1 B026 NON-UPWIND MONOTONICITY BASED FINITE VOLUME SCHEMES FOR HYPERBOLIC CONSERVATION LAWS IN POROUS MEDIA Michael G Edwards Civil and Computational Engineering Centre, University of Wales Swansea, Singleton Park, Swansea SA2 8PP, UK Abstract The focus of this paper is on the development of Mathematically robust convective flow approximation schemes that offer the benefits of an upwind formalism without actually upwinding. Development of robust schemes that remove upwind dependence would represent a significant step forward leading to a fundamental simplification of current methods. Introduction Standard reservoir simulation schemes employ single-point upstream weighting [1] (a first order upwind scheme) for approximation of the convective fluxes when multiple phases or components are present. The definition of the upstream schemes depends upon the direction of the local phase velocities. Higher order upwind approximations have also been developed [2-7]. The most successful high resolution schemes for systems of hyperbolic conservation laws in terms of actual front resolution, depend upon a characteristic decomposition of the system. The decomposition identifies characteristic wave components of a system enabling upwind approximations to be applied with minimum dissipation. A survey of schemes is given in [8]. For reservoir simulation mathematically more robust higher order Godunov schemes that are monotonicity preserving have also been developed for reservoir simulation [2, 7]. These schemes require a characteristic decomposition when applied to hyperbolic systems with multiple phases or components present. The decomposition leads to optimal upwind schemes where upwind directions are resolved according to characteristic wave components. However the decomposition adds further complexity compared to the standard scheme and additional computation is required to account for the decomposition matrices. Special treatment of stagnation points or “sonic points” where eigenvalues change sign is also required [2]. This paper presents novel robust schemes for reservoir simulation that permit reconstruction of stable lower order and higher order monotonicity preserving approximations while avoiding dependence upon both upwinding and characteristic decomposition. The formulation can also be used to simplify the higher order Godunov scheme locally in the presence of stagnation points (or if equal eigenvalues are detected), the details are beyond the scope of this paper. Here a new simplified scheme for reservoir simulation is presented that has no upwind (or upstream) dependence. 9th European Conference on the Mathematics of Oil Recovery — Cannes, France, 30 August - 2 September 2004 2 The schemes are formulated within a locally conservative finite volume framework and avoid dependence upon a characteristic decomposition by employing a form of Local Lax-Friedrichs (LLF) based flux [9]. Such schemes permit the construction of lower and higher order approximations without recourse to characteristic decomposition. This is achieved by using the local maximum eigenvalue of the hyperbolic system within the definition of the LLF flux. Flow Equations Without loss of generality in terms of the numerical methods applicability, the schemes presented here are illustrated with respect to two phase and three component two phase flow models, with unit porosity and where capillary pressure and dispersion are neglected. The conservation equations over Ω are written as ∂S ∂F ∂G ∫Ω ∂t dτ + ∫Ω ( ∂x + ∂y )dτ = M p 1 where F = (V1 , CV1 )T , G = (V2 , CV2 )T are the fluxes, (V1 ,V2 ) is the phase velocity vector (defined below), and S = ( S , SC )T , C = ( S , C )T are the vectors of conservative and primitive variables respectively, S is the miscible phase saturation and C the component concentration in the miscible phase, M p is a specified phase flow rate. The system considered here is comprised of an aqueous phase (water and polymer concentration) together with an oil phase. Phase quantities bear suffices a and o respectively. The velocities are defined via Darcy’s law, where for the aqueous phase velocity Va = (V1, V2 ) = − λa K (∇φ + ρ a g∇h ) In this work, the phase velocity is expressed in terms of total velocity VT = Va + Vo with Va = f ( VT − gλo ∆ρK∇h ) 2 where f = f ( S ) is the fractional flow, λa , λo are the phase mobilities and K is the permeability tensor. The saturation of the remaining phase is deduced from the volume balance equation, where saturations sum to unity. The incompressible flow condition leads to ∫ ∇ • VT dτ = M 3 which together with Eq. 2 are used to determine pressure and velocity, for further details of flow equations see [1]. For the initial value problem (IVP) field data is prescribed. For initial boundary value problems (IBVP), considered here in two-dimensions, an initial flow field is prescribed together with flux or pressure boundary values. Zero normal flow is imposed on solid walls. Locally Conservative Upwind Approximation First we shall consider schemes in one spatial dimension on a computational grid with discrete nodes xi = i∆x and time t n = n∆t . Before moving to systems we briefly recall that for a scalar equation of the form ∂s ∂f ( s ) + =0 ∂t ∂x 4 3 the standard explicit first order upwind scheme can be written as sin +1 = sin − ∆t n ( f i +1 / 2 − f in+1 / 2 ) ∆x 5 where the approximate flux is defined by f in+1 / 2 = ( 1 ( f in+1 + f in ) − λi +1 / 2 ( sin+1 − sin ) 2 ) 6 where ⎧⎪( f n − λi +1 / 2 = ⎨ i +1 ⎪⎩ f in ) /( sin+1 − sin ) ∂f / ∂s ( sin+1 − sin ) > ε ( sin+1 − sin ) ≤ ε This definition of wave speed ensures that shocks are captured with precision for any finite jump in si with λi +1 / 2 assuming the Rankine-Hugoniot shock speed across a mesh interval. In this form the first order scheme appears as a central scheme and is comprised of a central difference in flux with a central difference of a specific diffusion term. The scheme can be seen in its original upwind form by noting that for a positive wave speed the flux uses data to the left and reduces to f in+1 / 2 = f ( sin ) , otherwise f in+1 / 2 = f ( sin+1 ) and the flux uses data to the right as with single point upstream weighting. While the definition of Eq's. 5, 6 does not require any explicit sign dependence in the scheme, the upwind directions are clearly detected. This explicit scheme is the most fundamental scheme for scalar conservation laws in one dimension and is stable, locally conservative and monotonicity preserving subject to a maximum CFL condition of unity. This scheme also requires an entropy fix to disperse expansion shocks [8]. In order to apply such a scheme to a system of hyperbolic conservation laws of the form of Eq 1 the system is first decomposed into characteristic form. First we consider the system in one dimension. Decomposition is performed via the transformation ∆S = R ∆W 7 where R is the matrix of right eigen-vectors of the system Jacobian matrix A = ∂F / ∂S and the matrix of eigenvalues Λ is defined via Λ = R −1 AR 8 and ∆S, ∆W represent the respective conservative and characteristic variable increments. The upwind scheme is in effect applied to each characteristic wave component and the discrete system is recomposed into conservation form. The first order scheme for a system can be written as S in +1 = S in − ∆t n (fi +1 / 2 − fin+1 / 2 ) ∆x 9 with approximate flux defined by 9th European Conference on the Mathematics of Oil Recovery — Cannes, France, 30 August - 2 September 2004 4 fin+1 / 2 = ( 1 n (fi +1 + fin ) − R Λ i +1 / 2 R −1 (S in+1 − Sin ) 2 ) 10 provided the definition of the discrete eigenvalues Λ i +1 / 2 have an appropriate generalization of Eq. 6 such that conservation and exact shock speed are maintained. The CFL condition now applies with respect to the maximum eigenvalue of the system. Locally Conservative Non-Upwind Approximation in One Dimension The appearance of the matrix of eigenvectors Ri +1 / 2 in the system flux approximation is a consequence of upwinding on each characteristic component. This leads to an optimal diffusive operator in terms of numerical diffusion provided the shock jump criteria can be satisfied for the system in hand. If the matrix of eigenvalues is proportional to the unit matrix (equal eigenvalues) then the dependency of the discrete flux on the matrix of eigenvectors is removed, which essentially leaves a much simpler coefficient of artificial diffusion. The scheme can be simplified in this way without appearing to violate the crucial monotonicity preserving property of the scheme, subject to the CFL condition, by replacing the diagonal matrix of absolute eigenvalues Λ i +1 / 2 with the matrix Λ LFi +1 / 2 = max[x L , x R ] λij+1 / 2 I 11 j and the maximum absolute eigen-value of the system, (here max( ∂Va / ∂s , Va / s ) I ) over the local interval [xi , xi +1 ] is used. The discrete flux then reduces to fin+1 / 2 = ( 1 n (fi +1 + fin ) − Λ LFi +1 / 2 (Sin+1 − Sin ) 2 ) 12 The flux of Eq. 12 is called the Local Lax-Friedrich flux LLF. The non-upwind LLF scheme is then defined be Eq's. 12 and 9. Note that the scheme is still locally conservative. Extension to higher order accuracy is discussed in the next section. Since all eigenvalues in the diffusion operator are replaced by their maximum modulus, it follows that the price to be paid for this simplification is extra numerical diffusion. Higher Order Schemes Without Upwinding In the scalar case a higher order approximation is applied to the conservation variable. When an upwind scheme is applied to a system the higher order approximation is typically introduced wave by wave and applied to the characteristic variables, followed by recomposition to the conservative variables. In contrast, when using the discrete LLF flux Eq. 12, the extension to higher order accuracy is simpler to achieve since the non-upwind system formulations have no dependency upon characteristic variables in this definition of flux. Consequently a higher order approximation can be introduced for the left and right states respectively and be expressed directly in terms of the conservative variables. Alternatively the higher order expansions can also be applied to other sets of variables such as primitive or characteristic variables. In this work the primitive variables are selected. 5 In one dimension the scheme is expressed as a two-step process. First the higher order states are defined using a MUSCL formalism [10]. Higher order left and right hand side states are obtained by expansions about the states L and R, viz 1 S L = Si + PΦ ( ri++1 / 2 )(Cin+1 − Cin ) 2 1 S R = Si +1 − PΦ ( ri−+1 / 2 )(Cin+1 − Cin ) 2 13 where Φ ( ri++1 / 2 ) and Φ ( ri−+1 / 2 ) are flux limiters, or in this case slope limiters [10] and P is the transformation matrix between conservative and primitive variables. The slope limiters are functions of adjacent discrete gradients where ri++1 / 2 = (Cin − Cin−1 ) /(Cin+1 − Cin ), ri−+1 / 2 = (Cin+ 2 − Cin+1 ) /(Cin+1 − Cin ) 14 The slope limiters constrain the expansions to ensure that the higher order data remains monotonic. Details can be found in [10, 7 and 11]. The above flux of Eq. 12 is now applied to the higher order data so that the local Riemann problem is resolved with a higher order flux of the form fin+1 / 2 = 1 ((f (S L ) + f (S R )) − Λ LFi +1 / 2 (S R − S L ) ) 2 15 The flux of Eq. 15 is now used to integrate the system via Eq. 9. The higher order spatial scheme as defined by Eq's 13-15 and 9 which uses first order forward-Euler time stepping. The CFL condition of the scheme is dependent on the choice of limiter, typically an upper limit of 1/2 is selected. Extension to higher order time accuracy is achieved either by using the scheme of [6] or with the second or third order monotonicity preserving Runge-Kutta schemes of [12]. Note that the first order flux is recovered if the limiters are set to zero. Approximate Riemann Solvers Without Upwinding in Two Dimensions The higher dimensional extension of the above scheme is based on a generalization of the one dimensional discrete flux. The flow equations Eq. 1 are integrated in space and time over a discrete control-volume Ω CV with surface δΩ CV by direct use of the Gauss divergence theorem applied to yield a surface integral of divergence ∫Ω CV (S(t + ∆t ) − S(t + ∆t ))dτ = − ∫∆t ∫δΩ CV (Fdy − Gdx )dt + ∫∆t M p dt 16 The discrete scheme is developed from Eq 16, space only permits a brief summary here. The two-dimensional normal flux ℑ = ( F∆y − G∆x ) / ( ∆y , − ∆x ) is approximated in an analogous form to Eq 15. Higher order states are defined using a MUSCL based formalism [10], which essentially follows the above 1-D principle applied relative to each control-volume face. The higher order left and right hand side states are obtained by Taylor expansions and as in 1-D the expansions are constrained with slope limiters to ensure that the higher order data remains monotonic. For a structured grid limiting is performed in the co-ordinate directions. The limiting is based on the Fromm limiter Φ ( r ) = min(2,2r, (1 + r ) / 2) . Note as before, that the first order 9th European Conference on the Mathematics of Oil Recovery — Cannes, France, 30 August - 2 September 2004 6 flux is recovered if the limiters are set to zero. Formal time integration is achieved as above, employing either [6] or [12], without disturbing monotonicity subject to a reduced theoretical CFL limit in 2-D (as with standard schemes), although here up to 1/2 is used. Discrete approximations developed for general unstructured grids comprised of quadrilateral and-or triangular cells will be presented in a future report. RESULTS One Dimension The schemes are first compared for gravity segregation with reverse flow in one-dimension. Initial data consists of a vertical aqueous phase column above an oil column of equal volume (95% oil saturation) , separated by a diaphram which is removed at time t=0. The heavier miscible phase moves downwards, creating a shock followed by a constant state, contact discontinuity (due to polymer) and rarefaction wave. The oil moves upwards creating a shock at the opposite end of the rarefaction. The computed and exact water saturation solutions are shown in Fig’s 1-3 at time 1.25, (exact solution solid line, computed using 100 nodes square symbol) with the prescribed initial data S,C = 0.05, 0.1, x < 0.5 , S,C = 1.0, 0.7 otherwise. The first order scheme result is shown in Fig.1 and completely smears the solution, failing to detect the contact, constant state and shock caused by the miscible phase. The higher order LLF scheme and higher order Godunov (HOG) scheme results are shown in Fig’s 2 and 3 respectively. The comparison for water saturation shows that both the higher order schemes resolve the two shocks and contact discontinuity induced by the polymer concentration discontinuity. While the results of Fig. 2 indicate additional diffusion that is inherent in the LLF scheme in the region of the contact when compared to HOG, the non-upwind scheme achieves a considerable improvement in front resolution compared to the first order scheme. Two Dimensions The reverse flow problem is next considered in 2-D in the presence of a shale barrier with solid walls at the sides and top boundaries, and pressure specified on the lower boundary, the boundaries and initial interface is shown in Fig 4. All saturation contours are shown at the same output time, where the shock due to the upward moving oil phase is just being reflected from the top wall. The higher order scheme results are; HOG in Fig.5, HOG with LLF at sign change Fig. 6, and full LLF in Fig.7. The higher order schemes are each able to detect the shocks, expansion about the shale corner and the weaker linear contact wave. As in 1-D shocks are again well resolved, the sensitive linear contact wave is more spread. The simplified HOG scheme result shown in Fig 6 produces a similar result to HOG, with a slight increase of spread in the contact due to the local additional diffusion. The higher order LLF scheme result, Fig. 7, shows that the non-upwind scheme is able to capture the key flow features with similar resolution. A comparison the first order scheme result Fig. 8, shows that the base scheme fails to detect the contact and constant state as they pass the shale, and provides poor resolution of the lower shock and the slow moving shock directly above the shale. The non-upwind higher order LLF scheme Fig. 7 achieves a considerable improvement in front resolution compared to the first order scheme and detects all discontinuities that are present in the HOG solution. Conclusions A novel formalism is presented for hyperbolic conservation laws in reservoir simulation. The scheme has the following properties: 7 1) Upwinding and characteristic decomposition are circumvented, leading to a fundamental simplification of current methods. 2) A sound mathematical basis for stability and monotonicity is retained while avoiding both upwinding and characteristic decomposition. 3) Local conservation is maintained and the new formulation offers consistency in IMPES mode (or sequentially implicit mode), with respect to sign change in wave velocity and mass balance. 4) The new formulation can also be used to simplify the higher order Godunov scheme in the presence of sign change in wave velocity or in occurrence of degenerate equal eigenvalues. Results are presented for gravity driven flows involving two phase flow and three component two-phase flow, where the wave direction changes sign (reverse flow). Results obtained with the new schemes indicate a competitive comparison with the higher order Godunov scheme. Finally it is noted that the new formulation offers the benefits of being directly applicable to other systems of hyperbolic conservation laws without requiring a characteristic decomposition. Acknowledgement: Support of EPSRC grant GR/S70968/01 is gratefully acknowledged. References 1. 2. Aziz K and Settari A. “Petroleum Reservoir Simulation” Applied Science Publishers, London, 1979 Bell J. B. Colella P. and Trangenstein J.A. “Higher Order Godunov methods for general systems of hyperbolic conservation laws.” J. Comput. Phys 82 1989 362-397. 3. Blunt M. J. and Rubin B. “Implicit Flux-Limiting Schemes for Petroleum Reservoir Simulation”. J. Comput. Phys 102 1992 194-210. 4. Rubin B. and Edwards M.G. ''Extension of the TVD Mid-Point Scheme to Higher Order Accuracy in Time'' SPE 25265 Twelfth SPE Reservoir Simulation Symposium, New Orleans Louisiana USA, pp 375386, Feb 28 - Mar 3 1993 5. Edwards M G. Delshad M. Pope G. A. and Sepehrnoori “A High Resolution Method Coupled with Local Grid refinement for Three Dimensional Aquifer Remediation” In Situ, 23, 4, pp 333-377, 1999 6. Thiele M and Edwards M G “Physically Based Higher Order Godunov Schemes for Compositional Simulation” SPE 66403 SPE Reservoir Simulation Symposium Houston TX USA 11-14 Feb 2001 7. Edwards M.G. "A Higher Order Godunov Scheme Coupled With Dynamic Local Grid Refinement for Flow In a Porous Medium" Comput. Methods. Appl. Mech. Engrg , Vol 131, pp 287 - 308, 1996. 8. Godlewski E. and Raviart P. Numerical Approximation of Hyperbolic Systems of Conservation Laws App. Math. Sci. 118 Springer-Verlag, New York. 1996 9. Liu X. D. and Osher S. "Convex ENO High Order Multi-Dimensional Schemes without Field by Field Decomposition or Staggered Grids" J. Comput. Phys 141, 1-27 1998 10. Van Leer B. “Towards the Ultimate Conservative Difference Scheme, V. A second-order sequel to Godunov’s method. J. Comput. Phys. 32 1979 101-136. 11. Sweby P. K. “High resolution Schemes using Flux Limiters for Hyperbolic Conservation Laws” SIAM J. Numer. Anal. 21 1984 995-1011. 12. Shu C. W. and Osher S. Efficient Implementation of Essentially Non-Oscillatory Shock Capturing Schemes J.Comput. Phys, 77:439-471 1988. 9th European Conference on the Mathematics of Oil Recovery — Cannes, France, 30 August - 2 September 2004 8 1.2 1 0.8 1st order Exact 0.6 0.4 0.2 0 0 0.2 0.4 0.6 0.8 1 1.2 Fig 1 1st order versus exact solution 1.2 1 0.8 LLF HO 0.6 Exact 0.4 0.2 0 0 0.2 0.4 0.6 0.8 1 1.2 Fig 2 higher order LLF versus exact solution 1.2 1 0.8 HOG Exact 0.6 0.4 0.2 0 0 0.2 0.4 0.6 0.8 1 Fig 3 higher order Godunov versus exact solution 1.2 9 Water+polymer Fluid interface Shale barrier Oil Fig 4: Initial Conditions 9th Fig 5: Higher Order Godunov Fig 6: Higher Order Godunov (LLF) Fig 7: Higher Order LLF (non-upwind) Fig 8: First Order European Conference on the Mathematics of Oil Recovery — Cannes, France, 30 August - 2 September 2004 10