• No results found

Higher order, locally conservative temporal integration methods for modeling multiphase flow in porous media

N/A
N/A
Protected

Academic year: 2020

Share "Higher order, locally conservative temporal integration methods for modeling multiphase flow in porous media"

Copied!
8
0
0

Loading.... (view fulltext now)

Full text

(1)

Higher order, locally conservative temporal integration methods for

modeling multiphase ow in porous media

C.T. Miller a

,M.W. Farthing ay

, C.E. Kees az

,and C.T. Kelley bx

a

Center for the Advanced Study of the Environment

Department of Environmental Sciences and Engineering, Box 7400

University ofNorth Carolina

Chapel Hill, NC27566-7400, USA

b

Center for Research inScientic Computation

Department of Mathematics, Box 8205

North CarolinaState University

Raleigh,NC 27695-8205, USA

Traditionalapproaches for solving single-and multiphase uid ow in porous medium

systems rely on low-order methods in both space and time. Temporal approximations

often use either xed time-step methods or empirically based adaptive strategies that

adjust the time step based on the number of iterationsrequired for the nonlinear solver

to converge. Over the last several years, several single- and multiphase ow problems

havebeen solved using higherorder temporalintegration methods based onformalerror

estimation and control. This approach has led to robust and eÆcient solvers. In this

work, we summarize the recent advances in applying higher order temporal integration

methodsto a range of multiphase ow problems. We also extend this work to derive an

approximate solutionto Richards' equationusing mixed nite element methods inspace

and to formulate asolution approachfor three-phase ow.

1. INTRODUCTION

Overthe lastseveral years, advancementsinmodelingmultiphaseowandtransportin

porousmediumsystemshaveledtoincreasedeÆciency androbustnessinsimulatingsuch

systems; improved spatial and temporal discretizationschemes and advances in iterative

algebraicsolver approaches [1,2] have been largely responsible.

We have seen steady progress in temporal integration schemes. Usinglow-order

xed-time-step solution approaches for diÆcult multiphase ow problems is known to be

in-eÆcient [3]. Moreover, while heuristic time-adaption schemes are more eÆcient than

xed-time-step methods, they have given way to more sophisticated schemes: temporal

SupportedbyNIEHS grant2P42 ES05948,NSFgrantDMS-0112653,andNCSC

y

SupportedbyaDOEComputationalSciencesFellowship,NSFgrantDMS-0112653,andNCSC

z

UndertheauspicesofUCAR,VisitingScientistProgram,P.O.Box3000,Boulder,CO80307-3000,USA

(2)

integration methods estimateand controlthe temporal error through adaptingboth the

temporalstep size and the order of the temporal approximation. They are an especially

appealingclassofmethods[4],andconsiderableworkhasbeendoneonusingthem,along

with standardnite dierenceapproaches, toapproximate Richards'equation(RE)[3{8]

and, more recently, two-phase ow[9].

Since many subsurface systems are irregularly shaped, the ideal spatial discretization

approach would capture such features eÆciently. Because of this trait, nite element

approaches have received considerable attention in the literature, but standard

Bubnov-Galerkin methodsdonot conservemass locally. Mixednite element methods(MFEMs)

do not suer from this restriction, but only recently have these methods been combined

with adaptive higher order temporal integration schemes for saturated ow [10]. To

the best of our knowledge, the literature does not yet include treatments of multiphase

ow problems approximated using higher order adaptive step-size temporal integration

combined with MFEM approaches in space. We believe such approaches are promising.

Further, we know of no published report of a higher order temporalintegration

approx-imation for three-phase ow in porous media, althoughthis too seems like a potentially

useful advance.

Theoverallgoalofthisbriefworkistoadvancehigherordertemporalintegration

meth-ods to solve multiphase ow problems. Specically, our objectives are (1) to summarize

aneective and eÆcient approachfor adaptive, higherorder temporalintegration; (2)to

formulate ahigherorder MFEM modelforRE;(3)toillustratethe adaption of temporal

approximation order and step size for a eld-scale application; and (4) to formulate an

adaptivehigher order temporalapproximationfor three-phase ow inporous media.

2. TEMPORAL INTEGRATION

Low-order temporal integration approaches have been the standard for solving the

partial dierential equations (PDE's) that arise from water resources problems. As an

alternative, the method of lines (MOL) approach applies anapproximate method to the

spatial derivatives present in the original PDE and then approximates the resultant set

of ordinary dierential equations (ODE's) using the methods available for solving such

systems. Thesetwomethodsareequivalentwhenthesamelow-ordermethodsinspaceand

timeare usedtosolvethe originalPDE. However, decouplingthe solutionapproachusing

the MOL allows for the straightforward use of sophisticated time integration strategies

that have some attractive properties: excellent stability and accuracy, a range of orders

of approximation, adjustable step size, and built-in error estimation and control and to

meet user-specied criteria. MOL approaches clearly provide a robust and eÆcient way

to solve many water resources problems, althoughmany unanswered questions remain.

Choosing an integration strategy is one such question. We have investigated a range

of temporal integration approaches and advocate using a dierential algebraic equation

(DAE) approachbased on higher order,xed-leading-coeÆcientbackward dierence

for-mulas (FLCBDF's)[4,11]. We consider DAE's of the form

f[t;y(t);y 0

(t)]=0 (1)

(3)

variables with respect totime, y 0

. Applying a FLCBDF approximation toEquation (1)

yieldsa system of potentiallynonlinear algebraicequationsof the generalform

f(t;y; y+)=0 (2)

where isa constantthat depends onthe step size andorder of the approximation,and

isa constantrelated tothe predicted solution atthe unknown time step, y p

l +1

, wherea

predictor-corrector approachisused [4]. Equation(2) istypicallysolved usinganinexact

Newton method of the form

J m

Æ m

= f (3)

where J is the Jacobian of the nonlinear equations described by Equation (2), m is an

iteration index, and Æ is a Newton correction. The linear system solves needed to

com-pute Æ can be performed using direct methods for simple one-dimensional problems or

preconditioned Krylov-subspace methodsfor multidimensionalapplications [9].

The orderof the approximation and the step size can be chosen tomeet user-specied

absolute and relative error tolerances. Using a weighted norm approximation of the

dierence between the corrector and predictor solutionof the generalform

l =K

l kÆ

l k

W

<1 (4)

where istheerrorestimate,K

l

isacoeÆcientdependentontheFLCBDFapproximation,

and the subscript W denotes a weighted norm that depends on the user-specied error

tolerances. Complete details of this approachare available inthe literature [9,12].

We have found DAE approaches more eÆcient than the more traditional MOL

ap-proaches. The latter require the system of ODE's to be written in explicit form [3].

Moreover, DAE approachesbenetfromobject-oriented solutionmethods[9],and MOL/

DAEmethodscanbeused toderivemassconservativeapproximationsofmultiphase ow

by specifying explicit algebraicconstraints[9].

3. RICHARDS' EQUATION

We have shown that anMOL/DAE approachcan yielda robust and eÆcient

approxi-mationofRE[3]. Wehavealsoshownhowthistimeintegrationapproachcanbeextended

toderive a locallyconservativemixed hybridnite element(MHFEM) approximation to

single-phase ow in three spatial dimensions; such eorts also result in an eÆcient

solu-tion[10]. Here, weextendourpreviousworkbyderivinganadaptivetemporalintegration

approach forRE, which may be appliedto irregulardomainswhile maintaininga locally

conservative solution.

Although using a MHFEM discretization with a MOL/DAE approach showed several

benetsoverstandard approachesfor owinaconned, heterogeneousaquifer,anumber

of issuesshould be consideredbeforetranslatingits advantages for single-phaseow over

to RE.As mightbeexpected, these issues arise fromthe need to account forthe

nonlin-earity inherent in multiphase ow. Specically, the nonlinear nature of the constitutive

(4)

and spatial discretizations used as well as the techniques required to solve the resulting

nonlinear and linear systems.

Inadditiontothe standard MFEM [13], both the MHFEM[14] andthe enhanced

cell-centered dierence method (ECDM) [15,16] have been applied to RE. The ECDM can

alsobeseenasanonstandardnitedierencetechnique obtainedfromastandardMFEM

approachby usingcertainnumericalquadrature rules[17]. Likeothermorecommon

spa-tialapproximations, the MFEM discretizationshavevaried intheir choice of formulation

for RE, p-S-k relations,linear and nonlinear solution strategies, and time discretization.

The temporal approximations, however, have all been low order [13,14,16] with either

xed time step or empirical adaption. The low-order time discretization is of particular

concern for applying the MHFEMto the pressure head formof RE (PRE), whichcauses

signicantmass balanceerrors for certainproblems [18].

The PREcan begiven by

@^

@ +^

@

@ !

@

@t

+ r(q)^ =0 in[0;T] (5)

where and [0;T] are the physical and temporal domains, is the pressure head, ^is

a normalized density, and is the volume fraction for the aqueous phase. We use the

commonextension of Darcy's lawto variably saturated ow [19]

q= k

r ( )K

s

(r d)^ (6)

Here K

s

is the saturated hydraulic conductivity, k

r

is the relative permeability, and d is

a vector accounting for the acceleration of gravity. To describe the interdependence of

uid pressure, saturation, and therelative permeabilitywe use the p-S-k relationsof van

Genuchten [20] and Mualem [21], while we assume slight compressibility for the aqueous

phase density [22].

The boundary and initialconditions are given by

= b

on

D

;t2[0;T]

un=u b

on

N

;t 2[0;T]

= 0

in ;t=0 (7)

where u isthe mass ux.

TheECDM discretizationissuitablefor use with logicallyrectangularmeshes. Details

canbefoundin[17,15,23]. Briey,itformulatesanexpandedsystemforEquations(5){(7)

which includes anadjusted gradient

~

u = (r d)^ with

u = k^

r K

s ~

u (8)

It proceeds by identifying subdomains of over which the problem coeÆcients vary

smoothly. It forms local approximations of Equations (5){(8) over these subdomains

using the lowest order Raviart-Thomas space and certain numerical quadrature rules

[17]. Theselocalapproximationsarethen coupledby dening Lagrangemultipliersalong

(5)

We extended previous work by developing an a temporally adaptive ECDM solution

for a PRE. We used an available ECDM spatial discretization method [17], an

object-oriented FLCBDFDAE temporalintegration approach[4], and solved the resultant

sys-tem ofequationsusing apreconditionedgeneralized minimalresidual(GMRES) iterative

solver from the PETSc library [24]. Simulations were performed on an IBM RS/6000

SP supercomputer with 720 processors at the North Carolina Supercomputing Center

(www.ncsc.org).

As an example, we consider a typical unsatured ow problem for a hillslope domain.

The idealized hillslope geometry is dened in terms of a mapping F from a reference

domain ^

=[0;24][0;1] (meters) given by x

1 =x^

1 and

x

2 =

(

^ x

2

+0:1^x

1

sin(x^

1

=25) for x^

1 16

^ x

2

+1:477 otherwise

(9)

Giventhismapping,itisnaturaltoidentifytwosubdomainsof ^

forx^

1

<16andx^

1 >16.

The saturated conductivity and parameter values for the van Genuchten-Mualem p-S-k

relationswere then taken fromthe literature for a sand medium[25].

The initial value of was set to be at hydrostatic equilibrium with the water table,

locatedatx

2

=1=3. Noowconditionsweredened fortherightandbottomboundaries.

Homogeneous Neumann conditions were also used on the left boundary except for the

region given by x

2

1=3 where Dirichlet conditions were set to hydrostatic equilibrium

with the water table. The top boundary uxinto the domain wasgiven by

u b

= (

2sin(t) for5:33 x

2

13:33and t1=24

0 otherwise

(10)

A contour plot of at t = 0:1 on a 9060 mesh is shown in Figure 1. Figure 1

alsotraces the the choice of timestep and temporalapproximationorder overthe course

of the simulation for the DAE integrator with both full adaption and the integration

order restricted toone (BE-A).The data points shown are values averagedover every 10

time steps. The fully adaptive simulation required 336 seconds of total CPU time on 32

processors while the BE-A run required 1085.0 seconds. The BE-A run would be more

eÆcient than typical low-order heuristically based integration schemes and much more

eÆcient than xed-time-step low-order integration schemes. These results are consistent

with previous ndings documenting the advantages of the MOL/DAE approach used in

this work [3,10].

4. THREE-PHASE FLOW

Given a mass-conservative spatial discretization, we can obtain a mass conservative

temporal discretization for RE and two-phase ow models by employing a semi-explicit

index-oneformulation,aswe demonstrated in [22]. Here we briey outlinethis approach

for three-phase owmodels.

The semi-discretesystem of three-phase ow equationscan be writtenas

@

(

i;j

i;j )=O

di;j

(6)

0.1

0.12

0.14

0.16

0.18

0.2

0.22

0.24

0.26

2

4

6

8

10

12

14

16

18

20

22

0.5

1

1.5

2

x [m]

y [m]

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1

0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5

time [days]

average order DAE

average

t

10

4

DAE

average

t

10

4

BE−A

Figure1. left: Spatialdistributionofwaterphasevolumefractionforhill-slopesimulation.

right: Temporalintegration order and step size as afunction of simulation time for

hill-slope simulation.

where N is the number of discrete spatial nodes, O

d

is the approximate divergence of

uid phase mass ux at each node, and the subscripts w, n, and a are phase qualiers

correspondingtoawettingandanon-wettinguidphaseandair,respectively. Tosimplify

the presentation we have assumed that the form of the equations are the same on the

interior and boundary of the domain. We have also assumed that relations suÆcient to

obtainformalclosureareavailable;thatis,ateachspatialnodethereisathreecomponent

vector of unknowns, y

j

. Forexample, we could choose y

j =( w;j ; w;j ; g;j

) for common

closurerelations. Followingtheapproachgivenin[22],wederiveasemi-explicitindex-one

DAE fromeqn 11given by

@M i;j @t O di;j = 0 M i;j i;j (y i ) i;j (y i

) = 0

for i=w;n;a and j =1;:::;N (12)

where M represents the uid phase bulk density in the porous medium. Applying the

BDF time discretization aboveyields the nonlinear system

M j;i + i O dj;i = 0 M j;i j (y i ) j (y i

) = 0

for i=w;n;a and j =1;:::;N (13)

Thissystemcanbesolved eÆcientlyusingacombinationofstraightforwardalgebraic

ma-nipulationsandNewton'smethodasshownin[22]. Aslongaseqn13issolvedaccurately,

the numerical model is mass conservative in the sense that the discrete change in uid

(7)

5. CONCLUSIONS

The method of lines is a straightforward approach for applying sophisticated adaptive

temporalintegration methodstoPDE modelsofowand transportinporous media. We

havedemonstrated, in particular,that transient nonlinear models ofmultiphase owcan

besolvedeÆcientlyusingvariableorder,variablestepsize BDFmethodsinthecontextof

theMOL.Further,thesemethodscanbeappliedinconjunctionwithsophisticatedspatial

discretization techniques, such as MFEMs, toyield accurate, mass conserving, solutions

to complexmultiphase problems onirregulardomains.

REFERENCES

1. C. T. Kelley. Iterative Methods forLinear and Nonlinear Equations. SIAM,

Philadel-phia, 1995.

2. E. W. Jenkins, C. E.Kees, C. T. Kelley,and C. T. Miller. An aggregation-based

do-maindecompositionpreconditionerforgroundwaterow. SIAMJournal onScientic

Computing, 23(2):430{441, 2001.

3. M. D. Tocci, C. T. Kelley, andC. T. Miller. Accurateand economicalsolutionof the

pressure-head formof Richards'equation by the methodof lines. Advances in Water

Resources, 20(1):1{14,1997.

4. C. E.Kees and C. T. Miller. C++implementationsofnumericalmethodsforsolving

dierential-algebraicequations: Designand optimizationconsiderations. Association

for Computing Machinery, Transactions on Mathematical Software, 25(4):377{403,

1999.

5. C. T. Millerand C. T. Kelley. A comparison of stronglyconvergent solutionschemes

for sharp-frontinltrationproblems. InA. Peters, G.Wittum, B.Herrling,U.

Meiss-ner, C. A. Brebbia, W. G. Gray, and G. F. Pinder, editors, Computational Methods

in Water Resources X, volume Vol. 1,pages 325{332, Amsterdam, The Netherlands,

1994. Kluwer AcademicPublishers.

6. C. T. Kelley, C. T. Miller,and M.D. Tocci. Termination ofNewton/chord iterations

and the method of lines. SIAM Journal on Scientic Computing, 19:280{290,1998.

7. M. D. Tocci, C. T. Kelley, C. T. Miller, and C. E. Kees. Inexact Newton

meth-ods and the method of lines for solving Richards' equation intwo space dimensions.

Computational Geosciences, 2(4):291{309,1999.

8. G.A.WilliamsandC.T.Miller. Anevaluationoftemporallyadaptivetransformation

approaches for solving Richards' equation. Advances in Water Resources, 22(8):831{

840, 1999.

9. C. E. Kees and C. T. Miller. Higher order time integration methods for two-phase

ow. in press, Advances in Water Resources, 2002.

10. M. W. Farthing, C. E. Kees, and C. T. Miller. Mixed nite element methods and

higher-order temporal approximations. Advances in Water Resources, 25(1):85{101,

2002.

11. J. F.Kanney,C. T. Miller,and D.A.Barry. Comparisonoffully coupledapproaches

(8)

ValueProblemsinDierential-AlgebraicEquations. SocietyforIndustrialandApplied

Mathematics, Philadelphia, PA, 1996.

13. L. M. Chounet, D. Hilhorst, C. K. Y. Jouron, and P. Nicolas. Simulation of water

ow and heat transfer in soilsby means of a mixed nite element method. Advances

in Water Resources, 22(5):445{460,1999.

14. L. Bergamaschi and M. Putti. Mixed nite elements and Newton-type linearizations

for the solution of Richards' equation. International Journal for Numerical Methods

in Engineering, 45:1025{1046,1999.

15. Carol A. San Soucie. Mixed nite element methods for variably saturated subsurface

ow. PhD thesis, Rice University, Houston, Texas, 1996.

16. Carol S.Woodwardand ClintN.Dawson. Analysisofexpanded mixed niteelement

methods for a nonlinear parabolic equation modeling ow into variably saturated

porous media. SIAM Journal of NumericalAnalysis,37(3):701{724,2000.

17. T. Arbogast, C. N. Dawson, P. T. Keenan, M. F. Wheeler, and I. Yotov. Enhanced

cell-centered nite dierences forellipticequations ongeneralgeometry. SIAM

Jour-nal on Scientic Computing, 19(2):404{425,1998.

18. M.A.Celia,E.T.Bouloutas,andR.L.Zarba. Ageneralmass-conservativenumerical

solution for the unsaturated ow equation. Water Resources Research, 26(7):1483{

1496, 1990.

19. M. D. Tocci, C. T. Kelley, andC. T. Miller. Accurateand economicalsolutionof the

pressure-head formof Richards'equation by the methodof lines. Advances in Water

Resources, 20(1):1{14,1997.

20. M. Th. van Genuchten. A closed-form equation for predicting the hydraulic

con-ductivity of unsaturated soils. Soil Science Society of America Journal, 44:892{898,

1980.

21. Y. Mualem. A new model for predicting the hydraulic conductivity of unsaturated

porous media. Water Resources Research,12(3):513{522,1976.

22. C. E. Kees and C. T. Miller. Higher order time integration methods for two-phase

ow. Advances in Water Resources, 2002. In Press.

23. Ivan Yotov. Mixed Finite Element Methods for Flow in Porous Media. PhD thesis,

Rice University, Houston,Texas, 1996.

24. S. Balay, W.D. Gropp, L.C. McInnes, and B.F.Smith. PETSc homepage. Technical

Report http://www.mcs.anl.gov/petsc, ArgonneNational Laboratory,1998.

25. C. T. Miller, G.A. Williams,and C. T. Kelley. EÆcient and robust numerical

mod-eling of variably saturated ow in layered and porous media. In XII International

References

Related documents