Upload
james-hateley
View
21
Download
2
Embed Size (px)
DESCRIPTION
PPT from ccse Berkeley lab
Citation preview
Block-Structured Adaptive MeshRenement
Lecture 2Incompressible Navier-Stokes Equations
Fractional Step Scheme
1-D AMR for classical PDEs hyperbolic elliptic parabolic
Accuracy considerations
Bell Lecture 2 p. 1/27
Extension to More General SystemsHow do we generalize the basic AMR ideas to more general systems?
Incompressible Navier-Stokes equations as a prototype
Ut + U U +p = U
U = 0
Advective transport
Diffusive transport
Evolution subject to a constraint
Bell Lecture 2 p. 2/27
Vector eld decompositionHodge decomposition: Any vector field V can be written as
V = Ud +
where Ud = 0 and U n = 0 on the boundary
The two components, Ud and are orthogonal
U dx = 0
With these properties we can define a projection P
P = I (1)
such thatUd = PV
with P2 = P and ||P|| = 1
Bell Lecture 2 p. 3/27
Projection form of Navier-StokesIncompressible Navier-Stokes equations
Ut + U U +p = U
U = 0
Applying the projection to the momentum equation recasts the system asan initial value problem
Ut + P(U U U) = 0
Develop a fractional step scheme based on the projection form ofequations
Design of the fractional step scheme takes into account issues that willarise in generalizing the methodology to
More general Low Mach number models
AMR
Bell Lecture 2 p. 4/27
Discrete projectionProjection separates vector fields into orthogonal components
V = Ud +
Orthogonality from integration by parts (with U n = 0 at boundaries)
U p dx =
U p dx = 0
Discretely mimic the summation by parts:
U GP =
(DU) p
In matrix form D = GT
Discrete projectionV = Ud +Gp
DV = DGp Ud = V Gp
P = I G(DG)1DBell Lecture 2 p. 5/27
Spatial discretizationDefine discrete variables so that U , Gp defined at the same locations andDU , p defined at the same locations.
D : Vspace pspace G : pspace Vspace
Candidate variable definitions:
u,v
p
u
v
p u,v,p
Bell Lecture 2 p. 6/27
Projection discretizationsWhat is the DG stencil corresponding to the different discretization choices
Non-compact stencils decoupling in matrix
Decoupling is not a problem for incompressible Navier-Stokes withhomogeneous boundary conditions but it causes difficulties for
Nontrivial boundary conditions
Low Mach number generalizations
AMRFully staggered MAC discretization is problematic for AMR
Proliferation of solversAlgorithm and discretization design issues
Bell Lecture 2 p. 7/27
Approximate projection methodsBased on AMR considerations, we will define velocities at cell-centers
Discrete projectionV = Ud +Gp
DV = DGp Ud = V Gp
P = I G(DG)1D
Avoid decoupling by replacing inversion of DG in definition of P by astandard elliptic discretization.
Bell Lecture 2 p. 8/27
Approximate projection methodsAnalysis of projection options indicates staggered pressure has "best"approximate projection properties in terms of stability and accuracy.
DUi+1/2,j+1/2 =ui+1,j+1 + ui+1,j ui,j+1 ui,j
2x
+vi+1,j+1 + vi,j+1 ui+1,j ui,j
2y
Gpij =
pi+1/2,j+1/2+pi+1/2,j1/2pi1/2,j+1/2pi1/2,j1/2
2x
pi+1/2,j+1/2+pi1/2,j+1/2pi+1/2,j1/2pi1/2,j1/2
2y
Projection is given by P = I G(L)1D
where L is given by bilinear finite element basis
(p,) = (V,)
Nine point discretizationBell Lecture 2 p. 9/27
2nd Order Fractional Step SchemeFirst Step:
Construct an intermediate velocity field U:
U Un
t= [UADV U ]n+
1
2 pn1
2 + Un + U
2
Second Step:
Project U onto constraint and update p. Form
V =U
t+Gpn
1
2
SolveLp
n+ 12 = DV
SetU
n+1 = t(V Gpn+1
2 )
Bell Lecture 2 p. 10/27
Computation of Advective DerivativesStart with Un at cell centersPredict normal velocities at cell edges using variation of second-orderGodunov methodology un+1/2
i+1/2,j, v
n+1/2i,j+1/2
MAC-project the edge-based normal velocities, i.e. solve
DMAC(GMAC) = DMACUn+1/2
and define normal advection velocities
uADVi+1/2,j = u
n+1/2i+1/2,j
Gx, vADVi,j+1/2 = vn+1/2i,j+1/2
Gy
Use these advection velocities to define [UADV U ]n+1/2.
: u : v :
Bell Lecture 2 p. 11/27
Second-order projection algorithmProperties
Second-order in space and time
Robust discretization ofadvection terms using modernupwind methodology
Approximate projection formula-tion
Algorithm components
Explicit advection
Semi-implicit diffusion
Elliptic projections5-point cell-centered9-point node-centered
How do we generalize AMR to work for projection algorithm?
Look at discretization details in one dimensionRevisit hyperbolic
Elliptic
Parabolic
Spatial discretizations
Bell Lecture 2 p. 12/27
Hyperbolic1dConsider Ut + Fx = 0 discretized with an explicit finite differencescheme:
Un+1i Uni
t=F
n+ 12
i1/2 F
n+ 12
i+1/2
x
In order to advance the composite solution we must specify how tocompute the fluxes:
tf
xf xc
j1 j J J+1
Away from coarse/fine interface the coarse grid sees the average offine grid values onto the coarse grid
Fine grid uses interpolated coarse grid data
The fine flux is used at the coarse/fine interface
Bell Lecture 2 p. 13/27
HyperboliccompositeOne can advance the coarse grid
tf
(J1) J J+1
then advance the fine grid
tf
j1 j (j+1)
using ghost cell data at the fine level interpolated from the coarse griddata.
This results in a flux mismatch at the coarse/fine interface, which createsan error in Un+1J . The error can be corrected by refluxing, i.e. setting
xcUn+1J := xcU
n+1J t
fF
cJ1/2 + t
fF
fj+1/2
Before the next step one must average the fine grid solution onto thecoarse grid. Bell Lecture 2 p. 14/27
HyperbolicsubcyclingTo subcycle in time we advance the coarse grid with tc
tc
(J1) J J+1
and advance the fine grid multiple times with tf .
tf
tf
tf
tf
j1 j (j+1)
The refluxing correction now mustbe summed over the fine grid timesteps:
xcUn+1J := xcU
n+1J
tcF cJ1/2 +
tfF fj+1/2
Bell Lecture 2 p. 15/27
AMR Discretization algorithmsDesign Principles:
Define what is meant by the solution on the grid hierarchy.
Identify the errors that result from solving the equations on each levelof the hierarchy independently (motivated by subcycling in time).
Solve correction equation(s) to fix the solution.
For subcycling, average the correction in time.
Coarse grid supplies Dirichlet data as boundary conditions for the finegrids.
Errors take the form of flux mismatches at the coarse/fine interface.
Bell Lecture 2 p. 16/27
EllipticConsider xx = on a locally refined grid:
xf xc
j 1 j J J + 1
We discretize with standard centered differences except at j and J . Wethen define a flux, cfx , at the coarse / fine boundary in terms of f and c
and discretize in flux form with
1
xf
(
cfx
(j j1)
xf
)= j
at i = j and
1
xc
((J+1 J )
xc cfx
)= J
at I = J .
This defines a perfectly reasonable linear system but ...
Bell Lecture 2 p. 17/27
Elliptic compositeSuppose we solve
1
xc
((I+1 I)
xc
(I I1)
xc
)= I
at all coarse grid points I and then solve
1
xf
((i+1 i)
xf
(i i1)
xf
)= i
at all fine grid points i 6= j and use the correct stencil at i = j, holding thecoarse grid values fixed.
j 1 j J J + 1
Bell Lecture 2 p. 18/27
Elliptic synchronization
The composite solution defined by c and f satisfies the compositeequations everywhere except at J.
The error is manifest in the difference between cfx and(JJ1)
xc.
Let e = . Then he = 0 except at I = J where
he =1
xc
((J J1)
xc cfx
)
Solve the composite for e and correct
c = c+ ec
f = f
+ ef
The resulting solution is the same as solving the composite operator
Bell Lecture 2 p. 19/27
Parabolic compositeConsider ut + fx = uxx and the semi-implicit time-advance algorithm:
un+1i uni
t+f
n+ 12
i+1/2 f
n+ 12
i1/2
x=
2
((hun+1)i + (
hu
n)i
)
t
xf xc
j1 j J J+1
Again if one advances the coarse and fine levels separately, a mismatch inthe flux at the coarse-fine interface results.
Let ucf be the initial solution from separate evolution
Bell Lecture 2 p. 20/27
Parabolic synchronizationThe difference en+1 between the exact composite solution un+1 and thesolution un+1 found by advancing each level separately satisfies
(I t
2h) en+1 =
t
xc(f + D)
t f = t (fJ1/2 + fj+1/2)
t D =t
2
((uc,n
x,J1/2+ uc,n+1
x,J1/2) (ucf,nx + u
cf,n+1x )
)
Source term is localized to to coarse cell at coarse / fine boundary
Updating un+1 = un+1 + e again recovers the exact composite solution
Bell Lecture 2 p. 21/27
Parabolic subcycling
Advance coarse grid
tc
(J1) J J+1
Advance fine grid r times
tf
tf
tf
tf
j1 j (j+1)
The refluxing correction now must be summed over the fine grid timesteps:
(I tc
2h) en+1 =
tc
xc(f + D)
tc f = tc fJ1/2 +
tffj+1/2
tc D =tc
2(uc,n
x,J1/2+ uc,n+1
x,J1/2)
tf
2(ucf,nx + u
cf,n+1x )
Bell Lecture 2 p. 22/27
Spatial accuracy cell-centeredModified equation gives
comp = exact + 1comp
where is a local function of the solution derivatives.
Simple interpolation formulae are not sufficiently accurate for second-orderoperators
yc
yc
xc-f
xc-fxc
Bell Lecture 2 p. 23/27
Convergence resultsLocal Truncation Error
D Norm x ||L(Ue) ||h ||L(Ue) ||2h R P
2 L 1/32 1.57048e-02 2.80285e-02 1.78 0.842 L 1/64 8.08953e-03 1.57048e-02 1.94 0.963 L 1/16 2.72830e-02 5.60392e-02 2.05 1.043 L 1/32 1.35965e-02 2.72830e-02 2.00 1.003 L1 1/32 8.35122e-05 3.93200e-04 4.70 2.23
Solution Error
D Norm x ||Uh Ue|| ||U2h Ue|| R P
2 L 1/32 5.13610e-06 1.94903e-05 3.79 1.922 L 1/64 1.28449e-06 5.13610e-06 3.99 2.003 L 1/16 3.53146e-05 1.37142e-04 3.88 1.963 L 1/32 8.88339e-06 3.53146e-05 3.97 1.99
computed = exact + L1
Solution operator smooths the errorBell Lecture 2 p. 24/27
Spatial accuracy nodalNode-based solvers:
Symmetric self-adjoint matrix
Accuracy properties given by approximation theory
Bell Lecture 2 p. 25/27
RecapSolving coarse grid then solving fine grid with interpolated Dirichletboundary conditions leads to a flux mismatch at boundary
Synchronization corrects mismatch in fluxes at coarse / fine boundaries.
Correction equations match the structure of the process they arecorrecting.
For explicit discretizations of hyperbolic PDEs the correction is anexplicit flux correction localized at the coarse/fine interface.
For an elliptic equation (e.g., the projection) the source is localized onthe coarse/fine interface but an elliptic equation is solved to distributethe correction through the domain. Discrete analog of a layerpotential problem.
For Crank-Nicolson discretization of parabolic PDEs, the correctionsource is localized on the coarse/fine interface but the correctionequation diffuses the correction throughout the domain.
Bell Lecture 2 p. 26/27
Efciency considerationsFor the elliptic solves, we can substitute the following for a full compositesolve with no loss of accuracy
Solve c = gc on coarse grid
Solve f = gf on fine grid using interpolated Dirichlet boundaryconditionsEvaluate composite residual on the coarse cells adjacent to the finegrids
Solve for correction to coarse and fine solutions on the compositehierarchy
Because of the smoothing properties of the elliptic operator, we can, insome cases, substitute either a two-level solve or a coarse level solve forthe full composite operator to compute the correction to the solution.
Source is localized at coarse cells at coares / fine boundary
Solution is a discrete harmonic function in interior of fine grid
This correction is exact in 1-D
Bell Lecture 2 p. 27/27
Extension to More General SystemsVector field decomposition Projection form of Navier-StokesDiscrete projectionSpatial discretizationProjection discretizationsApproximate projection methodsApproximate projection methods2nd Order Fractional Step SchemeComputation of Advective DerivativesSecond-order projection algorithmHyperbolic--1dHyperbolic--compositeHyperbolic--subcyclingAMR Discretization algorithmsEllipticElliptic -- compositeElliptic -- synchronizationParabolic -- compositeParabolic -- synchronizationParabolic -- subcyclingSpatial accuracy -- cell-centeredConvergence resultsSpatial accuracy -- nodalRecapEfficiency considerations