two additional matrix-vector products per step (very cheap)
Solve for perturbation, pn - p*, using overlapping Schwarz PCG
tensor-product based local solves fast parallel coarse-grid solve using sparse-basis projection
Mathematics and Computer Science Division, Argonne National Laboratory
Initial guess for Axn = bn via projection ( A=E )
Mathematics and Computer Science Division, Argonne National Laboratory
Initial guess for Axn = bn via projection
|| xn - x*||A = O(tl) + O( tol )
two additional mat-vecs per step
storage: 2+lmax vectors
results with/without projection (1.6 million pressure nodes)
Mathematics and Computer Science Division, Argonne National Laboratory
Overlapping Schwarz Precondtioning for Pressure
(Dryja & Widlund 87, Pahl 93, PF 97, FMT 00)
z = P-1 r = R0
TA0-1 R0 r + Ro,k
TAo,k-1 Ro,k r
K
k =1
Ao,k - low-order FEM Laplacian stiffness matrix on overlapping domain
for each spectral element k (Orszag, Deville & Mund, Casarin)
Ro,k - Boolean restriction matrix enumerating nodes within
overlapping domain k
A0 - FEM Laplacian stiffness matrix on coarse mesh (~ K x K )
R0T - Interpolation matrix from coarse to fine mesh
Mathematics and Computer Science Division, Argonne National Laboratory
Two-Level Overlapping Additive Schwarz Preconditioner
Overlapping Solves: Poisson problemswith homogeneous Dirichlet bcs.
Coarse Grid Solve: Poisson problemusing linear finite elements on spectral
element mesh (GLOBAL).
Mathematics and Computer Science Division, Argonne National Laboratory
Overlapping Schwarz - local solve complexity
Exploit local tensor-product structure
Fast diagonalization method (FDM) - local solve cost is ~ 4d K N(d+1) (Lynch et al 64)
Mathematics and Computer Science Division, Argonne National Laboratory
2D Test Problem: Startup flow past a cylinder (N=7)
Resistant pressure mode, p166 - p25, (K=1488)Residual history
Mathematics and Computer Science Division, Argonne National Laboratory
Impact of High-Aspect Ratio Elements
Nonconforming discretizations eliminate unnecessary elements in the far field and result in better conditioned systems.
Iteration count bounded with refinement - scalable
Mathematics and Computer Science Division, Argonne National Laboratory
Fast Parallel Coarse-Grid Solves: x0 = A0-1
b0
Problem: coarse-grid solve important for conditioning
A0 sparse, relatively small (n ~ P)
A0-1
completely full
x0 , b0 - distributed vectors
all-to-all communication
(F 96, Tufo & F 01)
Any single element in b0 will have a nontrivial influence
throughout the domain, and hence, across all processors.
Mathematics and Computer Science Division, Argonne National Laboratory
Fast Parallel Coarse-Grid Solves: x0 = A0-1
b0
Usual approaches: Collect b0 onto each processor and solve A0x0 = b0 redundantly
Communication O(n), no parallelism. OK for P ”64.
Precompute A0-1
and store row-wise on processors. Communication O(n log P). OK for P ”512.
Current approach:
quasi-sparse factorization: A0-1 =XXT
OK for P ”10000.
x0 = XXTb0 - computed as projection onto sparse basis (parallel
matrix-vector product)
choose columns of X to be as sparse as possible
(F 96, Tufo & F 01)
Mathematics and Computer Science Division, Argonne National Laboratory
XXT Communication Complexity for 7x7 Example
Nested Dissection Ordering
Processor Map
n1/2 log P
X
Mathematics and Computer Science Division, Argonne National Laboratory
Intel ASCI Red Performance: Poisson problem on q x q mesh
q=63 q=127
latency*2 log P curve is best possible lower bound
Mathematics and Computer Science Division, Argonne National Laboratory
Weakly Turbulent Vascular Flows
Vascular flows AV-graft failure
importance of hemodynamic forces (shear, pressure, vibration) in disease genesis and progression
Stenosed carotid arteries turbulence a distinguishing feature of severely stenosed (constricted) arteries high wall-shear (mean and oscillatory) can possibly lead to
embolisms (plaque break-off) thrombis formation (clotting)
Turbulence computations 1-3 orders of magnitude more difficult than laminar (healthy) case
Mathematics and Computer Science Division, Argonne National Laboratory
Arterio-Venous Grafts
PTFE plastic tubing surgically attached from artery to vein (short-circuit)
Provides a port for dialysis patients, to avoid repeated vessel injury
Failure often results after 3 months from occlusion (cell proliferation) downstream of attachment to vein, where flow is weakly turbulent
Flow characteristics of the AV graft-vein junction
Over pulsatile cycle, 1000 ReG 2500 (ReV = 1.6 ReG)
Separation region downstream of the toe
Significant velocity and pressure fluctuations
(audible or palpable “bruit”)
LDA Flow Model (F. Loth, N. Arsalan, UIC)
Distal VenalSegment Inlet
GraftInlet
“toe”
Mathematics and Computer Science Division, Argonne National Laboratory
Partition bifurcation into 3 branches.
3-way partition avoids‘figure 8’ cross-section
Sweep each branch with standard hex-circle decomposition
Meshing Algorithm for Bifurcating Vessels
Mathematics and Computer Science Division, Argonne National Laboratory
Mathematics and Computer Science Division, Argonne National Laboratory
Coherent Structures in AV-Graft, Re=1820
Close-up of coherent vortical structures in PVSvisualized with the 2 criterion of Jeong & Hussain
(JFM’95)
Mathematics and Computer Science Division, Argonne National Laboratory
Coherent Structures in AV-Graft, Re=1820
Mathematics and Computer Science Division, Argonne National Laboratory
Mathematics and Computer Science Division, Argonne National Laboratory
Mathematics and Computer Science Division, Argonne National Laboratory
Close-up of in-plane velocity distributions at two axial slices
K=6168, N=7, Re=1820
Mathematics and Computer Science Division, Argonne National Laboratory
<u>
<u’>
<v’>
Experimental/Numerical Comparison at Re=1820
LDA measurements (Arsalan 99) SEM simulation results
Spanwise Vorticity ComparisonSpanwise Vorticity Comparison
Re = 1060
Re = 1820
Mathematics and Computer Science Division, Argonne National Laboratory
Hoodmean = 350peak = 1500 - 2500
Floormean = 600peak = 750
InstantaneousAveragerms
Hood
Floor
WSS along hood and floor (dynes/cm2)
axial distance from toe (Z/D)
position (cms)
time (secs) (715 Hz)
Axial velocity vs. time
Characterization of AV Graft Hemodynamics
Mathematics and Computer Science Division, Argonne National Laboratory
Summary
Spectral Element Simulations well suited to minimally dissipative flows exponential convergence in space third-order accurate in time
Capabilities Transitional-Reynolds number applications in complex geometries Weakly turbulent vascular flow simulations ~ 20 days/cardiac
cycle on 32 processors (700 Hz signal) Able to fully characterize hemodynamic environment for AV graft
Current / Future Algorithmic Issues improved linear solvers for complex domains (multigrid?)
improved mesh quality (smoothing) condition number can be strongly influenced by mesh deformation
fluid-structure interaction (linear elasticity)
Mathematics and Computer Science Division, Argonne National Laboratory
Mathematics and Computer Science Division, Argonne National Laboratory
Mathematics and Computer Science Division, Argonne National Laboratory
Mathematics and Computer Science Division, Argonne National Laboratory
Example: A=LU, A - tridiagonal
U U-1
PUPT X X = (PUPT)-1
Lexicographical(Std.) Ordering
O(n2) nonzeros
O(n log n) nonzeros
NestedDissectionOrdering
Mathematics and Computer Science Division, Argonne National Laboratory
Mathematics and Computer Science Division, Argonne National Laboratory
Mathematics and Computer Science Division, Argonne National Laboratory
Mathematics and Computer Science Division, Argonne National Laboratory
Mathematics and Computer Science Division, Argonne National Laboratory
Mathematics and Computer Science Division, Argonne National Laboratory
Mathematics and Computer Science Division, Argonne National Laboratory
Parallel Coarse-Grid Solves
Problem: x0 = A0-1
b0
x0 , b0 - distributed vectors, A0-1
completely full, all-
to-all communication
Usual approaches: Collect b0 onto each processor and solve A0x0 = b0 redundantly
Communication O(n), no parallelism. OK for P ”64.
Precompute A0-1
and store row-wise on processors. Communication O(n log P). OK for P ”512.
Current approach: compute sparse factorization of A0
-1 .
(F 96, Raghavan 98, Tufo & F 01)
Mathematics and Computer Science Division, Argonne National Laboratory
Cast x0 = A0-1b0 as x0 = XXTb0 (sparse matrix-vector product)
U U-1
PUPT X X = (PUPT)-1
StandardOrdering
O(n2) nonzeros
O(n log n) nonzeros
Tridiagonal Matrix Example
NestedDissectionOrdering
Mathematics and Computer Science Division, Argonne National Laboratory
Application -Hemisphere/Flat-Plate Mesh
K = 8168, coarse grid dimension n = 10142, P = 1024 XXT time < 1% of Navier-Stokes solution time
Mathematics and Computer Science Division, Argonne National Laboratory
Numerical SimulationsMesh Reconstruction from CT data
Axial spacing 1 mm
In plane resolution
~ 0.5 x 0.5 mm
Mathematics and Computer Science Division, Argonne National Laboratory
Axial Velocity Component vs. Time, Re=1820
Position (cm)
Time (secs)
715 Hz
Mathematics and Computer Science Division, Argonne National Laboratory
Hoodmean = 350peak = 1500 - 2500
Floormean = 600peak = 750
(in dynes/cm2)
Instantaneous
Average
rms
Hood
Floor
rms
WSSalong hood and floor
(dynes/cm2)
Axial distance from toe (Z/D)
Wall Shear Stress @ Re=1820
Mathematics and Computer Science Division, Argonne National Laboratory
Transient Wall Shear, Re=1820, Steady Inflow
Mathematics and Computer Science Division, Argonne National Laboratory
Stress Distribution on Venal Surface, Re=1820
Mathematics and Computer Science Division, Argonne National Laboratory
Validation: Arterio-Venous Graft Studies(Arsalan, Lee, Loth - UIC, Fischer - ANL)
AV Graft - Re=1820
Recommended