Chapter 2

CHA P TER II
Finite Element and Finite Difference Methods for Continuous Flows in Porous Media
THOMAS F. RUSSELL AND MARY FANETT WHEELER
1. Introduction. Miscible displacement is an enhanced oil-recovery process that has attracted considerable attention in the petroleum industry over the last 30 years. It involves injection of a solvent at certain wells in a petroleum reservoir, with the intention of displacing resident oil to other wells for production. This oil may have been left behind after primary production by reservoir pressure and secondary production by waterflooding. The economics of the process can be precarious, because the chemicals it requires are expensive and the performance of the displacement is by no means guaranteed. Complex physical behavior in the reservoir will determine whether enough additional oil is recovered to make the expense worthwhile. Mathematically, the process is described by a convection-dominated parabolic partial differential equation for each chemical component in the system. These equations are nonlinear and strongly coupled. By summing the component equations, one can obtain an equation that determines the pressure in the system; this nonlinear equation is elliptic or parabolic, according to whether the system is incompressible or compressible. Thus, in this problem one encounters elliptic, parabolic, and near-hyperbolic equations with complicated nonlinear behavior. It is an intricate problem to approximate numerically, and good numerical modeling is critical in the industry for accurate prediction of costly projects. This paper is concerned with several numerical procedures developed in the last few years for simulation of miscible displacement. This represents work of Douglas N. Arnold, David C. Brown, Bruce L. Darlow, Jim Douglas Jr., Todd Dupont, Richard E. Ewing, Richard P. Kendall, Thomas C. Potempa, Jean E. Roberts, Peter H. Sammon, and the authors. Theoretical and numerical results have been obtained for a wide variety of methods. This research has been mostly directed at a model system of two coupled equations for incompressible twocomponent single-phase displacement, but the techniques can be extended to more complex systems representative of realistic reservoir flow. In particular, we believe that similar procedures are appropriate for immiscible as welt as miscible displacement, and therefore for most problems of reservoir simulation. In the
35
Downloaded 24 Feb 2012 to 130.88.140.110. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
36
MATHEMATICS OF RESERVOIR SIMULATION
model problem, the dependent variable of most physical interest is the concentration c(x, t), 0 c ss 1, of the solvent in the Huid mixture. Figures 1 through 3 give a qualitative idea of the time evolution of level curves of c in a typical stable displacement. The displacement is on an areal five-spot pattern as in Fig. 3 of Chapter I. The paper is organized as follows. Section 2 discusses the physics of miscible and immiscible displacement at some length in order to justify our inclusion of physical diffusion-like terms in our numerical procedures. There has been some sentiment in the mathematica) literature that these displacement processes possess so little parabolic character that they should be modeled as discontinuous hyperbolic flows; we have taken a different view, and we present supporting reasoning here. In the next section we derive the partial differential equations that represent the chosen physical model. Section 4 surveys the usual timestepping methods considered in the petroleum industry for such problems and indicates why we decided to use sequential procedures. As a result, we treat the two model equations, for pressure and concentration, with different sets of methods described in 5 and 6. Before presenting these methods we note the procedures that have been used in the industry and point out some of their inadequacies. Methods for the pressure equation include standard and mixed finite elements with near-well logarithmic singularities removed or not. We demonstrate first-order convergence of the commonly used inconsistent blockcentered finite difference method by showing that it is equivalent to a certain
mixed method with a particular quadrature rule. For concentration, we introduce
interior-penalty Galerkin procedures on continuous and discontinuous spaces, a modified method of characteristics using finite differences or finite elements, and a finite difference method using cel) balances based on finite element concepts.
cl
o^
09
Production well
Injection well
FIG. 1 . Level curves of solvent concentration in the early stages of miscible displacement.
M ETHODS FOR CONTI N UOUS FLOWS
37
0 ^
'
Production well Injection well
F16. 2. Level curves of solvent concentration approaching solvent breakthrough.
Convergence theorems are cited for all procedures for both equations, together with numerical computations for the new methods. Section 7 shows how the pressure and concentration schemes have been tied together to solve the coupled system of equations, again presenting theoretical and numerical results. Section 8 indicates ideas and directions for future research, and the final section gives conclusions and the authors' opinions about the results obtained so far. This paper is primarily expository. Our principal objective is to show what has been done and why it has been done, and to give the reader a feel for the results.
Accordingly, we shall try to keep technical details to a minimum and to avoid the
o,l
c
. 0
Production well
Injection welf.
Fc. 3. Level curves of solvent concentration after solvent breakthrough.

38
lengthy proofs that pervade numerical analysis of coupled nonlinear systems. References for further details are indicated in the text. 2. Discussion of physical processes. As noted in Chapter I of this volume, reservoir simulation must begin with a physical model that adequately describes the significant flow phenomena. The nature of the physical model will affect the appropriateness of various possible mathematical and numerical models. Our mathematical models for the displacement processes considered in this chapter are partial differential equations of convection-diffusion type, whose solutions exhibit moving fronts that are sharp but continuons. The corresponding numerical methods we shall discuss are based on finite differences and finite elements, which, in their standard forms, treat diffusive processes effectively but have difficulty with convection. We shall present modifications of these standard forms designed to handle convection better without losing the ability to simulate diffusion rigorously. This section sets the background for these mathematical and numerical models by surveying the engineering literature to find out what processes should be considered in a physical model. It was pointed out in 3 of Chapter 1 that compositional, chemical, and thermal simulation combine aspects of miscible and immiscible displacement when viewed mathematically. Since our objective is to test new numerical ideas at a research level, we confine our discussion to miscible and immiscible displacement. We emphasize miscible displacement because most of our work has been in that direction, but we shall indicate why we believe that analogous ideas make sense for immiscible displacement. Miscible and immiscible displacement both involve convection, or physical transport, of fluids through the porous medium. At a macroscopic level, this process is governed by Darcy's law, (3.1) and (3.22) of Chapter I. We must stress the word "macroscopic"; at the microscopie level of a pore size, convection is highly irregular. In fact, under certain assumptions, Darcy's law has been derived rigorously from the pore-level NavierStokes equations by a homogenization process of local volume averaging. This has been done for single-phase [891, [911, [1131 and multiphase flow [90], [92], [65], [42]. As emphasized in [1131, the characteristic length of the local volumes used for averaging must be much greater than a pore length (10 -4 meters, say) and much less than a reservoir length (10 2 to 10 3 m). For single-phase flow, the permeability and pressure in Darcy's law are volume-averaged quantities; similarly, for multiphase flow, saturations and relative permeabilities are volume-averaged. This means that in a model based on Darcy's law, one cannot deal with variations on a finer scale than that of the volume averages; in particular, one cannot deal with discontinuities. Claridge [ 15], citing geological literature, points out that rocks have characteristic scales of heterogeneity ranging from the pore level to the reservoir level; in modeling, we must volume average over heterogeneities significantly smaller than grid size, so our averaging volumes must be of the
METHODS FOR CONTINUOUS FLOWS
39
order of the grid size. Thus, microscopic heterogeneities are lost when we pass to Darcy's law; the next scale of heterogeneity must be averaged in modeling, and we shall see that this can be modeled by dispersion; heterogeneities at or larger than grid size can be modeled directly with equation coefficients. If reservoir flow processes are in fact discontinuous (we contend below that they are not), they must not be modeled by fundamentally macroscopic principles. These observations show that flow in a porous medium is to be distinguished from other phenomena of fluid dynamics. Miscible displacement is also characterized by the processes of diffusion and dispersion, about which there is considerable literature. Diffusion of one fluid into another, which is the result of random motion of molecules, is governed by the Fick equation [71] (2.1
at
'"
aV --_ d
ax
ac
where V is the volume of one fluid on the positive-x side of a plane of cross-sectional area A normai to the x-direction, C is the concentration of that Huid, and dm is the molecular diffusion coefficient (units: length 2 /time) in a nonporous medium. In a porous medium, d, is multiplied by 1 IFO, where F is the electrical resistivity factor of the formation and 0 is the porosity [711. F can usually be roughly approximated by o 2 [36], so that the porous coefficient becomes bd, as in (3.33) of Chapter 1. This coefficient is typically of the order of 10 S cm 2 /sec or 10 ; ft 2 /day [45], which is quite small. Most of our discussion will deal with dispersion, a term that covers assorted physical phenomena. At a microscopic level, it is mechanical mixing caused by flow through tiny capillaries. In a single capillary tube, the theory of Taylor [99] leads to a diffusion-like term whose coefficient is proportional to the square of fluid velocity. This would extend to a bundle of straight capillary tubes. In a pack of sand grains, Brigham et al. [ 10] observe that a process that they cal] "eddy diffusion" will take place, in which flow channels will separate and meet again, with fluids being dispersed by virtue of traveling different distances. There the amount of dispersion should depend on the distance traveled by the fluids and not on their velocity; since a diffusion-like coefficient represents fluid movement per unit time and the time corresponding to a given distance will be inversely proportional to velocity, the coefficient must be proportional to velocity. Laboratory measurements with reservoir cores (rock samples) have found coefficients proportional to velocity to approximately the power 1.2 [ 10], [71]; this has been explained by the core combining the propertjes of sand packs and capillary tubes [ 10] or by diffusion failing to equalize concentration within single pores at high velocities [71]. Perhaps more significant for our purposes is dispersion at a macroscopic level. This is caused by macroscopic variations in rock permeability and porosity, and it
40
involves differences in velocities of individual fluid elements as they move through the rock [106]. Because of the qualitative similarity between this process and microscopic mechanical mixing, it is not surprising that a distributed (log-normal) permeability system behaves as if macroscopic dispersion could be represented by a diffusion-like term [1061 with coefficient proportional to velocity. A specific process of this type, observed in [57] and called Taylor's dispersion, concerns flow at different rates in adjacent horizontal rock layers combined with vertical dispersion between the layers, resulting in a large effective horizontal dispersion; a similar process had been studied by Taylor [99], [100]. Because macroscopic dispersion is most important for us, we shall use the exponent 1 instead of 1.2 in our work. In general, we should note that mechanical mixing and, by analogy, macroscopic dispersion have two coefficients, one parallel (longitudinal) to flow and the other normal (transverse); the longitudinal coefficient may be 30 times the transverse one [711, [57]. They are commonly denoted in the literature by a,, a t (units: length) and multiplied by the interstitial velocity v (the velocity of fluid movement; u/4, if u is the Darcy velocity) to obtain a diffusion-like coefficient. This coefficient is then substituted for D/4 in an equation like that obtained from (3.32) of Chapter 1 divided through by 0; thus, a, and a, are equivalent to d, and dr as defined in (3.33) of Chapter 1. We indicate next the potential importance of dispersion in applications of miscible displacement. The overreding significance of viscous fingering was enunciated in Chapter I, and the behavior of viscous fingers is primarily controlled by mixing [73]. Specifically, transverse dispersion can enhance the effectiveness of miscible displacement by suppressing fingers altogether [ 103], [73], [15], or it can influence the geometry of fingers [96]. Longitudinal dispersion blunts the tips of fingers [44], affecting their growth patterns. In a linear system, the minimum finger width that will not be damped out is proportional to I i, where e is the displacement efficiency (ratio of injected fluid to total fluid behind the leading edge of the front; equal to 1 for a piston-like displacement) and x is the distance traveled; a similar relationship hoids in a radial system [73]. The onset of fingers is also related to dispersion, as in the theories of Perrine [74], later supported by experiments [75], and Heller [48], in which a front was stable below a critical velocity proportional to 1 /a,. Dispersion can likewise be harmful to oil recovery by miscible displacement. A typical miscible process involves injection of a solvent slug that is miscible with oil provided that its concentration exceeds a minimum that is a function of pressure [52], [44], [115]. The slug will be followed by another material so that the slug, which is expensive, may be recovered; examples include propane followed by gas, alcohol followed by water [52], and a micellar solution followed by polymer water [ 115], [39]. Dispersion can dissipate the slug below minimum miscibility concentration, rendering the process immiscible [71], [44], [1151. As
METHODS FOR CONTI N UOUS FLOWS
41
discussed in Chapter 1, this results in much residual oil being left trapped in small pores by interfacial tension. The relationship between dispersion and ion exchange is extremely important for chemical flooding. Recovery efficiency of a chemical slug depends on phase behavior, mobility, and interfacial effects, all of which depend in turn on ionic environment [56]. Qualitative differences in ion concentrations are shown in [56] depending on whether dispersion is incorporated in simulation or not, and a dispersion-induced exchange process is described in which two fluids of similar ionic properties mix to create different properties. With laboratory experiments, Gupta [43] shows that the behavior of a field micellar (sulfonate/polymer) flood cannot be explained without significant dispersion. Sulfonate partitions into the oil phase because of ion exchange, while polymer stays in the water phase. Since the dispersion coefficient for sulfonate in the oil phase is an order of magnitude larger than water-phase coefcients, sulfonate breaks through at the production well before a nondispersive model would predict. Thus, sulfonate is produced earlier than polymer, and the entire process requires more sulfonate than would be the case without dispersion. If dispersion is potentially important, as we have seen, the question remains of whether it is significant, so that its effects are feit in practical problems. As we noted, molecular diffusion is quite small; if dispersion were no more important than diffusion, its effects would be of little interest. In the late fifties and early sixties, when there was a flood of interest in miscible recovery processes, some investigators took that view. In these experiments, linear tubes were packed uniformly with unconsolidated sand [6] or glass powder [ 1031, and a twodimensional quarter five-spot pattern (Fig. 3 of Chapter 1) was packed uniformly with glass beads [55]. The porous media of petroleum reservoirs, however, are consolidated rocks, and linear [45], [96] and five-spot [44] experiments in consolidated media led to the opposite conclusion that mechanical mixing (dispersion) dominates diffusion. This conclusion was based on the observation that the outfiowing concentration profile after a flood through a consolidated medium was independent of the flow rate (velocity); as we noted above, this implies a mechanical mixing model with dispersion coefficient proportional to velocity. Indeed, the experiments of Brigham et al. [ 10] found that the microscopic heterogeneities of consolidated rock will increase the dispersion coefficient by two orders of magnitude over its value for glass beads; this supports the findings of Habermann [44], in which the mixing zone was 8 to 20 percent of the area of the quarter live-spot, as opposed to those of Lacey et al. [55], where it was 1.6 to 5 percent. Discussions published with [55] indicated serious doubts of many readers as to whether dispersion was properly scaled from the laboratory to field reservoirs in that study. The work of Koch and Slobod [52] with an unconsolidated-sand tube much longer (123 feet) than those of the other experiments concluded that dispersion was the predominant mixing mechanism.
42
Thus, it appears to us that dispersion, even at the microscopic level, is vastly more significant than molecular diffusion alone. This viewpoint is advanced in relation to viscous fingering in [45] and [44], where it is noted that fingering increases the length of the interface by making it irregular, thereby increasing the effect of dispersion. Handy [45] finds that dispersion dominates diffusion even in this case, where dispersion acts mainly through the transverse a the smaller of its two coefficients. Additionally, we have the phenomenon of macroscopic dispersion, which came to be discussed in more recent literature. It has been pointed out [ 106], [58] that the scale of macroscopic dispersion is related to the scale of heterogeneities in the reservoir, so that small laboratory models are inadequate to measure this behavior and will therefore understate the dispersion that will occur in the field. Claridge [15] echoes this and identifies how heterogeneities and dispersion interact in the all-important process of viscous fingering. Even without largescale heterogeneities, he contends that a reservoir-scale displacement will show a mixed zone of 5 to 10 percent of the total system length; large-scale heterogeneities would increase this. Field measurements of macroscopic dispersion are hard to come by, since they are complicated by nonlinear flow, heterogeneities, and operational difficulties in the field [57]. A two-well test, which is the type of interest to us, is cited in [57] leading to an estimate of 8 feet for a,. As we shall see in 3, the Peclet number (ratio of dimensionless convection and dispersion coefficients) of a miscible displacement is L/a,, if L is a characteristic reservoir length and diffusion is neglected; in this case it is of the order of 100 to 300. A tracer test in the same field considered in [43] was approximately matched by a, = 3 feet [ 116]; the maximum produced slug concentration was actually a bit higher in the simulation than in the field, indicating that the actual value of a, might be slightly greater. An extensive review of field dispersion calculations for pollutant transport in aquifers has been published in France [58]. This article compiled tracer measurements from 27 sources and related dispersion coefficients to distance traveled and flow velocity. The authors found that velocity was not important, supporting the dispersion model [106] of macroscopic mixing, but that distance was important. Longer distance correlated with higher dispersivity. LallemandBarres et al. postulated that a critical distance existed beyond which this correlation would no longer hold (presumably beyond the largest scale of heterogeneity), but that most of their cited measurements did not reach that distance. For a sizable majority of their cases, the Peclet number was between 10 and 30. The theory of [106] leads to a Peclet number of approximately 2/S 2 for a log- normal permeability distribution, where the variante S 2 is roughly the differente of the arithmetic and harmonie average permeabilities divided by the harmonie average; in a field example, S 2 was found to be about 0.1, for a Peclet number of about 20. These similar results from widely varying sources indicate to us that convection-diffusion models of this range of Peclet number are appropriDownloaded 24 Feb 2012 to 130.88.140.110. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
43
ate for simulation of miscible displacement. Even without dispersion, research with graded-viscosity slugs [95], [54] has demonstrated reduced viscous fingering for field application, forcing one to consider modeling of continuous fronts. But the views we can find in current literature [57], [1161, [ 115] do not support the contention that molecular diffusion is the dominant mixing mechanism and dispersion is negligible. It could still be questioned whether conventional convection-diffusion-type equations properly average the physical behavior over all scales of heterogeneity. The theoretical work of Warren and Skiba [1061 led to this type of model, as did the laboratory study of Yellig and Baker [ 1 15]. In the Jatter study, dispersive mixing in short heterogeneous systems was found to appear homogeneous in longer systems, so that an effective dispersion coefficient could model it. This effective coefficient was found to increase with system heterogeneity or system length. This modeling concept is supported by Whitaker [ 1 12] and Gray [411, in which local volume averaging of the dispersion equation was carried out, leading to an equation in volume-averaged variables that has the same form as the standard convection-diffusion equation. The dispersion tensor contains terms proportional to velocity (mechanical mixing model of Brigham et al. [10]) and velocity squared (capillary-tube model of Taylor [991). In our work, we can model heterogeneities at or larger than mesh scale directly with variable coefficients in the partial differential equations. This will allow detailed study of such phenomena as macroscopic viscous fingering at mesh scale. Scales between the Darcy volume averages and the mesh will be modeled with the dispersion concepts justified by this discussion. We now turn our attention to immiscible displacement and outline our reasons
for advocating similar modeling techniques for it. Comparing (3.29b) and
(3.32b) of Chapter I, we see that the immiscible analogue of dispersion is capillary pressure. This pressure differente across a phase interface within a pore is a function of the interfacial tension and the curvature of the interface. We focus now on the importante of interfacial tension in the physics of immiscible displacement. An oil globule trapped in a pore is held there by interfacial tension. There is evidente [47], [93] that when such a globule moves, it does so in small discrete jumps that occur when the local pressure gradient overcomes the interfacial tension and collapses the interface. Between these jumps, the interface slowly deforms. To recover more oil by immiscible displacement, the quotient of pressure gradient and interfacial tension must exceed a critical value; once this happens, further increases in this quotient result in significantly greater oil recovery [104], [98], [93]. The ability to recover more oil means that the residual oil saturation, which is the maximum oil saturation for which the oil relative permeability (see (3.23) of Chapter I) is zero, has been decreased with decreasing interfacial tension. More generally, the entire relative-permeability function, on which the efficiency of immiscible displacement depends very strongly, is
44
affected by interfacial tension, both for steady [59], [2] and unsteady [1051, [70], [2] flow. These irreversible forces, as opposed to convection, are responsible for the hysteresis of capillary pressure and relative permeability [47] (the functions differ depending on whether the wetting phase displaces the nonwetting phase or vice versa). All of these considerations govern the microscopic efficiency of immiscible displacement. The macroscopic efficiency of immiscible displacement is governed by its stability or lack thereof (viscous fingering). Theories of the onset of instability have found it to occur above a critical flow velocity with wavelengths (finger widths) above a cutoff dependent on interfacial tension [83], [131, [64]. Peters and Flock [76] developed a dimensionless parameter, inversely proportional to interfacial tension, such that instability develops if the parameter exceeds a critical value. A difficult issue here is that interfacial tension acts at the microscopic pore level, while macroscopic fingering demands a macroscopic concept; probably the way to resolve this is through an effective macroscopic interfacial tension [40], justified by experiments displaying a cutoff wavelength. This may possibly be best understood in terms of volume averaging that would take account of the physical interfacial tension [941. As in the miscible case, this would preclude modeling that dealt with fronts of sharpness beyond grid scale. Another important development is the increasing view that miscible and immiscible displacement share many characteristics. Perkins and Johnston [72] conducted laboratory experiments that led to a concept of "immiscible dispersion." When water was injected into a bead-packed model saturated with oil and irreducible water, small fingers initiated, broke into a tree structure, and then deteriorated into a graded saturation zone. Side-by-side injection of water and oil into linear consolidated bead models gave rise to a transverse two-phase zone with saturation profile analogous to the concentration profile of miscible displacement; the dispersion coefficient was proportional to velocity, as in the miscible case. lt was theorized that this process, like miscible dispersion, was caused by stream splitting during convection. It would seem that the macroscopic dispersion concepts should carry over to immiscible displacement as well, especially since field tests such as those that give large dispersion coefficients [43], [57], [58] are really combinations of miscible and immiscible flow regimes. For both miscible and immiscible displacement, it seems to us that the complex physical behavior is largely a function of the mathematical regularizersdispersion and interfacial tension. These processes should not be neglected in modeling; they, and how they interact with rock heterogeneities, appear to be precisely what needs to be understood. By using methods based on finite differences and finite elements, which handle these processes well, and by incorporating heterogeneities (at scales greater than volume averages) into the dispersion and permeability coefficients of the differential equations, we hope to model the physics realistically and rigorously. In particular, we can demonstrate theorems that tell us that our procedures converge to the solutions of the
45
differential equations chosen as models. We feel that procedures based on propagation of discontinuities, while leading to some excellent mathematical and numerical work with application to other fields, are not appropriate for the problems of petroleum reservoir simulation. 3. Derivation of differential system for miscible displacement. Having discussed the physical phenomena that should be included, we are in a position to derive partial differential equations describing miscible displacement. Many of the necessary ideas are the same as those given in Chapter I for single-phase flow and two-phase immiscible flow. In particular, Darcy's law and conservation of mass are fundamental. Here, however, we have two components (denote them by subscripts o and s for oil and solvent) flowing together as one phase, so we use conservation of mass for each component instead of the phase. In addition, we must deal with Huid fluxes due to diffusion and dispersion of the two components into each other. An equation of state relating density to pressure is also important in this context, but for simplicity we shall assume that the fluids are incompressible. As in Chapter 1, consider a volume element V. Recalling the volume averaging discussed in 2, V should be at least as large as the scale of the averaging volumes. Using the notation of Chapter 1, the rate of mass accumulation of component i in V is
( 3.1 )
d
dt
;
cpc (X, t)d-x,
o, s,
where c c [0, 1) is the concentration of component i in the single phase. Darcy's law ((3.1) of Chapter 1) yields the volumetric convective flow rate of the mixture per unit cross-sectional area; restricting this to component 1, we obtain (3.2) k u; _ - c'. (V p -- pgV z) = c .
;
From Fick's law (2.1), as modified in 2 for porous media, we obtain the diffusive flow rate (3.3) u;.^^rr = nrVc,
In 2, we decided to use the first-power law for dispersion, giving rise to the dispersive flow rate
act
(3.4) , __ drjuj u;.disp = diiu( el, e, 3e,
where , and , are unit vectors in the directions parallel and perpendicular to u
(in three dimensions, a third term is added with coefficient d 1 ). IE' 0 is the angle
46
between , and the cartesian unit vector x , we use the relations

l = e x cos9 + sin8,
e, = ex sine + , cos 8,
}
cos 0 =
u X ul, sin8= u y,/lul,

; ; X r
to rotate coordinates and express (3.4) as
ux ac u^. ac u u . e., + e ..+ lul Iulax Iulav IuI u u ac u 0c\ + .. e.,+ drlu1 ,: lul lui , l ul ax l ulav
w
ul ,disp = dr 1 u I
(3.5)
d, ^
l u i
i
ac; 3c\ 2 ac uX ax + u x u}, a e.r + u x u 1, aX + u ti, a e V

;
_ d / 2 ac
uI
ax
u u. ac'
x j
uu.ac;u2ac;e.
x àx + Xay
ay X +
Note that a pth-power law (e.g., p = 1.2) would lead to 1 u 1 2-1) in the denominator of (3.5). Now we combine (3.1), (3.2), (3.3), and (3.5) with the divergente theorem as in Chapter 1, equating accumulation to influx and sources:
(^Pc^) dx = iv 0 (pciu + P ,.diI + pu 31
disp
) dx
(3.6)
+
;
c ; pgdx,
where c is the specified concentration of component i at an injection well and the resident concentration at a producer, and q is the volumetric flow rate per unit volume. We assume incompressibility of the fluids and the rock, and by using (3.5) we can rewrite (3.6) in the form
( 3.7 )
ac' D DV c uc dx = ,) ^6 t (
i dx , V c q
i= o, s,
where D is a diffusion-disperson tensor given by

~ 2 !
(3.8)
X y
D = bdml
d ux
u.ru y d, i
Ar Ar
A. 2
/^ h
uxu,.
X X
lul u u.
u 2
j^
lul u u. s
u2
This tensor is also derived in Peaceman [66] . In three dimensions, one can show that the analogous tensor is
4dm l+d,ju1T+d,ju1(1 T),

47
where
u T = u z U u.
;
u r uV u.Y u i
i
u,.u:
u,. u- u .u- u-
We would like to drop the integrations in (3.7) by collapsing V to an arbitrary point, but this requires justification since V is at least on the scale of the averaging volume. The work of Whitaker [ 1 12] and Gray [411 demonstrated that, in terms of volume-averaged variables, one can arrive at the same equation we would obtain if we dropped integrations in (3.7). Hence, we have the coupled system
(3.9) d at V. (DVc, u(>,) = c,q 1 = o, s.
,
The dependent variables are u, c,,, and c,,, with the additional constraint of C + c, = 1. The two equations can be called an oil equation and a solvent equation. A common manipulation with this system is to replace one of the equations (say the oil equation) by the sum of the two. When the equations are added together, the relation c + c,. = 1 causes the accumulation and difusiondispersion terms to drop out, leaving the system 3.1Oa -D k
(3.1Ob) 0
(c)
-=
ac at --- V (D(u)Vc -- uc) cq,
where c = c,, e = c,. The first equation (3.10a) now represents conservation of mass for the total fluid, the second for the solvent component. The first equation is frequently called the pressure equation because it is an elliptic equation for pressure with a concentration-dependent coefficient. The second equation may be considered a concentration equation because it is a (usually) convectiondominated parabolic equation for concentration with coefficients depending on pressure through the Darcy velocity. Because (3.1 Ob) depends on p only through u, we may also think of (3.1Oa) as a velocity equation. In (3.10a) the behavior of as a function of c is very important to viscous fingering, displacement efficiency, and ultimate oil recovery. This behavior depends on the mobility ratio ,/, _ 9(0)/ 4(1) = M. 1f M> 1, the displacement is said to have adverse mobility ratio and fingering is expected. This derivation was not restricted to two components. We could have obtained
J
a system of n equations in (3.9) for an n-component system flowing as one phase, and then used c, + c, + + c n = 1 to replace one of the n equations with a
48
pressure equation like (3.1Oa). By also considering multiphase flow, we could derive a compositional system like (3.34) in Chapter 1, with the u's of that equation containing diffusive-dispersive terms as well as Darcy terms. Within each phase, flow would be similar to that described by (3.10), but the phase volume fraction (saturation) would multiply all the terms and change with space and time, and there would be mass transfer between phases. Likewise, the derivation extends readily to compressible flow, as in Peaceman [67], rendering the pressure equation parabolic as noted in Chapter I. For simplicity in trying new methods, we shall confine our attention primarily to the system (3.10), keeping in mind that we want ideas that will extend conveniently to more complex problems in three dimensions. The system (3.10) is taken to hold over a bounded domain 9, representing a reservoir or part of a reservoir, through a time interval J = [0, T] . All of our computational work to date has dealt with c- R', but we have theoretical results in H and eventually expect to perform three-dimensional computations with these procedures. The system requires boundary and initial conditions; in reservoir modeling, the usual boundary conditions are no-flow, representing an axis of symmetry or an impermeable boundary, (3.11a)
(3.11 b)
uv=0,
xc 0S^, tc.J,
(DOc) v c(u v) = 0, x t - 8, t c- J,
or constant - pressure, applicable when the reservoir is repressurized at the
boundary during depletion, say by an aquifer: (3.12a) (3.12b)

p (X, t) = p o , x c- 3, t ci J,
appropriate condition for c.
In our work, we shall use no-flow conditions, which are the most common in practice. Note that (3.10)(3.11) determine p only up to an additive constant and that the compatibility condition
f qdx=f o u dx =
(3.13)
u v ds = 0,
which simply says that the total fluid input to an incompressible system with no-flow boundary must be zero, is forced. Finally, an initial condition
c(x, 0) = c o (x),
x c- 2,
is required, and initial values for p are determined by (3.1Oa) and (3.13). If the system were compressible, we would need an initial condition for p as well. The system (3.10)(3.11)(3.13) is in divergente form, and for certain numerical procedures we want to consider an alternative nondivergence form. Differentiating the convection term in (3.1 Ob) with the product rule and applying
49
(3.1Oa), we see that V . (uc) = u .Vc+ c(V . u) =u.Vc-}- cq,
so that (3.10) may be rewritten (3.14a)

(3.14b)
7 . uq,
ac
t --- 7 (D(u)Vc) + u Vc = ( -- c)q.
Note that, except at the flowing wells where q is nonzero, (3.14) is the same as (3.10). If a numerical approximation of u is divergence-free away from wells (for example, the mixed methods described in 5), the same equivalence will hold numerically. If, instead of (3.11), we use nonhomogeneous Neumann boundary conditions
u v=f, (DVc) v--- c(u v) -=g,
then we see that the concentration condition can be written

(3.15)
(DVc) v fc + g.
When a Galerkin procedure is applied to (3.14b), (3.15) crises naturally from Green's theorem. Theoretical treatment of this nonlinear boundary condition was
carried out in [3 1 ], and we shall discuss the issue no further in this paper. So far we have said little about the source terms of the differential system. In our computational work, we have used rums of Dirac delta functions, representing point sources and sinks. A moment's thought makes one realize that the pressure at a flowing well will then be + or - x, according as the well is an injector or a producer. Needless to say, infinite or negative pressures make little
sense to a petroleum engineer. The truc physical pressure around a well will resemble a truncated logarithmic singularity, as in (3.19) of Chapter 1. But for the purpose of testing numerical methods in an incompressible regime, it makes sense to see how well they can handle full singularities, since the real problems are almost singular. Once we move to compressible flow, with pressuredependent physical properties, something else wilt be done. For similar reisons, we are led to theoretical questions about regularity of singular solutions and convergence ratel of numerical schemes to singular solutions. Finally, we observe the degree to which (3.1 ob) or (3.14b) is convectiondominated, as defined by its Peclet number. 1f we neglect molecular diffusion by comparison to dispersion, and if we scale by a characteristic length L (usually the distance between corners for a live-spot pattern as in Fig. 3 of Chapter 1), a one-dimensional analogue of (3.14b) becomes
0 ^ - L aX, t
,
ac d,1 u 1 aèe u ae
+ pax (` -- c)q.
50
The scaled convection/dispersion ratio L/d, is the (longitudinal) Peclet number. In two dimensions, a transverse Peclet number L/d, also applies. In 2 we observed that practical problems should have Peclet numbers in the tens or hundreds, and this is the range considered in our computational work. 4. Possible time-stepping approaches. The system (3.10) or (3.14) is amenable to many possible time-stepping ideas. These fall into three major classes, as discussed in Chapter I: IMPES (implicit pressure, explicit saturation here concentration plays the role of saturation), fully implicit, and sequential. All three are in common use in the petroleum industry. Our methods are sequential. In this section we give some details of the three alternatives and indicate why we chose to pursue sequential methods. The IMPES method is applied to a system like (3.10) or (3.14) containing a pressure equation. in passing from time level n to n + 1, one first solves (3.1 Oa) for p"+', lagging all concentration dependencies to the old level c". If the system were compressible, with parabolic (3.1Oa), the time derivative of pressure would be approximated by the backward difference quotient (p" 1 p")/Ot; hence the notion of implicit pressure. The density p would be pressure-dependent and would multiply the terms of the equation as in (3.5), and in practical problems the viscosity is also pressure-dependent. Thus, the IMPES pressure equation encountered in industry simulation is, unlike (3.1Oa), nonlinear. Having obtained p", one then solves explicitly for c" 1 . We shall illustrate this in terms of finite differences in one dimension, using the block - centered scheme most popular in the industry. Let i 1, i, i + 1 index three blocks of lengths Ox , Ax ; , zx i+ , with centers located at x_,, x ; , x ; , ,. Define the distance between centers, e.g., zx 112 = '/2(x, + Ax; j ,), and let x ; _ , 1 2, x ;+ 1/ 2 be the interface coordinates (see Fig. 4). Now approximate (3.1 Ob) by
;
(4.1)
Ci
n+l
c 7
4,
n
Ot
n n n C 1 C1 ten+,2) Ctf I Ci 1 D ui -1/2 Ci- 112 ui+1/2ci+1/2 D(ui --I/ 1 ) x; 1 /2 ax; + I/2
Ax
^n n
where (neglecting the gravity term in u)

n+I n1
X r+ 1/2 -I
4.2
-n+^ _ _
ki+112
pi + 1 pi
(4
^^ r + 1/2)
4 .3) k- = 1 LX, + Ax1 ( harmonic average ), .3 ( g )^ r12 l^ II

Ox; + Oxi-f 1
ki Ki+ 1
ii
51
^- ^x
î--1
x.
-^__
1
Axi+l
Lxi
x i _ 1 xi-
xi
x1
Fic,. 4. Block-centered finite dijference coordinates.
cl , ,
i f pr
>p i'
:
(upstream value).
c ^, if p,"'' <p
We shall have more to say about block-centered differences and the velocity approximation (4.2) in 5, and about (4.4) (known in the industry as upstream weighting) in 6. For now, simply note that (4.1) can indeed be solved algebraically for cn `'. If we multiply (4.1) by and sum over 1, then all spatial terms and cancel (recall no-flow boundary conditions), and we obtain global conservation of mass:
(4.5)

,! c.n + 1 L Xt
=
t
(/i cx; +
t
c j q l"At.
The key to an IMPES formulation is the absence of saturation (here, concentration) time derivatives in the pressure equation. More complicated systems [67], [4], [1021 are handled by taking an appropriate linear combination of the component conservation equations to make these terms disappear. Computationally, IMPES is fast because only one variable (pressure) is solved for
implicitly via a system of linear or nonlinear equations. However, the explicit scheme in (4.1) is subject to a CourantFriedrichsLewy stability condition of the form l 1 Ot//x _ 1 (possibly more stringent with significant nonlinearities), since the equation is almost hyperbolic. In the industry, this is known as a "throughput" condition, because it says that no more than one pore volume of material can be put through a grid block in one time step. In our work, we are simulating rather complex enhanced recovery techniques that require fairly fine grid resolution, particularly near wells and steep fronts. Thus, IMPES would restrict our time steps severely, and for this reason we chose not to use it. Another reason is that we are interested in finite element methods, and the accumulation integral analogous to the first term of (4.1) would have the form
O e 'fl t
dx
for test function v; this is nonzero for basis functions on adjacent elements, thus coupling them and leading to an implicit system of equations, even when all
52
spatial terms are lagged at time level n. A finite element implementation of IMPES would have to lump the accumulation term such that it looked like its finite differente analogue. The fully implicit procedure does not require a pressure equation and usually proceeds from a system of the form (3.9). Essentially, one has two equations like (4.1), except that n is replaced by n + 1 in the spatial and source terms. This nonlinear system is solved simultaneously for p (equivalently, u) and c by Newton's method. Each nonlinear iteration, typically 3 to 5 per time step, involves the solution of fully coupled linear equations for p and c. Mass conservation follows as in the IMPES case, up to the residual tolerante in the Newton iteration. For a system as simple as (3.9), it would actually be simpler to reduce to (3.10) for fully implicit as well as IMPES; in more complicated multiphase problems, with nonlinear relative permeability coefficients in the convection terms, separate pressures differing by saturation-dependent capillary pressures, and variable density and porosity in the accumulation terms, this is no longer the case. Fully implicit time stepping eliminates the stability concerns of IMPES and makes accuracy the prime criterion setting time-step size, though one still has to take care that nonlinearities not prevent Newton's method from converging. For us, this procedure has two drawbacks. First, it is expensivelinear systems of order equal to the number of equations (two here, but greater in more complex problems) times the number of nodes must be solved several times at each time
step. Because of this, virtually all chemical and compositional models in industry
use the IMPES formulation. Second, when the equations are coupled and solved simultaneously, one is forced to use the same numerical method, grid, and time step for all equations. As we have seen, the equations in (3.10), if thought of as decoupled, exhibit quite different behavior, and one would like the flexibility of choosing separate and possibly different methods. For these reasons, we decided to use sequential methods. A sequential method, like IMPES, uses a system like (3.10) and solves implicitly for p, but it also solves implicitly for c in a decoupled fashion. This means that either c in (3.1 Oa) or u in (3.1Ob) must be lagged or extrapolated in time to decouple the equations. Since tne streamlines of a displacement should not change as rapidly as the movement of the front, we expect u to be smoother in time than c; hence, we chose to extrapolate u in (3.1Ob). Thus we solve first for c, then p (or u). In fact, this allows multiple concentration time steps for each pressure time step. Specific schemes will be discussed in subsequent sections, but a general picture has the form
(4.6a)
= cn + qn + I , 0 < < a < ^3 < . . . <^ ,
53
(4.6b)
V.
(c )
?P
n i
=9
rl i
where a and /3 are fractional indices (a = 0 and /3 = 1 if both equations use the same time step) and is a linear extrapolation from u" ' and . Conservation of mass again follows. With this decoupling, both equations are linear; for more complex nonlinear systems, a Newton iteration is needed for both, and its residuals determine the mass-conservation error. We note here that if a = 0 and 03 = 1 in (4.6), we could iterate the procedure by successive substitution to reach the fully implicit solution; that is, replace Juni' by the first computed u"'', solve for c'' and u"`' again, and so on. Thus, whatever stability and ratel of convergence we can prove for our sequential procedures wilt apply allo to the fully implicit method, provided that its Newton iteration can be shown to converge to a unique solution.
5. Methods for the pressure equation. Since we are using sequential time stepping, we can consider methods for (3. l 0a) and (3. I Ob) (or (3.14b)) separate-
ly. This section deals with procedures applied to the pressure equation (3.10a). As background, we begin with descriptions of some standard procedures. For economy of notation, we consider the problem
(5.1)
(5.2)
- V. ( a ( x )V p) - C . u = q,
--(aVp)t' uv=0, x
-x Q,
lower-order gravity terms do not affect the methods or the mathematical analysis in any significant way, though they do influence physical flow patterns. For simplicity, let SZ be a unit square in R. Take a(x) (representing k/) to be possibly discontinuous, but bounded below by a positive constant. As noted in 3, compatibility demands that f q dx = 0, and p is only determined up to an additive constant. We can add the additional condition f i p dx = 0 to fix p In the petroleum industry, finite differences are used almost universally. The version best known to mathematicians, called point-centered differences in the industry, consists of choosing meshes 0 == x 0 < x, < < x k = 1 and 0 = y o <y < < y, = 1, defining interfaces x,, , /2 = (x i + x,, 1 )/2 and yj , , /2 = (y1 + y1 , )/2, and approximating (5.1) at (x i , y,) by
.
(5.3)
Xi ^ 1/2 xi 1/2
1 P, ai + I /2.j
4 1,1
P.1 a P,1 P i
I /-',1
.j
xi t I xl
xi
xi -
- (similartermsiny-direction) =q ij ,
i 5k, 05js1.
This couples (x i , yj ) to its four immediate neighbors, and hence is called a five-point scheme. The boundary condition (5.2) may be handled by reflection, requiring that p ,, j = p, 1 for all j, and so on. The point-centered procedure is
used in a small minority of industry simulators.
54
The vast majority of working codes use the block-centered method mentioned in 4. This idea is derived rather naturally from physical considerations as foliows. Return to Fig. 4 in 4, and attach constant dimensions Dy and z to the blocks in the other two coordinates; for petroleum engineers, problems that are mathematically one-dimensional are still physically three-dimensional. Since aV p represents a Darcy velocity, or volumetric flux per unit cross-sectional area per unit time, we can express the influx to block i from block i + 1 in time Ot
as Vk= a ^ ^ ,^z
pit 1 Pl
yy Ozt.
OX + 2
The coefficient is regarded as a step function, constantly a and a, on blocks i and i + 1, respectively, and if one thinks of the time to flow from x , , to x ; as the sum of times from x ; , , to x, /2 and x ;+ , /2 to x, it makes sense to take a ; +, /2 as a harmonic average. Similarly, the influx to block i from block i 1 is
; ;
VL
_ ai
p; P1 1
Lxi /2
AYAZAt.
Finally, q; is the volumetric injection per unit volume per unit time into block i, so in time At we have
V, = q 4x DyzAt .
; ;
The steady-state equation (5.1) requires that

VR + VL + V1 = 0.
When extended to two-dimensional flow, this leads to a difference equation exactly like (5.3), except that the relationships between coordinates are different; in the point-centered case we have x ;+ , /2 x ; _, / 2 _ (x, x ; ,)/2, but this does not necessarily hold in the block-centered case. Also, the block-centered version of (5.3) holds for 1 ss i 5 k, 1 5 j ss 1 with x, /2 = 0, X k , , /7 = 1, y, /2 = 0, y, + 1 /2 = 1 being boundary coordinates, and no-flow boundary conditions are easily incorporated by setting a, /2 , 2 = a k } , /z , f = a ; ,, /2 = a ;. , + 1 /2 = 0. This convenience at the boundary and the physical concept of flow between cells are the reasons why petroleum engineers prefer block-centered to point-centered schemes. For uniform mesh, it is easy to see that the two procedures are identical except for their treatment of boundary conditions (point-centered puts nodes on the boundary, block-centered does not), and that their local truncation error is of order (0x) 2 . For smooth a and q, the analysis of Gerschgorin [38] yields O((Ox) 2 ) error estimates at the nodes (x , y ). The resuits are based on a maximum principle, and the exact solution is required to have four continuous derivatives. For Dirichiet boundary conditions. with point-centered differences,
; ;
55
stronger theorems needing at most two continuous derivatives have been demonstrated by Bramble, Hubbard, and Thomee for q = 0 [8], by Nitsche and Nitsche [62], by Kellogg [511, and by Bramble [7]. For nonuniform mesh, local truncation errors become interesting. Considering the simple equation p" = q, one can see for either approximation that, by Taylor expansions,
, x; x; xl
Both schemes divide by x ; , 112 x, , 1 , to approximate p"; in the point-centered

case, this is equal to (x ; , , x ; ,)/2 and results in first-order correctness, but in the block-centered case equality does not hold (for example, if blocks i 1 and i + 1 are Jonger than block i, x, ,,, --- x, , 1 , will be less than (x ; , , --- x,. , )/2) and the local approximation is inconsistent. This was first noted in the petroleum engineering literature by Settari and Aziz [86], [87], and there was concern
about the convergence of block-centered differences, though they continued in common use. Settari and Aziz suspected that the procedure was convergent in spite of its inconsistency, and this has recently been proved in one dimension by Weiser and Wheeler [ 108] and by Kreiss et al. [53]. We shall show that the block-centered scheme is equivalent to a certain mixed finite element method (to be discussed later in this section) with a particular quadrature rule, and first-order convergence follows, even for discontinuous coefficient in multiple dimensions. Manteuffel and White [60] have second-order convergence theorems for various point-centered and block-centered schemes on nonuniform meshes, but this is not available in manuscript form at this writing, and we are not sure what schemes, dimensions, coefficient types, and right-hand sides have been covered. This extends early work of Samarskii [84] on second-order convergence for finite difference methods that were locally first-order. Most of our work deals with finite element methods. Here we establish notation, describe standard Galerkin procedures, and recapitulate well known theorems that can be found in texts such as Ciarlet [ 14], Oden and Reddy [63], and Strang and Fix [97]. For simplicity we confine ourselves to tensor-product meshes of rectangles. Choose meshes J,:0 = x O < x, < < x k l and X ..0 = 1 as in the point-centered difference method. Define the < Yo <y, < piecewise-polynomial spacc
4't (J) = {v C'( [0,1
]) : v is a polynomial of degree s ron each subinterval of},
where q = 1 refers to discontinuous functions. Thus, continuous piecewise bilinears on the square would be the tensor-product space ./12(J,) 0 ,/1^^;(,^,.). A
56
basis for this space would be lv i wj : 0 ss i s k, 0 s j ss I}, where v ci.4't (A,) is 1 at x and 0 at all other nodes, and similarly for w^ c ,42o(zj. Define also the Sobolev space H" (S^) = jv: all partial derivatives of v of order s m are in L 2 (9) } and the norms
; ;
1V11 = 11V110 = IIVIIL2(52) = I v dx dy '^',

S2
lvii,,, Ii
= = ((rn
J'
)I/'.
Let (v, w) = f vw dx dy denote the usual L 2 inner product. In a Ritz variational method for (5.1) (5.2), one noten that p is the unique function (of mean value zero) in H'(Q) minimizing the functional (representing energy in structural problems) A(u) = '/2(aVu, Vu) (q, u), chooses a finite-dimensional subspace ./12 of H' (S^), and finds the approximation P that minimizes A in M. It is well known that this is equivalent to the Galerkin method: Find P c M (of mean value zero) such that (5.4) B(P, v) = (aVP, Vv) = (q, v), all vc ./I2, where (aVP, Vv) = f 2 aVP Vv dx. The Galerkin method is obtained by multiplying (5.1) by v and integrating by parts, using Green's theorem and (5.2). For nonselfadjoint equations, such as the convection-diffusion equation (3.lOb), the Ritz variational minimization can no longer be defined, but the Galerkin procedure can be extended. In general, the space of trial functions (where the solution P is sought) may differ from the space of test functions v, though the dimensions must be equal. These spaces may relate to meshes of almost arbitrary character, allowing highly irregular geometries to be treated. To compute with (5.4), choose a basis 0,, , O N for M. Setting v = I = 1, , N, yields N equations for the N degrees of freedom in P. Thus, basis functions /' and /' are coupled by the matrix coefficient (aV ' 1 , V,t' ). For general a, this must be computed by a numerical integration rule or by projecting a into a finite-dimensional piecewise-polynomial space. Further, if we take M = .ifl)(L\) O .M (c\,.) for example, this coefficient is nonzero if tJi, and are supported on any common subrectangle; since each is supported on the four rectangles meeting the node where it is 1, a fixed 1, will be coupled to nine basis functions, including itself, its four immediate neighbors, and its four diagonal neighbors. In three dimensions, the coupling would extend to 27 points instead of the 7 with standard finite differences. Numerical integration and more extensive coupling make standard Galerkin procedures costlier than finite differences, a principal reason they have not made inroads in petroleum reservoir simulation. To justify their use, one must show that they are better equipped to handle certain problems. The usual finite element advantages of more flexible geometry
METHODS FOR CONTI N IJOUS BLOWS
57
and better treatment of boundary conditions mean little in reservoir simulation, since geometric irregularities rarely go beyond tilting of geological layers (easily handled by finite differences), and boundary effects are normally unimportant or easily modeled. One specific application that does suggest finite elements is simulation of hydraulic fracture propagation and subsequent flow; these fractures are created in low-permcability formations to enhance production [ 50], leading to questions of rock mechanica and irregular geornetry. We shall not discuss this subject further here. We see that we shall have to demonstrate lome tangible benefits of finite element methods to offset their cost. First, we recall convergence rates for standard Galerkin procedures. In approximation theory, one shows that a piecewise polynomial P of degree r can approximate a smooth function p in L 2 (9) and H) such that
p PII -= a(h' 'IIrIL ). Iln - P (')(hÌli)ll, i).

where h is the largest linear dimension in the mesh undcrlying P. Then in Galerkin theory, one proves that the approximation selected by the Galerkin method comes within a constant multiple (independent of h and p) of minimizing the error. Putting these together, we sec that for piecewise-bilinear approximations, for example, we can expert second-order convergence of pressures and first-order of Darcy velocities; for piecewise biquadratics, we would have third-order and second-order, respectively. The bilinear results could be matched by finite differences if the theorems of Branible [7] could be extended to nonuniform mesh, Neumann boundary conditions, discontinuous coeflicients, and so on. However, assuming p 11 1 (12), the higher-order convergente of
biquadratics would have to make them more efficient than five-point finite
differences on a sufliciently fine mesh. Even with pressure singularities at wells, local estimates away from Wells could lead to similar conclusions. Thus, one can consider finite element methods with quadratic or higher-degree polynomials. Young [ 1 16] does this in a clever way. On a tensor-product mesh, he uses Lagrange polynomials of degree r, r .-= 1, 2, or 3, as basis functions and integrates the matrix coefficients with an (r 1 )-point Lobatto quadrature rule (exact for polynomials of degree 2 r -- 1). Since Bach Lagrange polynomial van ishes at all Lobatto points except one, the computational work of numerical integration is cut signiticantly. Yet every integral (a^ ,, V) would be computed exactly if a were constant, so by the theory of quadraturc error in Galerkin methods [97J, no orders of convergence are lost. For r 1, Young's method is the sarne as point-centered finite differences (thus proving second-order convergente, even on nonuniform mesh), but he found r = 2 to be more efficient for the reservoir problems tested (miscible displacement among theet), and r === 3 stilt more efficient. Hence it appears that Galerkin methods may indeed be competitive in reservoir simulation.
58
Our work relating to the pressure equation deals with other finite element concepts, removal of singularities and mixed methods. To describe removal of singularities for the pressure equation [19], [46], recall that the Laplace operator in two dimensions has the fundamental solution p(x, y) = 2^ log^^+ y , = 2 ^ log r
satisfying
l1
(5.5)
JAP = b(o.o),
where b. denotes the Dirac measure with mass at (x, y). Thus, if a is constant and q = Q(0,0 (5.1) has a solution
5.6 ( )
= x ^( ^ Y)
27ra
Q log r.
If we consider flow across the circle of radius R, (5.6) yields
-uvz = 2irR
on the circle, which simply verifies that a total flux of Q distributed radially must lead to a Darcy velocity (flux per unit cross-section) of Q/2irR. In general, a is not constant, but its value at a well determines the legding singular behavior of the pressure around the well. For miscible displacement, q will be of the form
(5.7) q = Q1b(xi, Y;),
where w is the number of wells and well i is at (x ; , y ; ). Let r; (x, y) be the distance from (x, y) to (x,, y i ). Then the singular part of the pressure is
(5.8) p'(x, y) =
1 N'
N'
' log r;(x, y) =
p'(x, y),
21r ;-, a
where a, = a(x i , y ; ). We seek a finite element method to obtain an approximation P" of the regular part of the pressure, P"(x, y) = p p'.
(5.9)
As we shall see theoretically and numerically, we can approximate p" much better than we could approximate p.
M ETHODS FOR CONTI N UOUS FLOWS
59
A calculation using (5.1), (5.5), and (5.8) shows that V. (a(x)Vp") = 0 (aVp') V. (aV p )
= Va . V p' + a p' + q Va . VP '
Va.
(5.10)
V(a
--
a,)Cpn
-
=jV ((a
IV. ((a
a;)Vp) - >(a ai)AP

-
ai)Vpt),
x 2;
,
for discontinuous a (permeability) this derivation, and the singular functions p;, should be confined to neighborhoods of the wells where a is continuous. Then p"
satisfies (5.10) and
(5.11)
(aV p")
= (aV p') v,
x 52 .
A variational form of (5.10)(5.l 1) is, by Green's theorem,
(aV p,,, V v)
(5.12)
_ ---
((a --- a ; )Vp , V v)

(aiVp, p)v ds, v H'W ).
;
The right-hand side of (5.12) is well behaved, because a a should cancel the 1 /r singularity of Vp , and Vp t' is zero near a boundary well if the boundary passes straight through. 1f SZ is a square, then Vp c' is zero on an entire boundary edge containing well i. 1f also a is constant near wells, then the right-hand side of (5.10) is smooth, and p" attains whatever degree of smoothness the domgin SZ permits in the theory of elliptic boundary-value problems. For a square, p" H 2 (Sl), so second-order convergence of piecewise-bilinear finite elements (finding P" in M, satisfying (5.12) for all v 12, .It2 = .M() XO M(^,.)) is established. 1f a -- a only vanishes to first order at well i, one can show that p" H , (52) for any c > 0, and that the above order of accuracy is attained if the next singular term of p is subtracted out. We skiall not define fractional-index Sobolev spaces here; see, for example, Adams [ 1 ]. In contrast, finite difrerences cannot mimic this in a straightforward manner. The second derivative in the last momber of (5.10) would have to be approxi;
60
mated by a difference quotient. 1f a is not constant, this may be singular at the well. The variational formulation enabies one to remove the second derivative and handle any singularities via integrale. The usual finite difference technique is to distribute the source term q over a grid block, compute a grid-block pressure that can be associated with an effective radius [68], [69], and use a well model such as (3.19) of Chapter 1 to obtain a pressure surface near the well. We have not attempted to analyze this theoretically, but it seems clear that it should be less accurate than the finite element method with singularities removed. The idea of mixed methods for the pressure equation is to approximate the pressure and velocity simultaneously in a variational method, rather than to obtain velocity by differencing (as in (4.2)) or differentiating (using aV(p' + P")) pressure. It is an elementary consequence of approximation theory that one order of convergence is lost in passing from pressure to velocity by differentiation. Additionally, the coefficient a may be rough (because of variations in viscosity, depending on concentration) or even discontinuous (because of permeability), but velocity is smooth. In the system (3.10), pressure is unimportant in the concentration equation; velocity is what is needed. In a compressible system, pressure would also appear, and in more complicated systems, viscosity and interphase mass transfer would also be functions of pressure. Because of these considerations, it would be desirable to have procedures that approximate botte pressure and velocity to the same order of accuracy. This is what mixed methods do.
We begin by separating (5.1) into two equations representing the two physical
principles that led to it:

(5.13) u = aVp, x 9 (Darcy's law),
(5.14)
7 u = q,
x c 9 (conservation of mass).
Let H(div ; 9) be the set of vector functions v ( (L 2 (Q)) 2 such that 0 v c L 2 (Z) and let (5.15)
V = { v (- H(div ; S2) : v v = 0 on a91.
Let W = L 2 (9), and denote the vector inner product on

(v, w)=v w dx,
S2
in the same way as the scalar inner product on L 2 . To obtain a variational form of (5.13)(5.14), multiply (5.13) by v C- V, divide by a, integrate, and integrate by parts to see that
(5.16) (a'u,v)(p,7v)=0, v V.
Next, multiply (5.14) by w c Wand integrate to obtain

(5.17) (7 u, w) = (q, w), w
C -_
W.
METHODS FOR CONTINIJOUS FLOWS
61
The system (5.16)(5.17) is then approximated by finite elements. For suitable finite-dimensional subspaces V h V and W,, W to be described next, the approximation is {U, P} ; Vh x Wh satisfying
(5.18) (5.19) (a'U,v)---(P,7v)=0, (7 U, w) _ (q, w), v, Vj,,
W ' W,,.
The finite-dimensional subspaces need certain properties in order for the convergence analysis of Brezzi [9], Falk and Osborn [35], and Raviart and Thomas [79] to hold. One of these is that div Vh ( Wh , which makes the somewhat strange nature of the following spaces, developed by Raviart and Thomas, understandable. On a rectangular mesh over the unit square described by I, and ,. as before, let the spaces of index r, r = 0, 1, 2. , be
W h = ^^ r
fr
(^t) O
(. 1 ),
V h
(5.20)
[`4r I ( , .t) O , ^ r I(^)] x [jfl r (,r) .IYtO
Vh={v=(v,,v)
J7 :z',(O,y)
v,(1,y) =0,v, (x, 0)=t' (x, 1) -o} _{v Vh:vv=0on3Sl},
where h measures the largest linear dimension in the mesh. Thus, the pressure approximation is discontinuous at mesh edges, and the x- (t'-) component of the velocity approximation is discontinuous at edges parallel to the .v- (v-) axis. Raviart and Thomas also developed spaces for use on triangular meshes, which we shall not discuss. For smooth right-hand side q, the analysis of [9], [35], [79]
yields the error estimates
(5.21)
lip
PII=O(h, I),
LIII _: O(hr 1),
I
VIL = I
where the norm in Vis defined by
(II vJ1 2 I) v II , t ii t ,
As desired, the errors in pressure and velocity are of the same order. For singular q, we can remove logarithms from the pressure as in the Galerkin case. However, it may be of more interest to remove terms of order Q/27rr from the velocity and obtain more accurate velocity approximations. We describe this idea, first suggested by D. N. Arnold, here. Represent the velocity vector u as the sum of a singular part u' and a regular part u", with
K'
( 5.22)
u' = Q , V (log r 1 2^r;I
62
using the notation from the Galerkin case. Then V u' = q, so that u" satisfies (5.23) Vu"=0, xc9, x:-aSZ.
(5.24) We also have
u"v= u'v,
(5.25) u" _ aV p u', x c SZ, and (5.23) through (5.25) replace the original (5.2), (5.13), and (5.14). As before, multiply (5.25) by v c_ V, divide by a, integrate, and integrate by parts, yielding (5.26) (a--'u", v) -- (p,V v) = _ (a 'u', v), v
--
L.-
V,
and multiply (5.23) by w Wand integrate to obtain (5.27) (7 u", w) = 0, w W. Then (5.24), (5.26), and (5.27) comprise a variational form that can be discretized with the RaviartThomas spaces V, W.. This is formulated in analogy with (5.18) and (5.19), except that U" is no longer sought in the spare Vh (normal component zero), but rather should lie in Vh and have normal component as close as possible to that of u'. This is accomplished by enforcing the condition
(5.28) (u' + U") v(v v) ds = 0, v
L:= V.
lf r = 0, then v v is constant on each boundary interval, so this says that U" v must take the average value of u' v on each interval. We remark that U" can be defined alternatively on 3 by interpolating iu' at appropriate Gauss points on 0 . For this procedure with singular q, Douglas, Ewing, and Wheeler have extended the analysis of [9], [35], [79] for consistency with earlier syntax to prove that (5.29) lip Pil = O(h log l /h), (5.30)
Il u ( u' + U ") 1) = O(h log 1/h).
Numerical results presented in [34] for r = 1 showed first-order convergence for p and u on I and second-order convergence away from singularities, as theory would predict. Without removal of singularities, velocity failed to converge on 9 (otherwise, results were the same). This is expected, because p c H' - (S1) and
u e H e(SZ) for any c > 0. At this stage it is appropriate to examine the linear matrix problem associated
with the approximation of (5.26)(5.27), since it is of a different form from the usual finite difference or Galerkin matrix. For simplicity we consider the case
63
r = 0. With the partitions 0 , and J, from before, let bases for subspaces be
.
^(A-r)
{v :1
ijk 1}.
v(X,,,)=din>>
{w: 1 = i =` k}, .1O (A . ) : JV` 1
wi (X) = 1 if X vj (yn) jrn'
<x <x,
j C 1 __ 1 },
M ?(&)Rw:1 5 j < 1},
W. (y) = 1
if v, , < y
l'1.
Then bases for (Vh ) .,, (Vh ), and Wh are, respectively, ^v`w`1, { and N umber the basis for (Vh ) i such that i changes rapidly, and for (Vh ),. such that j changes rapidly. A basis for V,, is ((Vh ), x 0) (0 x (Vh ), ). The resulting matrix problem has the form M, 0 - N,
(5.31)
R,
M,. N
U' = R,. P RR
N N 0 where
(a 'v` wj ,v w )
( ')i',j = (a w, Uj h v )
(zero u niess j j', i
i'
1),
(zerounlessi = r'jj --- j'I < 1) (zero unless j = j', i' = i or i ---- 1),
(Ni)^"j !j w, wl
a x(v "w)
'
(zero unless r = i', j' = j or j -- 1),
and R, R,., and RR contain the right-Band side of (5.26) (in R, and R,.) together with boundary integrals from (5.28). The numbering system makes M. and M,. tridiagonal, N, bidiagonal. By permuting rows and corresponding columns N,. can be made bidiagonal. Since M, and M,. are positive-definite and easily decomposed into lower- and upper-triangular factors, we can eliminate N and N; in (5.31) and see that (5.32)
AP = (N; M 'N, N,'.M . ' N,.) P

i
= RR 1V M -'R -- 1VVM 'R,..

k k
We solve (5.32), a system of order kl (essentially the same order as for a

continuous piecewise-linear Galerkin method), for Pand then back-substitute for Ur and U'. The matrix A is symmetric and positive-semidefinite, and we have used preconditioned conjugate gradient methods to solve the system. The choice of a good preconditioner is difficult and is a subject of current research. For higher-order spaces there are more diagonals in the matrix, but the basic ideas
are the same. The nodes for the discontinuous bases are placed at Gauss
64
quadrature points and appropriate Gauss rules are used to evaluate the integrals M,,, minimizing the work of generating the matrix. for M In this context it is worthwhile to note an analogy between the mixed method for r = 0 and block-centered finite differences. In both procedures, one obtains a pressure best regarded as constant on each grid block with a node at the center of the block. The finite differente method satisfies conservation of mass on each block, and the mixed-method velocity approximation satisfies the conservation equation 7 - U = q everywhere. To see the latter statement, note that 7 U" is constant on each block, and taking w in (5.27) to be 1 on a given block and zero elsewhere, it follows that V - U" = 0; since div Vh c Wh (V`" and W', defined in , (5.20)), (5.27) with w = 7 U" implies that on each cell 7 U" = 0 pointwise. To strengthen the analogy, we now wiste to show that the block-centered finite differente method ((5.3) with the relationships of Fig. 4 in 4) is the lowest-order mixed method ((5.18)(5.19) with Vh = Vh, Wh = Wh) with special numerical quadrature formulas. As in Fig. 4, partition [0, 1] with O ., : 0 = x, / , <x 31 , < < X k t 1/2 = 1 and /.,:0 = Y I/ 2 <Y3 / 2 < < Y, + 1/2 = 1 and set x i = (x i _ 1/ , + Xi- 1/2, OXi+ 1/2 = Xi rl - Xi, A.yj Xi t 1/2)/ 2, yj = (y1 /2 + Yj+ 1/2)/2, Oxi = Xi+ 1/2 = Yj+ 1/2 - Yj _ 1/2' Yj+ 1/2 = yj + 1 - yj . Multiply (5.3) by Ix i yj , the area of a cell, to obta i n
(5.33)
4yj((Ut)i+-1/2.j - (V) 1/2,j) + xi((U -)i,1, 1/2 - (U.)i.1 1/2) =Oxiyjgij^
where U t and U,. are velocity components defined by

(5.34a)
Pit I j
'
- pi,j -4 i/2
(UY) 1/2.j =
ai+ I/2,j OX
(5.34b)
(Ui,)i,1 1/2 =
Qi1/2
Pi j + 1-P
Yj ,/2
Let U be the unique function in Vh satisfying (5.34), and let P be the piecewise-constant function with cell values P, j . We show that {U, P} c Vh x Wh satisfies (5.18)(5.19) if the integrals are evaluated in a certain way. Note that the lelt-hand side of (5.33) is equal to
A Oxj
(U)
(U )
,
0- U
and the right-hand side is the midpoint-rule integral of q over the cell. Thus, U satisfies
(5.35)
(V - U, w) = (q w)M,M,.,
,
w t_ Wh,
where MX M V denotes midpoint-rule quadrature in both directions. Next, (5.34a)

65
implies that
1 1 - (Ax ; + x i , l )AYj
2 a,.
(Ut ), -
l',(p,, _ P, 1.,)
= 0,
which can be rewritten as
(5.36)
a
(a 'U,v, 112w^ )^ - p,
(v, ^n'i) 0,
where v, , , /2 and w' correspond to v;' (linear basis function) and w; (constant basis function) in the discussion of the mixed method matrix, and T M denotes trapezoidal integration in the x-direction tensored with the midpoint rule ir the y-direction. Similarly,
V -V
5.37)
( a 1 U ., w . v
_ p
(.
wv'
_ .
Combining (5.36) and (5.37), we obtain

(a 'L/, v, ). + (a 'Uk, v,.).1,1, - (P, 7 t') 0, v V.
(5.38)
By (5.35) and (5.38), the block-centered pressure and associated vetocity satisfy the mixed-method equations (5.18) and (5.19), provided that the quadrature rules MM, TM, and M,T. are used as indicated. The boundary conditions also agrec. Because of the artihce a, J, , = a h /2 . ] _ = a,, , /, = 0 in the block-centered method, (5.34a) and (5.34b) yield t% v
0, which is the condition imposed by the spacc V,; in the mixed method. The relationship between block-centered differences and the mixed method with r = 0 is now demonstrated. One can use (5.35), (5.38), error estimates for quadrature rules, and the analysis of mixed methods to obtain (5.21) with r = 0, i.e., if {U, P} is obtained by block-centered ditferences and (5.34), then (5.39) lip - Pl + Iu - U
=- 0(h).
This confirms mathematically the appropriateness of harmonically averaging the
coefficient a, which was motivated physically in the derivation of block-centered differences Erom cell balances. As noted in the discussion of Young's method with r == 1 [116], a similar analogy exists between point-centered difrerences and the bilinear Galerkin procedure with quadrature rule T,T,..
6. Methods for the concentration equation. This section considers approximation schemel for the concentration equation (3.1Ob) or (3.14b). We begin by
cataloguing the standard methods used in the petroleum industry, together with
66
some alternatives that have been discussed in the literature. Then we introduce the procedures that are the subject of our work. Since we deal here with the concentration equation alone, assume that a Darcy velocity u is known, so that (3.1 Ob), (3.11 b), and (3.13) are
(
6.1)
uc
t
q x Q
,
tLJ
(6.2)
(6.3)
(DVc)-vc(uv)=0,
c(x, 0) = co (x), x: 9.
xc 0Q, t c J.
-
The usual industry approaches to the system (6.l)(6.3) entail block-centered finite differences with upstream weighting (see (4.1) through (4.4) ), leaving out the diffusion-dispersion term (because, as we shall see, upstream weighting gives rise to numerical dispersion that exceeds physical dispersion on practical meshes). Time stepping may be either implicit or explicit; because our methods are implicit, we write the implicit scheme:
Nl n n 1
C i.j Wi,j
C i.j
ui+ 1/2,1C 1 t I/'.j 1'i I;' jLi I/ 2 1 +
nel
0t
Axi
nt1 nul
(6.4)
u i.j+ 112C,.j +-1/2
U1 1/2Ci j--I/2
,.nul
n+l
= cij qi.j
where
(6.5) ci t 1/2,j =
nt I Ci^j
if (ui, 1/2.j).0 > 0, if ( u i + 1/2.j ).i < 0,
,n f
C i I,1
and similarly in the y-direction for c l/z If u = (u,., u,.) and are constant, then as noted in (4.4) of Chapter 1, the upstream weighting (6.5) corresponds to a numerical dispersion term of
(6.6) 7.(DnumVC) = 4) 20x + 4) 20y 2
-
Upstream weighting is used in the industry because it suppresses nonphysical oscillations in the finite difference solution. As is well known, central differences and standard Galerkin procedures exhibit oscillations (also known as overshoot) when applied to convection-dominated flows (flows of high Peclet numbersee the discussion in 3), especially at the trailing end of moving fronts. For several views on this topic, see the collection of papers in [49] . Petroleum engineers working in the field do not like to see oscillatory solutions, preferring any smooth solution to even a fairly accurate one with smalt wiggles. The standard response of reservoir model developers to this is to live with numerical dispersion, perhaps
attempting to use grid spacing such that the coefficients in (6.6) correspond
roughly to physical dispersion.
67
This approach is fraught with dangers. To understand the situation better, we shall attempt to represent (6.6) in terms of physical dispersion models. Suppose Ax = Dy = h, and let v = u/o be tule interstitial velocity. Then (6.6) can be written in tensor form as h I l'J
(6.7) Dnum =
If 0 is the angle between v and the x-axis, we see from (3.8) that a physical dispersion model (neglecting difl'usion because D,,,,, is obviously velocitydependent) has the form
cos 2 O
cos sin()
cos 0 sin 0
sin() s i n' f -cos H sin 0 --- cos fl sin 0 cos2 H
,
Dph y . = (viI v 1
(6.8)
For different values of 0, we seek (Y, and (Y, such that Dnun1 -- Dph,.. For general 0 this is not possible because, even with symmetry, there are three equations in two unknowns. However, we do obtain [ 107]
ir
(6.9)
H _, :
(Y^=(Y,=
h
=
0.354 h.
We see considerable differences, depending on whether the velocity is parallel (0 = 0) or diagonal (0 = 7r/4) to the grid. Consider Figs. 3 and 4 of Chapter I. In the diagonal case, significant transverse dispersion (h/2 f) impedes the flow from the injector to the producer; in the parallel case, no such dispersion exists. This, in addition to any preferential flow along grid paths in a five-point difference scheme, makel the discrepancies in Figs. 4 and 5 of Chapter 1 understandable. Similar anomalies have been observed in thermal recovery by steam displacement [16]. We note also the implications for problems with high Peclet numbers, which may be in the tens or hundreds (see 2). If the mesh is uniform and the domain length is L, then N = L/h is the number of grid intervals in one direction. Using a, = h/2, the Peclet number with numerical dispersion is
(6.10) Penun, = L/((Y,) nun , = 2L/h = 2N.
Thus, accurate simulation of a problem with Peclet number in the hundreds, in
three dimensions with several layers, could require 100,000 grid blocks. Adaptive
68
refinement could reduce this greatly, but it seems more fruitful to look for better methods first. One alternative suggested in [101] is two-point upstream weighting, in which (6.5) is replaced by (for uniform mesh)
3/201
(6.11)
'/2 Ci-1,
ur+ 1/2
> 0,
c,, I/2 _
312Ci+1 /2C12,
ui+1/2 <
0-
The analysis leading to (6.6) would make this appear to be a second-order approximation, but in the presence of a sharp front it could be expected to be little better than single-point upstream weighting. Two-point methods have been helpful for certain problems, but as shown in [114] (see Fig. 5 of Chapter I), not for adverse-mobility-ratio displacements. Yanosik and McCracken [114] simulated such problems with a nine-point finite difference scheme in two dimensions, using upstream weighting. Their method is based on the well-known fourth-order approximation of the Laplace operator on a uniform grid,
.-+ +4 I., P^ ^ I.^ P^.^ ^ Pi.j + I ) ,2 [-2opii (P1 -+-+ 6.12
( ) 6h
+
1(p,1_1 + PiI.;+I + Pi+i.j i + P+I,;+1)]5
which one obtains by approximating 2/3 of the Laplacian by the usual five-point operator and the other '/3 by the same operator rotated by 45 0 on a scale of hh.
They did essentially the same thing with the convection term 7 (uc) = 0 (c(k/)cp) of (6 .1), using upstream weighting for the coefficient c multiplying
each of the eight differences between p and its neighbors. Thus, their numerical dispersion is 2/3 of D num from (6.7), plus a rotated term; it turns out to be (6.13) h 4 1 v_,1 + (v + v,.1 + I v.
=
Dnum
12_
v 1
v,,1 I v.,
v,.1
v.Y + v, (
1 v.V
__
v I
4 v, + 1 v
v, I + I va v^,
They modified the scheme suitably for nonuniform grids. For uniform grids, we can do the same decomposition into physical dispersion terms as before, and we find
8 = p, (6.14)
7r
, 0..
:a1
2 , a!
0=
4 , 4 ,...:a l =
7r 37r
h^0.471h,a,=
h^0.236h.
The difference between 0 = 0 and 0 = ir/4 is about one fifth of what it was for five-point differences. Reduced dependence on grid orientation should be expected, and this was observed. Numerical dispersion, however, is as severe as before.
69
Watts and Silliman [107] experimented with a refinement of each rectangular grid block into four triangles, by introducing the two diagonals into Bach rectangle. They used midpoint weighting (central differences) and a physical dispersion model with a, = a,. The triangles allow flow in diagonal directions, and the dispersion is not grid-oriented. Also, the triangles can be numbered such that the nonzero pattern of the system of linear equations is a subset of that for standard five-point differences; this makes the procedure attractive for implementation in existing simulators. For mobility ratio 41, rather coarse mesh (10 x 10), and fairly high physical dispersion (the units are not elear from the paper), grid orientation was very mild. It would be interenting to observe the behavior for lower dispersion, and particularly for the physically realistic case of a, a,. If the method could give sharp, nonoscillatory fronts in such cases, it would be very attractive. We consider next what can be done with fairly standard finite element methods in this area. As in 5, we can multiply (6.1) by a test function v H), then use Green's theorem and (6.2) to obtain the variational form ac
(6.15)
I) at , v
+ (DVcc uc, Vv) __ (cq, v),

.
t J,
c(x, 0) ( 0(x),
x 52.
Then we choose a finite-dimensional subspace .W of H'(2), approximate C O by (by, for example, interpolation or projection in the L' inner product--the latter conserves mass), and replace the time derivative (ac/at )" ' by (say) a backward diference quotient over time step Jt to get an implicit time-stepping
scheme: Find C',C 2 , 44 such that (6.16)
n C "r I , v + (DVCrrt I uCn, I, Cv) = (C'rr;' q " `1 ^t
1'
t'),
all v
..'f'(.
It is instructive to compare the accumulation terms of (6.16) and the finite difference method (6.4). In (6.16), neighboring nodes are coupled by integration over common elements, while they are not coupled in (6.4). On a uniform mesh in two dimensions, it can be shown that the piecewise-bilinear finite element matrix is what would be obtained with finite differences if the additional dispersion term
h23 h4 a aa(. 6 t (c-) + 36 t ax'a y

were added to the left-hand side of (6.1). Thus, finite elements are slightly more
dispersive than finite differences.
The theory of Galerkin methods for parabolic equations shows that, for smooth
70
coefficients (including q) and domais (hence smooth solution c), the error c C is optimal in space (order h'+' in L' for polynomials of degree r) and time (first-order). To be more specific, for w(x, t) defined on x J, we define the norms
wil L'(J;fI(t)) =
w(., t) i H' (s1>dt
-,
WII L'(J;H-'(l)> = ess sup IIw(, t )IIt1"(s >

0t-T
Then the error satisfies [22, 109] max liC C il K

,i
h ^^ ^
il u hl I: (J.H"'M +
Ii a^^
at 1. (J.H c.2
)
(6.17)
+KAt
at` L-(J:L(W)
a2u
where K is independent of h, Jt, and u. A few comments about this error estimate are in order. lf (6.1) is strongly convection-dominated, a heuristic argument of the type leading to (6.6) and using c> a/at u a/ax shows that the one-sided time difference gives rise to a numerical dispersion term with coefficients a, =1 v { Jt/ 2, a, = 0, where v = u/ p. This corresponds to a physical model and hence is not grid-oriented, but it does require small time steps to avoid excessive numerical dispersion. (Note that an explicit method would have the same dispersion with negative sign, hence would be subject to a stability constraint on the time-step size.) Further, the nonsymmetric convection term causes difficulties in the error analysis, as terms arising from it have to be bounded by a combination of diffusive-dispersive terms with small coefficients and accumulation terms with large coefficients, these accumulation terms are ultimately eliminated by Gronwall's lemma, but the constant in (6.17) acquires exponential growth in time. Thus (6.17), white true, is not entirely satisfactory for convection-dominated flow. The oscillations and/or numerical dispersion characteristic of simulation of such problems are testimony to that. Young [ 116] applied his LagrangeLobattoGalerkin procedure (see 5) to the concentration equation as well as to the pressure equation, using a physical dispersion model with a, = a,. Because of the accumulation integral (1/ Ot)(bCn', v) in (6.16), which would not be computed exactly with constant coefficients, one order of accuracy in space could be lost. This is likely more than offset at practical levels of discretization by the efficiencies of the quadrature,
described in 5. The accumulation integral is nonzero only if the same basis function is substituted for C" + ' and v, so lumping is achieved and IMPES time
stepping could be considered. For all polynomial degrees tested, including bilinears which are the same as point-centered five-point finite differences, Young found that oil recoveries from diagonal and parallel grids (as in, for
71
example, Fig. 5 of Chapter 1) converged to the same answer as the grids were refined. This is in contrast to the findings of [ 16] with upstream weighting and shows that grid-orientation sensitivity is due in large measure to the griddependent dispersion of upstream weighting, not merely to the Jack of diagonal flow channels in five-point differences. Botter results were found with biquadratics and bicubics. Small time steps were used, as dictated by Peclet numbers from about 50 to 300, and grids were fine enough to resolve fronts; presumably coarser grids would lead to oscillations. Our work with the concentration equation involves interior penalties with finite elements, a modified rnethod of characteristics with finite differences and finite elements, and a cell-balance finite element scheme that can be likened to the YanosikMcCracken nine-point difference procedure. The interior penalty Galerkin procedure was motivated by early computational work of Douglas and Dupont, described in [ 19], with continuous biquadratic polynomials (i1() x .M. 2 (J,.)) and of Settari et al. [88] with continuously dif erentiable bicubic (ifl (J,) x M ; (.1,.), Hermite bicubic) polynomials for concentration. As shown in Figs. 2 and 3, high concentrations of the displacing solvent "finger" toward the production well in a typical displacement, creating a solution after breakthrough that is dinicult to approximate with piecewise polynomials on a coarse grid. Along a boundary of the quarter five-spot meeting the production well, the solution is qualitatively as pictured in Fig. 5. Along the diagonal from injector to producer, however, the concentration is close to 1. Allowing themselves on the order of 100 nodes for a quarter five-spot (viewing it as a small part of a larger simulation with dozens of wells), Douglas and Dupont observed a cusping behavior of the C" quadratic solution along the boundary for a 5 x 5 grid (Fig. 6). In contrast, the (` cubic solution was smeared out (Fig. 7). Grid-orientation sensitivity of oil recovery was noted in both cases; the parallel grid produced more with cubics and the diagonal grid produced more with quadratics. It appeared that a level of smoothness in some sense between C' and C' would yield a better approximation scheme.
production well F1 ;. 5. Qualitative wncentration projile along edge of diagonal grid
ft rive-spot.
72
production well
Fu,. 6. C'usping of C' quadratic approximation.

.
1 n the calculus of variations, one speaks of adding (positive) penalty terms to a functional in order to bias the minimizing function in some fashion. For the self-adjoint problem (5.1)(5.2), an equivalent variational form exists as noted in 5, and penalty terms could be added to bias the approximate solution from a Ritz method. For the convection-diffusion equation (6.1) through (6.3), the analogy is no longer perfect, but the same type of process can be carried out. The idea is to measure failure to be C' by the size of jumps in normal derivatives across mesh boundaries, and to control these jumps by adding penalty terms containing them to the standard method (6.16). Let E 0 and E, denote, respectively, the sets of interior and boundary edges in the finite element mesh. Fix a unit normal vector v on each edge (for a rectangular mesh, use _, or ,.), taking the outward normal on boundary edges. 1f e (- E, and x e- e, define the jump of a function f at x across e as
)
[J ](x) = lim [J(x ev(x)) f(x + cv(x))].

tio
production well
FIG. 7. Smearing of'C cubrc approximation.
73
For x e E, let [f](x) = f (x). Now take a nonnegative penalty function u, (x, t) (usually constant on edges), and add to the left-hand side of (6.16) the
term
,n+^ (6.18) J,(C ,v) _ EIei U,
[
a anv
'
}
dx.
If (6.1) through (6.3) is regarded as a periodic problem with the no-flow boundary condition representing reflection, as is the case for the live-spot miscible displacement problem, then the boundary edges may also be thought of as interior edges with jumps in aC"''/av multiplied by 2. In this case, add also
(6.19) J(C ,,^ ,v)_ >2e c,
aCn'` v
dx
to (6.16). The edge length ( e 1 appears in (6.18) and (6.19) to scale the resulting matrix terms to the size of those in (6.16); 1 e 1 may be regarded as part of a,. Note that J,(v, v) >- 0 and J a (v, v) >- 0, so the extra positive-semidefinite term does not worsen the behavior of the linear system arising from (6.16). Note also that, if c is smooth (in L'(J;H 2 (52) ), say), then the normal-derivative jumps of c are zero, so that J,(c(t), v) = J d (c(t), v) = 0 for any v .4t. Thus the penalty should not alter the fact that C wilt converge to c as the mesh and time steps are refined; it will change the nature of C for particular meshes. Douglas and Dupont [23] proved that, if M C'(S^) is capable of good approximation (e.g., for .M _ x ./^0 _ r ), .M 1 C'(SZ) consists of quadratic splines, which can still approximate to order h 3 in L 2 (SZ) andh' in N'()), then an error estimate like (6.17) holds with K independent of u,. Further, the theory demonstrates a bound independent of Q, for J, (C" r', C" t ') and J i,(C" }', C" +' ), so that as a, is increased on an edge, the numerical normal-derivative jumps on that edge must decrease. By manipulating a 1 , one can lix the smoothness of C'' anywhere between C (u, = 0) and C' (a, _ + (L). N umerical computations by Douglas and Dupont with moderate values of a, yielded solutions between the cusped and smeared ones, as desired. Additionally, for reasons that are not intuitively clear, grid-orientation sensitivity was practically eliminated. The penalties were added to the original code by R. E. Bank. One possible interpretation relates to dispersion. Young [ 116] theorized that he did not observe grid orientation, even with five-point differences, because unlike previous investigators he used a physical dispersion model. In their original grid-oriented computations without penalties, Douglas and Dupont used a physical diffusion model (d,,, > 0, d l = d, = 0 in (3.8) ). In this model the coef kient does not grow with velocity, so that the Peclet number approaches infinity near wells. Interior penalties introduce mesh-dependent dispersion around a front by forcing the numerical solution toward continuous differentiability when C l polynomials may be unable to match the sharpness of the front.
Perhaps the penalties added enough dispersion near the production well to make
74
j +2
j +1
j - 1
j - 2 i-2
FIG. 8.
i-i
i+l
i+2
Coupling for continuous bilinear interior penalty procedure.
up for the lack of physical dispersion and the coarseness of the grid. Later computations have shown that the grid used by Douglas and Dupont was too coarse to give good approxirnation without numerical dispersion. One must pay a price in computational work for using interior penalties. For simplicity, consider piecewise-bilinear basis functions. Without penalties, (6.16) couples the basis function v (l at (x , y^), 0 at other nodes), represented by the solid dot in Fig. 8, to itself and its eight neighbors represented by circles. Now note that v ^ has nonzero normal-derivative jump on all of the edges marked by double lines. Similar considerations for other nodes show that J, couples v ;^ to all basis functions at nodes marked by X's. Effectively, we now have a 21-point scheme instead of a 9-point one. This has serious implications for the amount of space needed to store the resulting system of linear equations in a computer and for the amount of processor time to generate the system and solve it, especially if the system is large and an iterative linear-solution technique is being used. In three dimensions (with penalties on jumps across faces), the coupling goes from
; ;
27 nodes to 81.
In addition to starting from C spaces and penalizing normal-derivative jumps to approach C', one can consider starting from C ' (discontinuous) spaces and penalizing function jumps to approach C. This requires care in reaching an analogue of (6.16), because the discontinuities force us to integrate by parts one element at a time and keep track of the resulting edge integrals. Let T index the elements, with (f, g) . = T fg dx. If the average of f at x E- e across edge e c E o is
{. }(x) =
'/2 lim [f(X ev(X)) + f(X + ev(X))1,

c10
then [ fg] = [ f ]{g} + {
and (assuming D and u are continuous)
75
(7 (DVC ,n I uC'n ' 1 ), v)
(7 (DVC~n +- I _ UCn + I ), V v )
(6.20)
--
r f:,,
ac, 1
av
3C,,r . l
{v} -^ [v] dx
av
f
f
u0v([C^nuI] {v}+ {CCnh[v])dx.
Noting that [c' ' '1 _ [acn t ' /av] = 0 for the true solution c, we can make the edge terms symmetric by replacing [a(-,n ` I la,l^Vi with [en, l ] {av/v}, without affecting convergente of C to c. We can define (r, and J analogously to a, and J,, obtaining as our procedure
)
C n ,1 _ C n
, v + I (DV C,n r 1
] n1 Iv `
J
uCn 1 Vv) r
(6.21)
aC ^'
DQC
[v]) d^.
av
1
av
+ icti 1} [v]) dx
n; l n, l
u- VOCn
1 {v}
lei
Uo[^
] [v]dx
( C
q ,v),
./fl.
4t could be 41 , (^ Y ) 0 .iW , (O,.) for any r > 1. The scale factor e I ' appears in Jo for the same reason as before. No J term is needed because [ Cn ' ] = 0 on c3 with a reflection boundary condition. This type of method was analyzed theoretically for elliptic problems by Wheeler [ 1 10] and for parabolic problems by Arnold [3]. Its principal advantage is that, because the basis functions on each element are independent of those on other elements, the choice of mesh is very flexible. For example, two elements may intersect in part of an edge rather than a complete edge, as is usually required. This allows local refinement in a tensor-product mesh without having to
extend refining lines across the domgin. For dynamic refinement in travelingfront problems, this could be useful. Also, the accumulation matrix in (6.21) is block diagonal, hence is easily invertible; since it is the only term with coefficient 1 /At, it may be a good approximation of the whole matrix and may serve as an efficient preconditioner for conjugate-gradient solution of the linear equations. The thrust of the theoretical results was that, even on such an irregular mesh, there is a minimum value cr such that for a 0 => aT an error estimate like (6.17) holds.
76
FlG. 9. Coupling for discontinuous bilinear interlor penalty procedure.
Like its continuous counterpart, this penalty procedure can involve considerable extra computational work. With piecewise-bilinear functions, each element has Tour basis functions; in the continuous case, only one per node was required. With higher-order polynomials this growth is moderated (e.g., with cubics, from 9 to 16 per node), but in three dimensions it becomes worse. The extra edge integrals demand computer time and programming effort. For bilinears, the function marked by the solid circle in Fig. 9 is coupled to the 16 indicated functions, as opposed to 9 in the continuous case. By adding J, to (6.21), it is possible to combine the two penalty methods. This would allow a discontinuous underlying space, with large i and moderate a, leading to solutions like those obtained by Douglas and Dupont. We next describe the modified method of characteristics, which is applied to the nondivergence form (3.14b) of the concentration equation. As in (6.1), assume that u is known. The idea of the method is to time-step along the characteristics of the hyperbolic problem obtained by ignoring diffusion and dispersion in (3.14b). This treats the convection part of the problem, which is the difficult part numerically; then finite differences or finite elements are used for the diffusion-dispersion part, for which they are well-suited. The time-stepping direction is given by a unit vector T(x) defined by
(6.22) + + 1 u(x) 2 _ ^--at u X + u v ay ax aT
a a
Using (6.22), we can rewrite (3.14b) in the form

77
( 6.23 )
2 + IUf 2 a c _v. (DVc) (^ c.)q. ( c) ^

aT(x>
Then (6.23), which has the form of a diffusion-dispersion equation, is approxi mated by backward-differenced finite differences or finite elements.
The crucial aspect of the method, both theoretically and computationally, is the approximation of the "time" derivative 3c/8r. Many methods based on characteristics fix a point x ( SZ at time level n and ask where it will go at time level n + 1; these moving point or front tracking methods must then solve at time level n + 1 on a mesh of irregular or unpredictable character. In two dimensions this is difcult, and in three it is still much more difficult. Our method takes the opposite view, fixing a point at level n 4- 1 and asking where it came from at level n. Thus the solution mesh at level n 1 is controlled by the method, not the flow; it can be fixed for all time or adjusted to changing flow patterns in an uncomplicated way. Two- and three-dimensional flows present no difficulty in this framework. Our numerical work so far has used fixed mesh. Accordingly, consider (x, t ; SZ x J. 1f the velocity operating on the Huid particle between t" and t' can be well approximated by u(x, t" ), let u * _ u(x, t";') and define
(6.24)
z = x
u*
0(x)
Vit.
This approximates the characteristic through (x, t" +') by its tangent at (x, t ' 1) For certain applications, with rapidly varying velocity, u" will have to be chosen more carefully. For a function f, set f(x) = f(x). Then form the backwarddifference approximation
aC' n I
en i '(x) cn
(-X )
j'
^t 1 + 0 which is equivalent to
(6.25)
* 2 8e"' I c"+
1(x)
_C,n(x)
0^ + u
(x)
Lt
Thus, a finite difference approximation of (6.23) can take the form

c 1. 1
(6.26)

'
0t
{differente formula for v (DVc) +' ^=:_(C C) 'q 1 ,
where the diffusive-dispersive differente formula is necessarily a nine-point scheme if D is a full-tensor physical dispersion model. Analogously, a finite element scheme is
ntl cn
l
(6.27)
î
dt
v + (DV C
+1,
vv) = ((C C)n l 1 ^n ^ 1, v),
v - .^Yt..
78
_Note that in either method, the convection term moves to the right-hand side in C", so that the system of linear equations is symmetric and positive-definite. This makes the method ideal for iterative linear-solution algorithms in large problems. In (6.26), ; = C"((x , y )) does not, in general, lie at a grid point, so some type of interpolation is necessary to evaluate it. The finite element ca s e requires no interpolation since the elements already do it, but the integral (4C", v) must be evaluate d by a suitable quadrature rule. The choice of this rule is not obvious, because C" is not the type of function considered by the usual theory; it will have corners at points interior to elements that are mapped back to element edges by approximate characteristics. Because of this, we have used Lobatto rules in our numerical work, desiring information from every element in the characteristic image of a given element. In either method, approximate characteristics may cross aQ; if they do, the no-flow boundary condition can be used as a reflection to continue C. With the above notation, we can see that this method fails into the classof fractional-step procedures, with the method of characteristics taking C" to C", then finite differences or finite elements taking ^^n to C" +' . The characteristic step does not conserve mass, and we shall say more about this in 7. Characteristics are not new to the petroleum literature; Garder, Peaceman, and Pozzi [37] developed a moving-point method for miscible displacement almost 20 years ago, but as we noted above, our procedure is quite different. For a single equation like (6.1) with smooth data, Douglas and Russell [27] analyzed this procedure theoretically. For finite differences on nonuniform grids with linear interpolation for C", they found the expected error of order h + \t at the grid points. For uniform grids, they were able to replace h by h 2 if a form of quadratic interpolation was used. For the finite element method, they obtained an optimal estimate like (6.17). For botte methods, the usual 1 a2elat2 ((Ot temporal error estimate was replaced by 1( a 2 c/aT 2 11 At, which should be much smaller because the truc solution changes much less rapidly along characteristics that follow the flow than it does at a fixed point in space. The Galerkin proof did not need the techniques for treating nonsymmetric terms that inflate the constant in (6.17). Thus, there is reason to expect much larger time steps to be feasible and accurate with this method than with standard schemes. It was noted that the theory extended to nonlinear problems. Ewing and Russell [30] extended the finite element theory of Douglas and Russell [37] to a three-level second-order backward-difference method, obtaining the same optimal spatial error together with the expected 1 error in time. Their techniques could be extended further to higher-order backward-difference methods. High-order time stepping methods do not make sense for traveling-front problems in conjunction with standard procedures, because the time derivatives of the solution are large. However, by time stepping with the flow, we can make high-order methods worthwhile. Pironneau [77] analyzed essentially the same single-step finite element scheme for the NavierStokes equations; his proof obtained one order of
; ;
79
accuracy less in space than the proof in [27], but he allowed the analogue of diffusion and dispersion to go down to zero. Numerical results for single parabolic equations in one dimension have been reported by Ewing and Russell [30] and Russell [82] . Neuman [611 used the same ideas in numerical computations for contaminant transport. A constantcoeficient convection-diffusion equation was approximated by single-step and multistep methods [30], and the nonlinear Burgers' equation was treated by single-step schemes [82]. In each case an exact solution was available to allow comparison of errors for various methods. lf the spatial mesh was too coarse to resolve a front, characteristics did not improve the solution much. Once the grid was fine enough (perhaps three intervals across a front), characteristics enabled
much larger time steps to be used with no loss of accuracy. In some cases, time steps could be two orders of magnitude larger than in standard methods. The answers with characteristics were also remarkably free of oscillations and numerical dispersion. The multistep method improved over the single-step scheme when the mesh was very fine, so as to practically eliminate spatial error. With Burgers' equation, no accuracy was lost if the nonlinearity was lagged along characteristics, avoiding a systein of nonlinear equations at each time step; this gives hope that the method may be able to handle more complicated nonlinearities in the same way. The straight characteristics of these example equations allowed the method to conserve mass; we shall say more about its nonconservativeness for miscible displacement in 7. The cell-balance finite element procedure was introduced by Potempa [78]. It is a hybrid of ideas from block-centered finite differences and piecewise-bilinear finite elements. In 5, we saw that block-centered differences follow from conservation of mass on each grid block in a domain. Now consider the variational form (6.15) of the concentration equation, without Green's theorem: ( 6.28)
ac v 7 DV c -- u c v
)
Actually, (6.28) holds for v ( L 2 (52), and if we choose v to be the characteristic function x ; , of grid block [x ; ,, x ; ] x [_y 1 ,, y^] == R ;1 , (6.28) becomes precisely the statement of mass conservation for R. In the finite element method, we do not use x as test function, but rather (for example) a continuous piecewise bilinear function v,1 that is 1 at (x ; , y 1 ) and zero at all other nodes. Since v ; is supported on SZ - R ;1 R 1 + 1 R ; , , ; R ; , I ^ , ,, we can regard (6.28) with v = v i, as a weighted mass balance on SZ ;I . Potempa considers the region in R 3 under the graph of v ;j and views (6.28) with v = v ;1 as a mass balance on that region. We shall think instead of a mass balance on 52,1 "^ R 2 with measure v.(x, y)dx dy. Potempa's method mimics block-centered differences, but uses the regions Q ;^ with weighted measures rather than the standard grid blocks; it is written as a finite difTerence scheme. To mimic block-centered differences, we must define the mass in a region Sl;^
80
and the rate of flow (using a Darcy velocity) from one weighted region to another. For incompressibie miscible displacement, using volume instead of mass, the volume of the solvent component in g ij is
(6.29) y'''
Xo
f` 4cv ;j dx dy.
For the inter-region flow let X; and Yj be the piecewise-linear functions such that vtj (x, y) = X ; (x)Yj (y). Consider an area element dx dy in R 1 ,j+1 , and analyze flow in the x-direction from 9 i, to Sl ;+ ,, j and Z,1+1 (no x-direction flow from to SZ ;, j+ I ) under velocity u x . The interstitial velocity is u r /4, and in time dx/(u X /4) this area element will lose its mass to a like element to its right. This mass (volume) of all components is 4 dx dy, and its weight in Sl ; j will change by av j /ax dx. The proportion Yj of this flow will go into 5^ ;+ ,, j , and the remaining proportion 1 Yj = Yj+ I will go into 5^ ;+ ,, j ,. . The flow rate from St ; j to 5Z ;+J due to the element dx dy is
=
(k dx dy) ((ay ;j /ax)dx)Yj ay;..
u X Yj dx dy. dx/(u x /<p) ax
Thus, the total flow rate is

(6.30)
Similarly,
)
r I^j =
R ^.^ U R,. ^.; ^
av
uX
ax
Yj dx dy.
6.31

rÎ.j+I _ ij
u avij Y+I + u y x ax j
x^ a +I
dx dY.
Rates of flow to the other six neighbors of (x i , yj ) are obtained by symmetry from (6.30) or (6.31). Set I to zero. These quantities, multiplied by a time step Ot, are analogous to VR and VL of the derivation of block-centered differences in 5. Potempa then sets up an upstream-weighted finite difference scheme along the lines of (6.4), of the form (mass balance for solvent component only)
(6.32)
4> v 1 dx dy
n+i n C,,, Cnj ^t +I +I

i+k,j+! ,-, n+1
Ci+k/2,jt//2 +
~n+ I n+ v;jC q 1 dx dy
k--I !3-I
where
if r ,^k,j,' > 0,
(6.33) Ci+k/2,j+1/2 =+.c.j+^
C,+k,j+/
if r
< 0.
Because the accumulation term is lumped, this scheme can be made explicit also.
81
From the way the method was derived, it is clear that it conserves mass. Additionally, Bell, Shubin, and Wheeler [51 have shown that the implicit version satisfies a maximum principle. Thus, it will have no problem with overshoot, and likely will not experience any nonphysical oscillations. In fact, if u, and u,. are constant the procedure reduces to the nine-point scheme of Yanosik and McCracken [ 1 141 described Barlier. The reduced grid-orientation sensitivity and considerable numerical dispersion of that procedure are shared by Potempa's method; the legding part of its numerical dispersion term is (6.13). Potempa's philosophy in developing this procedure was to reduce grid orientation within a framework easily implemented in existing petroleum industry simulators. The method does this, because it differs Erom that of Yanosik and McCracken only in its use of vetocity integrals, rather than pressure differences, to determine flow coefTicients. These integrals are easily calculated by inexpensive quadrature rules; for example, u can be evaluated at the midpoint of a cell and the rest of the integral evaluated exactly. It is interesting that Potempa's flow concept extends to nonrectangular meshes, triangles, for example. This has enabled computations to be done for the leven-spot miscible displacement shown
Jo 0
F iG. 10. Finite element eelt-balance procedure for the inverted seven-spot pattern.
82
in Fig. 10. For highly nonuniform meshes, resuits have not been as satisfactory as for uniform grids, probably because of the nature of (6.13) in that case; thus, the method is presently not suitable for accurate simulation of displacements with high Peclet number because of its numerical dispersion (see (6.10)). It should be possible to modify it for better treatment of nonuniform meshes. The implicit version of the method has been proved to converge for smooth coefficients to the solution of (6.1) by Bell, Shubin, and Wheeler [5], provided that a physical diffusion term is incorporated in (6.32). The error estimate is max 11 C" Cn II = o(h + Ot), where C" is bilinearly interpolated to a function. It follows from this work that the method of Yanosik and McCracken, and any other method with a similar type of mesh-dependent diffusion and a maximum principle, is also convergent. 7. Theoretical and numerical resuits for the miscible displacement system. In this section we shall describe results, both theoretical and computational, for various combinations of the techniques defined in 5 and 6 for the miscible displacement problem. We begin by specifying test problems that have been the focus of most of the experimental work. Figure 3 of Chapter 1 displays a five-spot pattern of injection and production wells. Assume that the reservoir is horizontal, of thickness 1 foot (otherwise, divide the three-dimensional quantities by the thickness to obtain a twodimensional model), and initially saturated with oil (c 0 (x) = 0 in (3.13)). Let
the distantie between an injector and the nearest producers be 1000 fi feet, so
that the diagonal grid is a 1000-foot square. Impose a constant flow rate of 200 ft 3 /day on each well, a uniform porosity of 0.10 on the reservoir, and a viscosity of 1 centipoise on the oil. Each injector governs a 2000-ft-square five-spot of pore volume (0.1 0)(2000)2( 1) = 400,000 ft 3 , so that a pore volume is injected in 2000 days. The viscosity of the Huid mixture is assumed to obey the quarter-power law
0 011
^',oIvent
1/4 = 1/4 + 1/4 ' U lood solvent
which can be written in the form

(c) = [( 1 c) + M'" 4 e] ^1
where e = csoivent and M is the mobility ratio all /, o , Vent . The mobility ratio, diffusion-dispersion coefficients d e,, d,, and d,, and permeability k wilt vary from experiment to experiment. If d, = 0, then the longitudinal and transverse Peclet numbers are 1000/d, and 1000/d 1 , respectively. If d, = d, = 0, there is a single isotropic Peclet number that varies areally, approaching infinity at wells, zero at the nonweil corners of the diagonal grid, and 56/d4 at the center. We shall follow Young [ 116] and refer to the central Peclet number in such cases.
For M> 1, the more mobile displacing solvent may make multiple unstable channels into the oil, in the phenomenon of viscous fingering. The greater the
83
value of M, the more unstable the displacement front. Even if the front is stable, greater values of M cause the solvent to preferentially move toward the production well, reducing oil recovery; see, for example, [28]. As noted in 2, dispersion and heterogeneity are also controlling factors here. It is fairly clear heuristically, and presumably could be made rigorous, that the mathematical solution of a homogeneous problem should not have multiple fingers; the point on the front directly between the injector and the producer is closest to the producer, so should move the fastest, and other points should move more slowly as one traces away from the fastest point. Nonzero dispersion should make the problem well-posed in the lense of continuous dependence on data, for arbitrarily high mobility ratio. Thus, we take the view that any multiple fingers observed in simulation of homogeneous problems are due to numerical instabilities and not to the modeling of physics. We shall be seeking single, stable fingers in homogeneous problems and multiple fingers in heterogeneous problems with adverse mobility ratio; those fingers will be duc to modeling of physics. In addition to viscous fingering, our numerical studies will be concerned with grid-orientation sensitivity, numerical dispersion, overshoot and undershoot (or lack thereof as dictated by maximum principles), and material-balance errors (or conservation of mass). Traditionally, grid orientation has been measured by differences in recovery curves (e.g., Fig. 5 of Chapter 1) for diagonal and parallel grids. We feel that this is not adequate, because heterogeneities may make the location of concentration level curves in the reservoir (e.g., Fig. 4 of Chapter 1) fust as important as outlet concentration at a production well. Hence, we present areal concentration maps along with recovery curves. Numerical dispersion is evaluated rather subjectively by comparing different methods for the same
problem and by varying grid spacing and time-step size. Overshoot (undershoot)
is the greatest amount by which concentration exceeds 1 (is less than 0) at some suitably chosen time. Schemes based on the divergence form (3.10b) of the concentration equation should conserve mass; taking v = 1 in (6.16) yields (7.1)
f c"
fr'dx
dx
fCdx + Let
2
(`n + 1 q n . 1
`7
dx.
For the nondivergence form (3.14b), one obtains (7.2)
(k frdx -}- (1 t
f it
[(C,
C)n r 1
( ti
V Cn + 1 ] dx.
`7
If V u = q and u - v = 0, as is the case if u is obtained from a mixed method, then
f
(7.3)
u.
VCn1Idx
_f
r_
V. (uCn+l) +
cn I(u -
Cn
+l(7 . u)] dx
S1 cnt I q n+ I
_ _ i1
v)ds
dx
fC
, 1 qn ,
dx,
84
and (7.2) reduces to (7.1). If u is obtained otherwise, then the nondivergence form does not conserve mass, and material-balance errors will be used as a criterion of acceptability of numerical solutions. The modified method of characteristics defined in (6.27) is nonconservative even if u comes from a mixed method, because the convection term in (7.2) is handled by approximations to characteristics instead of a variational integral. Overshoot and undershoot have appeared in the early stages of displacements, when the front is nearly radial around the injection well (assuming uniform, isotropic permeability) and very steep (see Fig. 1). Methods without excessive numerical dispersion have difficulty representing such fronts in reasonable cartesian grids. If M is large, these numerical errors may lead to nonphysical viscous fingers. To alleviate these problems, many of our experiments have assumed radial flow in the early stages, solving a one-dimensional equation cheaply until the front is far enough away from the injector to be representable in a two-dimensional mesh. We derive here the equation that has been used. With incompressible radial flow, the Darcy velocity is u(r) = q/21rr, so that no pressure equation is needed. There is no transverse dispersion, so we have D = 4d m + d,u = 1dm + d lq/2irr. Consider an annulus A around the injection well of inner radius r o and outer radius r,. An equation balancing accumulation, convection, diffusion, and dispersion of solvent on A (no sources) is
acdx =A qc( o) 2 r r qc( 1)
at
rDa (o) + 2irr a (r ) c c ,D ar 1 o ar r
(7.4)
=
faA ar
i ic p ds.
)
ac
Using the polar-coordinate representation V f = (1 / r) (0/ar) (rf(r)) of the divergence operator, we can use the divergence theorem in (7.4), rewrite in polar coordinaten, and drop integrations to obtain
(7.5) 2i ^
a c r ac a 1rrDarc \= 0, at ar
q
r > r welf
t > 0.
The boundary condition at the injection well is an influx of cq, which is represented by
(7.6) 21rrwe11
ac a r (rwe11) gc(rwell) =
q,
and we can require c = 0 at some distance R far from the well. The initial
condition is complete saturation with oil. We can handle the point-source case by letting r We --i 0, which yields ( 7.7a)
a 21rr d + d } ac 2 ^rr a ^ m i9 r ^ at ar ^
^
--
c = 0,
r > 0, t>0,
METHODS FOR CONTINUOUS FI.OWS
85
(7.7b) (7.7c)
c- d 1 c, ar c=0,
r==0, t>0,
r= R, t>0,
(7.7d) c=0, r>0, t = 0.
It is a simple matter to solve (7.7) by finite elements, using the Neumann condition (7.7b) when integrating by parts. Note that, because of dispersion at the well, the flowing concentration c and the resident concentration c may differ; this can happen at production wells also, but our experiments have neglected it so far. The theoretical results for finite elements will refer to spaces of piecewise polynomials of degree r and s for the concentration and pressure approximations, respectively. These will be defined on meshes of maximum diameter h,, and h. The proofs of the theorems in the cited references simplify to a single parameter h; it is routine to specialize the arguments to obtain estimates like O(h^ +' + h;,'' ) from D(h' + h''' ), and we make no further note of this here. As noted at the end of 4, it may be worthwhile to use different time steps ztp > At a.; this will be mentioned where appropriate. All of the theorems, unless noted otherwise, assume that the source and sink term q is smoothly distributed rather than a combination of Dirac measures. This lays a firm foundation for the methods and should demonstrate, as indicated by the numerical results of [34], the expected rates of convergence away from wells. Around wells, these rates wilt not be observed in practice. The remeinder of this section is organized by concentration method. Within each concentration method, we consider various pressure methods and the theoretical and numerical results associated with each combination. The concentration methods, with pressure methods in parentheses, are five-point finite differences (five-point finite differences), the finite element cell-balance scheme of Potempa (standard Galerkin), a standard Galerkin method (standard Galerkin, mixed), continuous interior penalties (standard Galerkin), discontinuous interior penalties (standard Galerkin, mixed), and the modified method of characteristics with finite elements (standard Galerkin, mixed) and nine-point finite differences (five-point finite differences, mixed). The combined time stepping is in the manner of (4.6) unless otherwise indicated. The five-point difference scheme, with upstream weighting in the concentration equation, is a combination of (5.3) (block-centered grid) and (6.4). This would be the sequential method used in the petroleum industry and is included Tere for completeness. We know of no theoretical convergente results for the coupled system; indeed, there is evidente that, at least at practical levels of discretization, it fails to converge to the same answer for different grid orientations [ 16]. For constant permeability, M = 100, and d, = d, = d, = 0 (numerical dispersion only), Fig. 1 1 displays diagonal and parallel concentration maps and
86
........
...
.:;
'f
Concentration at .5 pore volume, diagonal grid, 20 x 20 mesh.

O 0
Concentration at .5 pore volume, parallel grid, 28 x 28 mesh.
Ut
n ^
O W W O U W W r
0
G
.00 .B
.75 i .
.S0
PORE VOLUMES INTECTED

- 20 x 20 diagonal grid _........_ 28 x 28 parallel grid
Recovery curve. FIG. 11.
Five-point upstream-weighting finite differente procedure for the five-spot pattern witti
M= 100.
recovery curves. The fronts are considerably smeared, and the sensitivity to grid orientation is clear. The method does conserve mass, and because of the upstream weighting it should be easy to show that it satisfies a maximum principle, so that overshoot and undershoot do not occur. Analogous results are shown in Fig. 12 for the finite element cell-balante
METFIODS FOR CONTINUOUS FLOWS
87
Concentration at .5 pore volume, diagonal grid, 20 x 20 mesh.

0 0
ui f-,
V
W D> O U
[.i
win
0 C)
^8 .75 L . 1S 1.50
POR E VOLUMES 2NTECTED E
20 x 20 diagonal grid
28 x 28 parallel grid
Recovery curve. F K. 12. Finite element cell-balance ( Potempa) for the Jive-spot nastern with M 100.
procedure (6.32) with Galerkin pressure (5.4). This shares the virtues of the difference method and, as expected from 6, reduces grid orientation because its nine-point dispersion from upstream weighting is nearly rotationally invariant (see (6.14)). Its Peclet number, like that of the difference method, is estimated by (6.10). Fig. 13 shows an inverted seven-spot pattern with no-flow boundaries, a problem that permits one to study grid orientation in a single plot. Figures 14
88
%O
01
FIG. 13. inverted seven-spot pattern.
and 15 were generated by the difference and cell-balance methods, respectively, and the advantage of the diagonal flow allowed in the cell-balance scheme is evident. Theoretical analysis of the cel1-balance scheme by Belt, Shubin, and Wheeler [5] was noted in 6. One can easily write an analogue of the Galerkin procedure (6.16) for the nondivergence form (3.14b) of the concentration equation. This, coupled to (5.4),
FIG. 14. Concentration at .5 pore volume (five-point upstream-weighting finite difference procedure for the inverted seven-spot pattern, 19 x 22 grid, M = 100).
METZ-IODS FOR CONTI N UOUS FLOWS
89
.... .......
+.
i r i
rom ' ' ^
, '.`
FIG. 15. Cvneentra!ion al .5 gore volume (finite element cel!-balance procedure for the inverted seven-spot pattern, 19 x 22 grid, M = 100).
was analyzed by Ewing and Wheeler [32]. For a continuous-time version, in which the time derivative in (6.15) is not replaced by a difference quotient, they proved that
C - CII1.'(J:l(^!1) + hIIO(P
P)III_'(J:1.'(il)) + hII 7 ( (, ^^)II L'(J:l_'(11)}
(7.8)
O(hrt1 + h''' + h''` ')

O(h`' 1 + h`)
if d,=d,=0, r + s 3, if d, Oor d,
0, s >_ 2.
Convergence for piecewise-linear polynomials for both pressure and concentration was not demonstrated. They also analyzed discrete-time schemel, obtaining estimates like (7.8) with additional error terms of ,!t for backward differencing and (At ) 2 for CrankN icolson. Analyses of efficient time-stepping algorithms for the standard Galerkin approach can be found in Ewing and Russell [31]. In particular, they investigated the use of different time steps for pressure and concentration, with ;^t p an integer multiple of At,. and velocity extrapolated as in (4.6). For backward differencing, they found time-truncation errors of O(^t i . + (At,,) 2 ) in addition to the spatial errors of (7.8), so that one should choose At n = O((^t.) 1 / 2 ). The use of preconditioned conjugate gradient iteration for the systems of linear equations, iterating only long enough to stabilize the time-stepping procedure, was considered in this paper; the idea had been discussed for a single nonlinear parabolic equation in [24]. In certain cases, Ewing and Russell showed that a fixed number of iterations per time step, independent of h and Ot, was sufficient for stability and optimal convergence. Their arguments also lead to an improved estimate in
90
(7.8); the term h' can be replaced by h 2 1 log h 1, if d, = d, = 0 and r = s = 1. This convergence of linear polynomials for both unknowns has not been proved
for the more physical case of tensor dispersion.

Numerical resuits with standard Galerkin methods were cited in 6 as motivation for the interior-penalty procedures. Cusping (a form of overshoot and undershoot) and grid orientation made the results unsatisfactory. These results used Dirac measures as source terms at wells and subtracted out pressure singularities as in (5.12). Ewing and Wheeler [33] were able to analyze this theoretically for the restricted case of M = 1 (unit mobility ratio), in which viscosity is independent of concentration, pressure does not change with time, and (3.14b) becomes a linear parabolic equation with singular source. Their resuits were, for continuous time and any e> 0,
u uit =
(7.9)
l ic Cii L(J:L'(s2)) =
O(h' E)
! O(h^ 2
if d, = d, = 0, ) ifd, Oord, 0.
These theorems depend on regularity results of Sammon [85], who proved that for any c> 0 and r<oo,
IIe i L (J:N' `(SZ)) + llc i i- (J;n" (12 + IICII L'(Jli' `(W) + IIP III.'(J.H '(W)
2 I l
(7.10) ac
2
+ II uIIL'(J.L '(SI))+
The nondivergence analogue of (6.16) coupled to the mixed method (5.18)(5.19) was analyzed by Douglas, Ewing, and Wheeler [25]. With smooth source and continuous time, they derived optimal rates of convergence for concentration, pressure, and velocity:
l ic C L-(JL2()) = O(h si
(7.11)
r+ l
h s ol ),
c+ Ii u - X 11 L'(J:V) + 11 P - P11 L'(J.W) = O(h t I + h ! ),
where V and W are as in 5 and the spaces Wh and Vh of (5.20) are used. For singular source and M = 1, they obtained 11 u Uil = O(h 1 log h 1), together with concentration estimates as in (7.9). In another paper [26], they considered an efficient time-stepping procedure for this combination of methods. This procedure takes &t p to be an integer multiple of At a., but instead of extrapolating velocity as in (4.6) it extrapolates to the middle of the current pressure time step. Compensating terms involving the difference between the two extrapolations are added to the right-hand side, with concentration replaced by an extrapolation. The result is that over a pressure time step, the concentration matrix remains constant; it can be factored once by a direct method and used several times. This efficient direct method is an alternative to the iterative methods analyzed by
91
Ewing and Russell [311. Time-truncation errors of O(t,. + (11 t) 2 ) were obtained in addition to the spatial errors of (7.11). No numerical experiments have been conducted with this particular combination. The continuous interior-penalty Galerkin procedure (nondivergence analogue of (6.16), with (6.18) and (6.19) added) for concentration was analyzed by Wheeler and Darlow [ 1 1 1 ] in combination with a standard Galerkin method (5.4) for pressure. Continuous-time and discrete-time cases were considered, in particular, the use of different time steps for pressure and concentration. The analysis, which combines ideas of Douglas and Dupont [23] and of Ewing and Wheeler [32], involves a restriction on the time rate of change of o, in (6.18) and (6.19); it suffices to assume that there exist constants a* > 0 and K such that
h sup J `_ a, (x' t) at (x ' t) : 0
Q*
(7.12)
+ sup
cs, (x, t) - (x, 1)
aQ , at
1 : u, (x, t)
>_
Q* 5 K.
An estimate of the form of (7.8) for d, = d, = 0 was demonstrated, with appropriate time-truncation errors for the discrete-time schemes. Numerical results with subtraction of singularities were cited in 6; the behavior of standard Galerkin methods was improved upon significantly. The discontinuous interior-penalty method (6.21), analyzed for a single parabolic equation by Arnold [3] as noted in 6, has not been analyzed as part of a coupled system. Numerical results with the standard Galerkin method (5.4) for pressure (with singularities removed) have been reported by Darlow [17] and by
Douglas, Wheeler, Darlow, and Kendail [28]. In the Jatter paper, mobility ratios
of 1, 10, and 41 were used, with cd = 1 ft`/day, d, = d, = 0 (Peclet number 56), and various permeability patterns. In accordance with theory, it was found that u o (actually Q od m ) had to exceed a minimum value to stabilize the scheme; values of the order of 100 ft 2 /day were found to make the approximate solution nearly continuous. If a, was also made large, the piecewise-bilinear solution became nearly C'; thus, any edge with large a, was effectively removed from the mesh, and self-adaptivity concepts could be tested in a comparatively simple program with a fixed mesh. On edges to be kept in the grid, best results were obtained with u, = 0.5 ft 2 /day. As one would expect, solutions were insensitive to removal of edges in parts of the grid away from the front. The study used the nondivergence form of the concentration equation, concentration time steps of 8 days (0.004 pore volumes injected) with (effectively) an underlying uniform 20 x 20 grid, and radial initialization (7.7) to 0.1 pore volumes injected. The efficient time-stepping scheme analyzed in Douglas, Ewing, and Wheeler [26] was also used. For uniform permeability, material-balance errors were of the order of 1 %, an acceptable level, and grid orientation (as measured by recovery curves) was very slight. The penalty Q, introduces some numerical dispersion, not clearly
92
understood. Overshoot and undershoot were significant (approximately 9%), even for this rather modest Peclet number. Runs were made with nonuniform permeability, and satisfactory results were obtained with variations up to 10 to 1 (1000 millidarcies in half of the domgin, 100 in the other half). However, the method did not perform well for variations of 100 to 1, and convergence problems were encountered with a mobility ratio of 100 [1 7J. Some of these difficulties were alleviated by using the mixed method (5.18)(5.19), with velocity singularities removed as in (5.26) through (5.28), in place of the standard Galerkin method for pressure. These calculations by Darlow, Ewing, and Wheeler [18] used the conservative divergence form and bilinear polynomials in (6.21), with r = 1 in (5.20). Figure 16 presents diagonal and parallel concentration maps and recovery curves for M = 100, 4d m = 1 ft 2 /day, d, = d, = 0 (Peclet number 56), and uniform permeability of 100 md. Radial initialization was used to 0.1 pore volumes injected. Material-balance errors of about 0.5% arose from inexact solution of the linear equations by preconditioned conjugate gradient iteration. Grid orientation is clearly absent, but numerical dispersion and overshoot/undershoot are as serious as for the Galerkin pressure. The gains from the mixed method were the ability to handle higher mobility ratio, as shown here, and greater variations in permeability; the latter is to be expected, since one no longer has to multiply by the rough permeability coefficient to obtain velocity. The overshoots in these methods are not instabilities, because they do not grow with time. They are reflections of the inability of a particular grid to resolve a moving front without additional numerical dispersion. The modified method of characteristics (6.27) reduces these difficulties by, in effect, stopping the front; the question of overshoot comes down to how well the front can be projected into a piecewise-polynomial space. Coarse grids will still cause overshoot, but it will not be as serious as in the standard Galerkin and interior-penalty schemes. The combination of the finite element modified method of characteristics with a standard Galerkin procedure for pressure was analyzed by Russell [80]. Results of the form of (7.8) were derived, with time-truncation error of O(Ot) depending on the characteristic derivative 9 2 c/ar e as in 6. Numerical resuits were reported by Russell [81], using biquadratic polynomials with singularities removed for pressure and bilinears for concentration. The code was actually based on discontinuous bilinears with penalties, but the penalties were not helpful and were suppressed (i.e., a o was made large and Q, was set to zero); effectively, it was an inefficient continuous bilinear code. Radial initialization was used to 0.01 pore volumes injected, with sufficiently fine grid around injection wells to handle the front thereafter. Because of the smaller time-truncation error, large time steps of 40 to 80 days (0.02 to 0.04 pore volumes) could be taken without loss of accuracy, as was demonstrated by varying h and Ot in different runs. The accumulation term (^Cn, v) in (6.27) is evaluated by numerical integration; as noted in 6, Lobatto rules (usually Simpson's rule) were used. An
93
Concentration at .5 pore volume,
diagonal grid, 10 x 10 mesh.

O O
> O C W W O
U
WW C N
PORE VOLUNIES INTECTED

._.^,_.. l)iagonal grid .........' Parallel grid Recovery curve.
FIG. 16. Intertor penalty Galerkin procedure for the /ive-shot pattern witti M = 100.
/T
1.S0
exception was made in corner celis containing wells, because it is easier if no quadrature point falls at a welf. For points near wells, u* in (6.24) must be computed carefully because the vetocity field (extrapolated from two previous pressure time steps) changes rapidly in space. In [811, an attempt was made to do this semi-analytically by using the knowledge of the singular behavior of the
94
velocity; this proved to be unsatisfactory and was replaced in the later numerical work of Ewing, Russell, and Wheeler [291. The results, all with uniform permeability, showed very little grid orientation, numerical dispersion (as seen by varying h and At), or overshoot (0.5% for M = 100, 4dm = 1 , d, = d l = 0 on a 20 x 20 diagonal grid). A problem with M = 41 and longitudinal and transverse Peclet numbers of 100 and 1000 was handled successfully. Material-balance errors were another matter. While generally around 3%, they ranged as high as 10% in certain cases. Most of these errors occurred after solvent breakthrough at the production wells, indicating that the semi-analytic method for tracing characteristics from quadrature points near production wells was inadequate. Indeed, take v = 1 in (6.27) and assume that C = C at wells (this is truc at production wells, and also at injectors after a short time). The result is
(7.13)
S2
^Cn+ I f bCn) (x) dX =X dx,
which shows that mass-conservation error is determined entirely by how accurately one follows the characteristics of the velocity field and integrates (PC", and by the accuracy of the velocity field itself. This motivated a better scheme for computing u* in (6.24), and suggested the mixed method (5.18)(5.19) for pressure and velocity. It should be straightforward to analyze (5.18)(5.19) coupled to (6.27) by combining the results of Douglas, Ewing, and Wheeler [25] with those of Russell [80]; this has not been carried out. Numerical computations have been reported by Ewing, Russell, and Wheeler [29], combining the pressure code of Darlow, Ewing, and Wheeler [ 18] and the concentration code of Russell [81] with penalties removed. Characteristics from quadrature points near wells were traced in segments corresponding to micro-timesteps, each of these involving a trapezoidal predictor-corrector approximation of a chord along the characteristic. The size of each micro-timestep was selected by an error estimator using the inverse-radius behavior of the velocity field; details appear in Ewing, Russell, and Wheeler [29]. The desired results were achievedgrid orientation, numerical dispersion, and overshoot were still minimal, and mass-balance errors were reduced to the range of 0.5% in most runs. Figure 17 shows diagonal and parallel concentration maps and recovery curves for uniform permeability, M = 100, cd, = 1 ft 2 /day, dl = 0 foot, and d, = 0 foot. With the mixed method, it was possible to model random permeability variations of six orders of magnitude, as is shown by the physical fingers in Fig. 18 for the same mobility ratio and dispersion coefficients. The same problem was simulated by the cell-balance scheme (6.32) with physical dispersion replaced by numerical dispersion, with results pictured in Fig. 19. It is elear that the cell-balance method is much more dispersive for this example.
95
........
...
'^
^.
.............................'
Concentration at .5 pore volume,
diagonai grid, 20 x 20 mesh.
Lfl
n L) ^
O Ln W .
0 U W.
Lfl
.Be
.75 1.L3
PORE VOLUMES INTECTEO
_..""''-
20 x 20 diagonal grid 28 x 28 parallel grid

Recovery curve.
Fic,. 17. Modijied method of characteristics for the live-spot pattern with M = 100.
The mixed method also improved the stability of the procedure, helping to suppress nonphysical fingers. Figures 20 and 21 show concentration maps for M = 41, uniform permeabi1ity, od _= 0, d, = 10 Peet, and d, = 1 foot. These were computed by the finite element modified method of characteristics for concentra-
tion on 32 x 32 and 64 x 64 grids, respectively, and a standard Galerkin method

96
.1
FIG. 18. Concentration solution at .3 pore volume, mixed method and modified method of characteristicsrandom permeability and M = 100 (five-spot pattern).
for pressure. The mixed method was able to obtain results like those in Fig. 21 with a 32 x 32 concentration grid. The finite difference modified method of characteristics (6.26) has been analyzed by Douglas, coupled with either a five-point differente procedure [21 ] or a mixed method [20] for pressure. Preliminary unpublished computations of Douglas and Roberts have found significant grid orientation. The finite difference version (6.26) becomes complex if a grid point falls at a production well, since there is no unique characteristic emanating from the well; some type of averaging over various directions is needed. Only a first effort has been made at treating this, and this may be the source of the grid orientation.
FIG. 19. Concentration solution at .3 pore volume, finite element cell-balance procedurerandom permeability and M = 100 ( live-spot pattern).
97
F IG. 20. Concentration solution at .5 pore volume, M = 41, 32 x 32 concentration grid.
8. Future directions of research. The research reported to date has answered many questions, but it poses many others. We suggest possible lines of future investigation here. We consider ideas applicable to the various methods introduced in 5 and 6, and then discuss more general ideas. In the pressure methods, we need better treatment of near-well singularities (near-singularities in the truc physical problems), both theoretically and computationally. All of our high-order convergence theorems assume that the source and sink terms of (3.10) or (3.14) are bounded; for singular terms, the only proofs to date are for the unphysical assumption of concentration-independent viscosity. This theory needs to be ext.ended to the physical case. Computationally, removal
Ft6. 21. C oncentration sol ution al .5 pore volume, M = 41, 64 x 64 concentration grid.
98
of singularities works well if the coefficient a = k/ is constant, in which case the singularity can be removed from pressure, or if flow is radial, allowing removal from the velocity in the mixed method. A sensible procedure that has not been tried is to remove from velocity around an injection well, where flow is radial if permeability is constant and isotropic, and from pressure around a production well, where flow is not radial after solvent breakthrough. In cases of high mobility ratio, the viscosity contrast will make k/ vary by direction from the producer, and singularity removal becomes more difficult; perhaps a directional dependence should be incorporated into the theory. The methods should be easily extendable to anisotropic permeability (k ,, k,. distinct in (3.2) of Chapter I) by replacing circles by ellipses. For highly heterogeneous problems, the subtracted singularities may have to be cut off in small neighborhoods of the wells. The mixed method demands better techniques for solving its system of linear equations. So far we have used a conjugate gradient algorithm preconditioned by the diagonal of the matrix A of (5.32). This has been quite slow and better preconditioners are needed. The analogy between block-centered differences and mixed methods has suggested the use of the block-centered difference matrix as a preconditioner, and early results show satisfactory convergence in approximately ten iterations, independent of mesh size and heterogeneity. Another possibility is to use an alternating-direction iteration as a preconditioner either for the mixed method itself [111 or for the block-centered preconditioner. The modified method of characteristics for concentration needs accurate velocities in order to maintain small mass-conservation error, motivating the use of mixed methods for pressure. It would be of interest to carry out experiments on how accurate these velocities need to be and what order mixed method is most efficient for obtaining them. We would like to see performance comparisons between the lowest-order method (r = 0) on a fine grid and the next (r = 1) on a coarser grid. The question of conservation for the modified method of characteristics is an overriding one, and we should consider ways of post-processing the procedure to make it conservative. The conservation error in a typical time step is of the order of 10 -4 , so it should be possible via some projection or extrapolation of the numerical solution to compensate for this. The mixed method allows easy treatment of random permeability variations down to grid level. Since the interaction of heterogeneity, dispersion, mobility ratio, and instability (viscous fingering) is of such interest as noted in 2, we could fix a permeability distribution on a moderately coarse grid and then observe the behavior of the solution as the grid is refined. This refinement might reach levels requiring considerable expense, but it could shed a great deal of light on the fundamentals of the problem, perhaps showing the way for subsequent theoretical work. As an example, consider the plots in Figs. 20 and 21, generated for mobility ratio 41 and homogeneous permeability with a standard Galerkin method for pressure and the modified method of characteristics for the concentration. The multiple-fingered case used a 32 x 32 mesh for concentration, while
.
99
the stable case used 64 x 64. Otherwise the two runs were the same. Because the modified method of characteristics avoids numerical dispersion, oscillation, and grid orientation on reasonable grids, it should be possible to model the physics with a fair degree of accuracy. Another convergence study is suggested by some fragmentary numerical results with mixed methods for a single elliptic equation. These have pointed to possible superconvergence (convergence at a higher rate than the possible global order of approximation) of velocities at Gauss nodal points. This type of convergence is well known for Galerkin procedures at element nodes, and theoretical and numerical investigations for mixed methods would be worthwhile. Many workers in this area have thought that it would be interesting to somehow combine finite elements near wells with finite differences elsewhere in the reservoir, because finite differences are (aster but finite elements are better equipped to handle complicated near-well flow. Potempa's finite difference method based on finite element cells might provide ideas in this direct ion. There is no reason why that procedure demands upstream weighting; its numerical dispersion could be curtailed by use of midpoint weighting combined with a physical dispersion model discretized in the usual ways. The idea of flow between finite element cells would allow one to set up a global finite difference scheme, with standard block-centered flow coefficients away Erom wells and finiteelement-cell flows near wells. A somewhat neglected area in our numerical work is that of time-step selection. We have found that the modified method of characteristics permits much larger time steps than other procedures, but this has not been systematically quantihed. It would be useful to develop automatic time -step selectors that would push each method to its limits and/or balance spatial and temporal errors. Analogously, in the space variables adaptive grid refinement or generation should be considered. The advantages of the modified method of characteristics do not appear unless the grid is fine enough to resolve fronts, and in large-scale problems this may require adaptive mesh. Finally, the procedures will have to be extended to more complex problems than two-component single-phase displacement. The mixed method will be more complicated for multiphase flow because each phase has its own pressure and Darcy velocity; ideas like those of Chavent et al. [ 121 are one way of handling this. It should be possible to extend the modified method of characteristics to compressible, multiphase, multicomponent flows by noting that a Huid particle in a phase moves according to the interstitial velocity of that phase, which is its Darcy velocity (including the relative permeability as a multiplier) divided by the product of its saturation and the rock porosity. This allows transport of component compositions by the method of characteristics, after which phase behavior (mass transfer) could be updated and dispersion incorporated via finite differences or finite elements.
100
9. Conclusions. This research was motivated by the significant difficulties encountered by standard finite differente and finite element methods in simulating enhanced recovery processes of petroleum engineering. These processes lead to convection-dominated parabolic partial differential equations strongly coupled to elliptic or parabolic equations, as exemplified by the two-component model of miscible displacement. Finite differente procedures used in the petroleum industry smear traveling concentration fronts far bevond the level of physical dispersion, and replacement of upstream weighting by midpoint weighting would cause unacceptable nonphysical oscillations. Answers can be severely dependent on the orientation of the computational grid. Standard Galerkin methods can exhibit the same grid-orientation sensitivity for high-mobility-ratio displacements on coarse grids, yield nonphysical cusp-like solutions on coarse grids, and oscillate seriously even on rather fine grids. Dominant convection forces all of these methods to use small time steps. The interlor-penalty Galerkin procedures developed to attack grid orientation and cusping and to provide flexibility in grid refinement appear now to be too cumbersome and costly for the benefits they provide. The discontinuous version increases the order of the system of linear equations needed to attain a given accuracy of approximation, and the continuous one increases the bandwidth and density of nonzero entries in this system. Time steps must still be small, penalties on normal-derivative jumps add numerical dispersion, and even on rather fine grids solutions show overshoot and oscillations. Grid orientation can be alleviated by virtually any finite element method or non-upstream-weighted finite difference method on meshes adequate to resolve traveling fronts, and similarly cusping appears to have been an indication that the mesh was too coarse to handle the problem in question without numerical dispersion. If numerical dispersion is needed, it can be added more cheaply and without grid orientation by increasing the physical dispersion coefficient. The flexibility in grid refinement offered by the discontinuous penalty procedure may be advantageous in certain cases, but the authors have found the method cumbersome to program and expensive to run, and it would seem that there must be easier ways to achieve the same gains. The mixed method for pressure and velocity looks very promising if better ways of solving the associated system of linear equations can be found. The scheme treats highly heterogeneous problems, with permeabilities varying by six orders of magnitude, as easily as homogeneous problems. It has been shown theoretically and numerically to calculate velocities more accurately than the usual procedures that compute pressure and then differente or differentiate. The modified method of characteristics for concentration also looks very promising, provided that mass conservation can be kept under control and that the scheme can be generalized to more complex problems. Mass conservation is dependent on accurate velocities, making the coupling of this procedure with mixed methods particularly attractive. Theoretically and numerically, the
101
method has been shown to lose no accuracy with time steps much larger than those of other methods, perhaps one to two orders of magnitude larger, provided that the spatial grid is fine enough to resolve fronts. Answers show essentially no grid orientation, numerical dispersion, or overshoot. For problems with very high Peclet number, say in the thousands, adaptive grid refinement will be needed on the current generation of computers. Methods with mesh-dependent numerical dispersion, such as the upstream finite difference method based on finite element cell balances, are too dispersive for accurate simulation of problems with Peclet number in the hundreds unless adaptive refinement is used. They can provide qualitative results in an efficient and uncomplicated manner. The concept of finite element cell balances extends to meshes of arbitrary geometry. Overall, we feel that significant progress has been made in attacking the numerical difficulties typical of realistic petroleum reservoir simulation. The authors and others will continue these investigations and attempt to bring the ideas of this paper closer to practical application.
Acknowledgments. A great many of the ideas in this paper are due, in one form or another, to Jim Douglas, Jr. His influence on both of the authors' views of
reservoir simulation problems has been considerable. The assistante of T. C. Potempa in writing and running a block-centered finite difference code, and in producing plots with his finite element celI-balance procedure, was very valuable. We are indebted to Amoco Production Co., Arco Oil and Gas Co., Exxon Production Research Co., Getty Oil Co., IBM Corp., Marathon Oil Co., and Mobil Research and Development Corp., which supported much of this research. The first author thanks R. L. Christiansen and H. Kazemi for helpful discusslons
and the management of Marathon Oil Company for permission to publish this paper. The second author was supported in part by the National Science Foundation under grant MCS 80-03025.
REFERENCES
[1] R. A. ADAMS, Sobolev .Spaces, Academic Press, New York, 1975. [21 J. 0. AMAEFU1_l'. AND L. L. HANDY, The effect of interfacial tensions on relative oil/water
permeabilities of consolidated porous media, Soc. Pet. Eng. J., 22 (1982), pp. 371-381. 131 D. N . ARNOLD , An interiornenalty finite element method with discontinuous elemenis. Ph.D. thesis, Univ. Chicago, 1979; SIAM J. Numer. Anal., 19 (1982), pp. 742-760.
[4] K. Aziz_ AND A. SFTTARI, Petroleum Reservoir Simulation, Applied Science Publishers,
London, 1979.
[51 J. Bcu., G. SIIUBIN, AND M. F. WuE_rI.E^R, Analysis of a new method for computing the flow of miscible fl uid s in a porous medium, to appear.
[6] R. J. BI.ACKWEi.I , J. R. RAYNE, AND W. M. Ti RRY, Factors influencing the efficiency of
miscible displacement, Trans. Al M E, 216 (1959), pp. 1-8.

102
[7] J. H. BRAMBLE, On the convergente of difference approximations for second order uniformly
elliptic operators, in Numerical Solution of Field Problems in Continuum Physics, Vol. II, American Mathematical Society, Providence, RI, 1970, pp. 201-209. [8] J. H. BRAMBLE, B. E. HUBBARD, AND V. THOME, Convergente estimates for essentially positive type discrete Dirichlet problems, Math. Comp., 23 (1969), pp. 695-710. [9] F. BREZZI, On the existence, uniqueness and approximation of saddle-point problems arising Erom Lagrangian multipliers, RAIRO Anal. Numer., 2 (1974), pp. 129-151. [ 10] W. E. BRIGHAM, P. W. REED, AND J. N. DEW, Experiments on mixing during miscible displacement in porous media, Soc. Pet. Eng. J., 1 (1961), pp. 1-8. [11] D. C. BROWN, Alternating-direction iterative scheures for mixed finite element methods for second order elliptic problems, Ph.D. thesis, Univ. Chicago, 1982. [ 12] G. CHAVENT, J. JAFFRE, G. COHEN, M. DUPUY, AND 1. DIESTE, Simulation of twodimensional waterflooding using mixed finite element methods, SPE 10502, 6th SPE Symposium on Reservoir Simulation, New Orleans, 1982, pp. 147-158. [ 13] R. L. CFuUOKE, P. VAN MEURS, AND C. VAN DER POEL, The instability of slow, immiscible, viscous liquid-liquid displacements in permeable media, Trans. Al ME, 216 (1959), pp. 188-194. [ 14] P. G. CIARLET, The Finite Element Method for Elliptic Problems, North-Holland, Amsterdam, 1978. [ 15] E. L. CLARIDGE, Discussion of the use of capillary tube networks in reservoir performance studies, Soc. Pet. Eng. J., 12 (1972), pp. 352-361. [ 16] K. H. COATS, W. D. GEORGE, AND B. E. MARCUM, Three-dimensional simulation of steamflooding, Soc. Pet. Eng. J., 14 (1974), pp. 573 -592. [ 17] B. L. DARLOW, A penalty-Galerkin method for sol ving the miscible displacement problem, Ph.D. thesis, Rice Univ., Houston, 1980. [ 18] B. L. DARLOW, R. E. EwING, AND M. F. WHEELER, Mixed f nite element methods for miscible displacement problems in porous media, SPE 10501, 6th SPE Symposium on Reservoir Simulation, New Orleans, 1982, pp. 137-145; Soc. Pet. Eng. J., to appear. [ 19] J. DOUGLAS JR., The numerical solution of miscible displacement in porous media, in Computational Methods in Nonlinear Mechanics, J. T. Oden, ed., North-Holland, Amsterdam, 1980, pp. 225-238. , Simulation of miscible displacement in porous media by a modified method of [20] characteristic procedure, Numerical Analysis, Dundee 1981, Lecture Notes in Mathematics 912, Springer-Verlag, Berlin, 1982. [21] , Finite differente methods for two-phase incompressible flow in porous media, SIAM J. Numer. Anal., 20 (1983), pp. 681--696. [22] J. DOUGLAS JR. AND T. DUPONT, Galerkin methods for parabolic equations, SIAM J. Numer. Anal., 7 (1970), pp. 575 -626. , Interior penalty procedures for elliptic and parabolic Galerkin methods, Computing [23] Methods in Applied Science, Lecture Notes in Physics 58, Springer-Verlag, Berlin, 1976. [24] J. DOUGLAS JR., T. DUPONT, AND R. EWING, Incomplete iteration for time-stepping a Galerkin method for a quasilinear parabolic problem, STAM J. Numer. Anal., 16 (1979),
pp. 503-522.
[25] J. DOUGLAS JR., R. E. EwING, AND M. F. WHEELER, Approximation of the pressure by a mixed method in the simulation of miscible displacement, RAIRO Anal. Numer., 17
(1983), pp. 17-33.
[26]
[27]
, A time-discretization procedure for a mixed finite element approximation of miscible displacement in porous media, RAIRO Anal. Numer., to appear. J. DOUGLAS JR. AND T. F. RUSSELL, Numerical methods for convection-dominated diffusion problems based on combining the method of characteristics with finite element or finite difference procedures, STAM J. Numer. Anal., 19 (1982), pp. 871-885.
f
103
[28] J. DOUGLAS JR., M. F. WHEELER, B. L. DARLOW, AND R. P. KENDALL, Self-adaptivefinite
element simulation of miscible displacement in porous media, SIAM J. Sci. Statist. Comput. to appear. [29] R. E. EwING, T. F. RussELI., AND M. F. WHFELER, Simulation of miscible displacement using mixed methods and a modified method of characteristics, SPE 12241, 7th SPE Symposium on Reservoir Simulation, San Francisco, Nov. 1983. [30] R. E. EwING AND T. F. RussE.LL, Multisten Galerkin methods along characteristics for convection-difusion problems, in Advances in Computer Methods for Partial Differential Equations IV, R. Vichnevetsky and R. S. Stepleman, eds., IMACS, Rutgers Univ., New Brunswick, Ni, 1981, pp. 28 36. [311 ------- , Efficient time-stepping procedures for miscible displacement problenis in porous media, SIAM J. Numer. Anal., 19 (1982), pp. 1-67. [32] R. E. EwING AND M. F. WEIEt.t_t:R, Galerkin methods for miscible displacement problems in porous media, STAM J. Numer. Anal., 17 (1980), pp. 351-365. , Galerkin methods for miscible displacement problems with point sources and [33] ----- links--unit mohility ratio case, Proc. Special Year in Numerical Analysis, Lecture Notes #20, Univ. Maryland, College Park, 1981, pp. 151---174. , Computational aspects of mixed_finite element nmethods, in 1 MACS Transactions on [34] Scientific Computation, Vol. 1, R. S. Stepleman, ed., North-Holland, Amsterdam, 1983. [35] R. S. FALK AND J. E. OSBORN, Error estimates for mixed methods, RAIRO Anal. Numer., 14 (1980), pp. 249 277. [36] T. C. FRIK, ed., Petroleum Production Handbook, McGraw-E-li11, New York, 1962, pp. 23-36. [37] A. O. GARDER, D. W. PF.ACEMAN, AND A. L. Pozzi, Numerical calculation of multidimensional miscible displacement by the method of characteristics, Soc. Pet. Eng. J., 4 (1964), pp. 26-36. [38] S. GERSCtiGORiN, Fehlerabschtzung fiir das DifJerenzenverfahren zur Lsung partieller DijJerentialgleichungen, Z. Angew. Math. Mech., 10 (1930), pp. 373-- 382. [39] W. B. GOGARTY, Status of surfactant or micellar methods, J. Pet. Tech., 28 (1976), pp. 93-102. [40] S. B. GORr.LL AND G. M. Ho is , A theorl of the optima! policy of oi! recovertiv by secondarvv displacement processes, SIAM J. Appl. Math., 43 (1983), pp. 79--88. [41] W. G. GRAY, A derivation of the equations for multiphase transport, Chem. Eng. Sci., 30 (1975), pp. 229. 233. [421 W. G. GRAY AM) K. O'Niit.t., On the general equations for flow in porous media and their reduction to Darcl''s luw, Water Resources Res., 12 (1976), pp. 148-154. [43] S. P. GUPTA, Dispersive mixing effects on the Sloss field micellar system, Soc. Pet. Eng. J., 22 (1982), pp. 481-492. [44] B. HABERMANN, The efficiency of miscible displacement as a function of mobility ratio, Trans. Al M E, 219 (1960), pp. 264--272. [45] L. L. HANDY, An evaluation of dif usion effects in miscible displacement, Trans. AI ME, 216 (1959), pp. 382- 384. [46] L. J. HAYta, R. P. KEN DAI.L, AND M. F. WOLF EE R, The treatrvent of sources and sinks in steady-state reservoir engineering simulations, in Advances in Computer Methods for Partial Differential Equations l 1, R. Vichnevetsky, ed. IMACS, Rutgers Univ., New Brunswick, NJ, 1977, pp. 301-306. [47] J. P. HELLER, The interprefalion of ,nodel experiments for the displacement of (luids through porou s media, Al Ch E J., 9(1963), pp. 452-459. [48] ------- -, Onset of instability patterns between miscible jluids in porous media, J. Appl. Phys., 37 (1966), pp. 1566-1579. [49] T. J. R. Huc,EIES, ed., Finite Element Methods for Convection Dominated Flows, ASME,
,
New York, 1979.

104
[50] [51]
Pet. Tech., 34 (1982), pp. 2229-2232. error estimate for elliptic difference equations on a convex polygonal, SIAM J. Numer. Anal., 3 (1966), pp. 79-90. [52] H. A. KOCH JR. AND R. L. SLOBOD, Miscible slug process, Trans. AIME, 210 (1957), pp.
R. B. KELLOGG, An 40-47.
H. KAZEMI, Low permeability gas sands, J.
[53]
[54]
[55]
[56]
[57] [58]
T. A. MANTEUFFEL, B. K. SWARTZ, B. WENDROFF, AND A. B. WHITE, More aspects of supratruncation error convergence on nonuniform meshes, presented at STAM National Meeting, Denver, June 1983. C. R. KYLE AND R. L. PERRINE, Experimental studies of miscible displacement instability, Soc. Pet. Eng. J., 5 (1965), pp. 189-195. J. W. LACEY, J. E. FARIS, AND F. H. BRINKMAN, Effect of bank size on oil recovery in the high-pressure gas-driven LPG-bank process, J. Pet. Tech., 13 (1961), pp. 806-816. L. W. LAKE AND F. HELFFERICH, Cation exchange in chemical flooding: Part 2The effect of dispersion, cation exchange, and polymer/surfactant adsorption on chemical /lood environment, Soc. Pet. Eng. J., 18 (1978), pp. 435-444. L. W. LAKE AND G. J. HIRASAKI, Taylor's dispersion in stratified porous media, Soc. Pet. Eng. J., 21 (1981), pp. 459-468. A. LALLEMAND-BARRES AND P. PEAUDECERF, Investigation of the relationship between the value of the macroscopic dispersiveness of an aquifer medium, its other characteristics and the measurement conditionsbibliographic study, Bull. BRGM, Ser. 2, 3, #4 (1978), pp.
H.-O. KREISS,
.
277-284. [59] M. C. LEVERETT,
Flow of oil-water mixtures through unconsolidated sands, Trans. AIME, presentation, STAM National Meeting, Denver, June
[60]
132 (1939), pp. 149-171. T. A. MANTEUFFEL AND A. B. WHITE, 1983.
An Eulerian-Lagrangian numerical scheme for the dispersion-convection equation using conjugale space-time grids, J. Comp. Phys., 41 (1981), pp. 270-294. [62] J. A. NITSCHE AND J. C. C. NITSCHE, Error estimates for the numerical solution equations, Arch. Rational Mech. Anal., 5 (1960), pp. 293-306. [63] J. T. ODEN AND J. N. REDDY, .4n Introduction to the Mathematical Theory of Finite Elements, John Wiley, New York, 1976. [64] H. D. OUTMANS, Nonlinear theory for frontal stability and viscous fingering in porous media, Soc. Pet. Eng. J., 2 (1962), pp. 165-176. [65] J. G. PATEL, M. G. HEGDE, AND J. C. SLATTERY, Further discussion of two-phase flow in porous media, AIChE J., 18 (1972), pp. 1062-1063. [66] D. W. PEACEMAN, Improved treatment of dispersion in numerical calculation of multidimensional miscible displacements, Soc. Pet. Eng. J., 6 (1966), pp. 213-216. [67] , Fundamentals of Numerical Reservoir Simulation, Elsevier, Amsterdam, 1977. [68] , Interpretation of well-block pressures in numerical reservoir simulation, SPE 6893, 52nd Annual Fall Technical Conference and Exhibition, Denver, 1977; Soc. Pet. Eng. J., 18 (1978), pp. 183-194. [69] , Interpretation of well-block pressures in numerical reservoir simulation with nonsquare grid blocks and anisotropic permeability, Soc. Pet. Eng. J., 23 (1983), pp. 531-543. [70] J. PEAZ, P. REED, AND J. C. CALHOUN, Relationship between oil recovery, interfacial tension, pressure gradient in water-wet porous media, Prod. Mon., 19 (1955), No. 7, pp. 34-38. [71) T. K. PERKINS AND 0. C. JOHNSTON, A review of diffusion and dispersion in porous media, Soc. Pet. Eng. J., 3 (1963), pp. 70-84. [72] , A study of immiscible fingering in linear models, Soc. Pet. Eng. J., 9 (1969), pp. 39-46. [73] T. K. PERKINS, O. C. JOHNSTON, AND R. N. HOFFMAN, Mechanics of viscous fingering in miscible systems, Soc. Pet. Eng. J., 5 (1965), pp. 301-317.
[611 S. P. NEUMAN,
105
[74]
R. L. PF RRINE_,
[75]
[76] [77] [78] [79]
The development of stability theory for miscible liquid-liquid displacement, Soc. Pet. Eng. J., 1 (1961), pp. 17-25. --, A unified theory for stable and unstable miscible displacement, Soc. Pet. Eng. J., 3
(1963), pp. 205-213.
[80] [81]
[82]
[83]
The onset of instability during two-phase immiscible displacement in porous media, Soc. Pet. Eng. J., 21 (1981), pp. 249-258. O. PIRONNEAU, On the transport-diffusion algorithm and its application to the A'avierSiokes equations, Numer. Math., 38 (1982), pp. 309 ---332. T. C. POTEMPA, Finite element met hods for convection dominated transport problems, Ph.D. thesis, Rice Univ., Houston, 1982. P. A. RAVIART ANI) J. M. THOMAS, A mixedfinite element method for 2nd order elliptic problems, in Mathematica) Aspects of the Finite Element Method, Rome 1975, Lecture Notes in Mathematics, Springer-Verlag, Berlin, 1977. T. F. Russt.LL, An incompletely iterated characteristic finite element method for a miscible displacement problem, Ph.D. thesis, Univ. Chicago, Chicago, 1980. - ----------, Finite elements with characteristics for two-component incompressible miscible displacement, SPE 10500, 6th SPE Symposium on Reservoir Simulation, New Orleans, 1982, pp. 123--135. -- -- , Galerkin time stepping along charaeteristics for Burgers' equation, in 1 MACS Transactions on Scientific Computation, Vol. 1, R. S. Stepleman, ed., North-Holland, Amsterdam, 1983, pp. 183---192. P. G. SAFFMAN AND G. 1. TAYLOR, The penetration of a (luid into a porous medium or Hele-Shaw cel) containing a more viscous liquid, Proc. Roy. Soc., A245 (1958), pp.
E. J. PETERS AND D. L. FLOCK, 312-329.
[84] A. A. SAMARSKII, Local one dimensional scheures on non-uniform nets, Zh. Vych. Nat., 3, 3 (1963), pp. 431--466. (In Russian.) [85] P. H. SAMMON, Numerieal approximations for a miscible displacement process in porous media, to appear. [86] A. SUTTARI AND K. Aliz, Use of irregular grid in reservoir simulation, Soc. Pet. Eng. J., 12 (1972), pp. 103-114. [87] , Use of irregular grid in eylindrical coordinates, Soc. Pet. Eng. J., 14 (1974), pp. 396-404; Discussion, pp. 405-408. [88] A. SETTARI, H. S. PRICF AND T. DUPONT, Development and application of variational methods for simulation of miscible displacement in porous media, Soc. Pet. Eng. J., 17 (1977), pp. 228--246. [89] J. C. SLATTERY, Flow of viscoelaslie,Huids through porous media, AIChE J., 13 (1967), pp. 1066-1071. [90] ----, Multiphase viscoelastic flow through porous media, AIChE J., 14 (1968), pp. 50-56. [91 ] -- --, Single-phase flow through porous media, AIChE J., 15 (1969), pp. 866-872. [92] --, Two-phase f ow through porous media, AIChE J., 16 (1970), pp. 345-352. [93] - -, Interfacial efJects in the entrapment and displacement of residual oil, AIChE J., 20 (1974), pp. 1 145--1 154. [94] - - , Interfacial eJjrects in the recovery of residual oil by displacement, in 1982 report to our industrial consortium, Dept. of Chemical Engineering, Northwestern Univ., Evanston, I L, 1982. [95] R. L. SLOBOD AND S. J. LESTL, Use of a graded viscosity zone to reducefingering in miscible phase displaeements, Prod. Mon., 24 (1960), No. 10, pp. 12-19. [96] R. L. SLOBOD ANI) R. A. THOMAS, Effect of Iransverse diff asion on fingering in misciblephase displacement, Soc. Pet. Eng. J., 3 (1963), pp. 9--13. [97] G. STRAND ANI) G. J. Fix, An Analysis of the Finite Element Method, Prentice-Hall, Englewood Cliffs, Ni, 1973.
106
[98] J. J. TABER,
Dynamic and statie forces required to move a discontinuous oil phase Erom porous media containing both oil and water, Soc. Pet. Eng. J., 9 (1969), pp. 3-12. [99] G. I. TAYLOR, Dispersion of soluble matter in solvent flowing slowly through a tube, Proc. Roy. Soc., A219 (1953), pp. 186-203. [ 100] , Conditions under which dispersion of a solute in a stream of solvent can be used to measure molecular diffusion, Proc. Roy. Soc., A225 (1954), pp. 473-477. [1011 M. R. TODD, P. M. O'DELL, AND G. J. HIRASAKI, Methods for increased accuracy in numerical reservoir simulators, Soc. Pet. Eng. J., 12 (1972), pp. 515-530. [ 102] G. W. THOMAS, Principles of Hydrocarbon Reservoir Simulation, International Human Resources Development Corp., Boston, 1982. [ 103] C. VAN DER POEL, Effect of lateral diffusivity on miscible displacement in horizontal reservoirs, Soc. Pet. Eng. J., 2 (1962), pp. 317-326. [ 104] 0. R. WAGNER AND R. 0. LEACH, Effect of interfacial tension on displacement efficiency, Soc. Pet. Eng. J., 6 (1966), pp. 335-344. [ 105] J. E. WARREN AND J. C. CALHOUN JR., A study of waterflood e ficiency in oil-wet systems, Trans. AIME, 204 (1955), pp. 22-29. [ 106] J. E. WARREN AND F. F. SKIBA, Macroscopic dispersion, Soc. Pet. Eng. J., 4 (1964), pp. 215-230. [ 107] J. W. WATTS AND W. J. SILLIMAN, Numerical dispersion and the origins of grid orientation effecta summary, Paper 96C, presented at the 73rd Annual Meeting of AIChE, Chicago, 1980. [ 108] A. WEISER AND M. F. WHEELER, Block centeredfinite difference methods, to appear. [ 109] M. F. WHEELER, A priori L 2 error estimates for Galerkin approximations to parabolic partial differential equations, SIAM J. Numer. Anal., 10 (1973), pp. 723-759. [ 110] , An elliptic collocation finite element method with interior penalties, SIAM J. Numer. Anal., 15 (1978), pp. 152-161. [ 1 1 11 M. F. WHEELER AND B. L. DARLOW, Interior penalty Galerkin methods for miscible displacement problems in porous media, in Computational Methods in Nonlinear Mechanics, J. T. Oden, ed., North-Holland, Amsterdam, 1980, pp. 458-506. [ 112] S. WHITAKER, Diffusion and dispersion in porous media, AIChE J., 13 (1967), pp. 420-427. , Advances in theory of (luid motion in porous media, Ind. Eng. Chem., 61 (1969), [ 113) No. 12, pp. 14-28. [ 114] J. YANOSIK AND T. MCCRACKEN, A nine-point, finite difference reservoir simulator for realistic prediction of adverse mobility ratio displacements, Soc. Pet. Eng. J., 19 (1979), pp. 253-262. [ 115] W. F. YELLIG AND L. E. BAKER, Factors affecting miscibleflooding dispersion coefficients, J. Canad. Pet. Tech., 20 (1981), No. 4, pp. 69-75. [ 116] L. C. YOUNG, A frnite-element method for reservoir simulation, Soc. Pet. Eng. J., 21 (1981), pp. 115-128.
f

Chapter 2

Uploaded by

Copyright:

Available Formats

Chapter 2

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chapter 2

Uploaded by

Copyright:

Available Formats

CHA P TER II

MATHEMATICS OF RESERVOIR SIMULATION

M ETHODS FOR CONTI N UOUS FLOWS

Production well Injection well

F16. 2. Level curves of solvent concentration approaching solvent breakthrough.

Fc. 3. Level curves of solvent concentration after solvent breakthrough.

MATHEMATICS OF RESERVOIR SIMULATION

METHODS FOR CONTINUOUS FLOWS

MATHEMATICS OF RESERVOIR SIMULATION

METHODS FOR CONTI N UOUS FLOWS

MATHEMATICS OF RESERVOIR SIMULATION

METHODS FOR CONTI N UOUS FLOWS

MATHEMATICS OF RESERVOIR SIMULATION

METHODS FOR CONTINUOUS FLOWS

cpc (X, t)d-x,

(3.4) , __ drjuj u;.disp = diiu( el, e, 3e,

MATHEMATICS OF RESERVOIR SIMULATION

between , and the cartesian unit vector x , we use the relations

u X ul, sin8= u y,/lul,

to rotate coordinates and express (3.4) as

ac; 3c\ 2 ac uX ax + u x u}, a e.r + u x u 1, aX + u ti, a e V

where D is a diffusion-disperson tensor given by

4dm l+d,ju1T+d,ju1(1 T),

METHODS FOR CONTINUOUS FLOWS

ac at --- V (D(u)Vc -- uc) cq,

MATHEMATICS OF RESERVOIR SIMULATION

or constant - pressure, applicable when the reservoir is repressurized at the

boundary during depletion, say by an aquifer: (3.12a) (3.12b)

appropriate condition for c.

METHODS FOR CONTI N UOUS FLOWS

(3.1Oa), we see that V . (uc) = u .Vc+ c(V . u) =u.Vc-}- cq,

so that (3.10) may be rewritten (3.14a)

t --- 7 (D(u)Vc) + u Vc = ( -- c)q.

then we see that the concentration condition can be written

MATHEMATICS OF RESERVOIR SIMULATION

where (neglecting the gravity term in u)

4 .3) k- = 1 LX, + Ax1 ( harmonic average ), .3 ( g )^ r12 l^ II

METHODS FOR CONTI N UOUS FLOWS

Fic,. 4. Block-centered finite dijference coordinates.

MATHEMATICS OF RESERVOIR SIMULATION

METHODS FOR CONTINUOUS FLOWS

MATHEMATICS OF RESERVOIR SIMULATION

The steady-state equation (5.1) requires that

METHODS FOR CONTINUOUS FLOWS

Both schemes divide by x ; , 112 x, , 1 , to approximate p"; in the point-centered

]) : v is a polynomial of degree s ron each subinterval of},

MATHEMATICS OF RESERVOIR SIMULATION

1V11 = 11V110 = IIVIIL2(52) = I v dx dy '^',

METHODS FOR CONTI N IJOUS BLOWS

p PII -= a(h' 'IIrIL ). Iln - P (')(h`Ili)ll, i).

MATHEMATICS OF RESERVOIR SIMULATION

If we consider flow across the circle of radius R, (5.6) yields

' log r;(x, y) =

M ETHODS FOR CONTI N UOUS FLOWS

a;)Vp) - >(a ai)AP

A variational form of (5.10)(5.l 1) is, by Green's theorem,

((a --- a ; )Vp , V v)

MATHEMATICS OF RESERVOIR SIMULATION

principles that led to it:

Let W = L 2 (9), and denote the vector inner product on

[`4r I ( , .t) O , ^ r I(^)] x [jfl r (,r) .IYtO