You are on page 1of 18

# Journal of Computational Physics 229 (2010) 4813–4830

Contents lists available at ScienceDirect

Journal of Computational Physics
journal homepage: www.elsevier.com/locate/jcp

A fully coupled 3-D mixed ﬁnite element model of Biot consolidation
Massimiliano Ferronato *, Nicola Castelletto, Giuseppe Gambolati
Dept. Mathematical Methods and Models for Scientiﬁc Applications, University of Padova, via Trieste 63, 35121 Padova, Italy

a r t i c l e

i n f o

Article history:
Received in revised form 3 March 2010
Accepted 14 March 2010
Available online 17 March 2010
Keywords:
Mixed ﬁnite elements
Biot consolidation
Coupled solution

a b s t r a c t
The numerical solution to the Biot equations of 3-D consolidation is still a challenging task
because of the ill-conditioning of the resulting algebraic system and the instabilities that
may affect the pore pressure solution. Recently new approaches have been advanced based
on mixed formulations. In the present paper a fully coupled 3-D mixed ﬁnite element
model is developed with the aim at alleviating the pore pressure numerical oscillations
at the interface between materials with different permeabilities. A solution algorithm is
implemented that takes advantage of the block structure of the discretized problem. The
proposed model is veriﬁed against well-known analytical solutions and successfully experimented with in realistic applications of soil consolidation.

1. Introduction
Poro-elasticity denotes the coupled process between mechanics and ﬂow in porous media. Its theoretical basis goes
back to the mid 1920s when Terzaghi described analytically the one-dimensional (1-D) consolidation of a soil column under a constant load [1]. In 1941 Biot generalized Terzaghi’s theory to three-dimensional (3-D) porous media [2] by establishing the mathematical framework which is usually termed as poro-elasticity. Even though sophisticated, and perhaps
more elegant, approaches have been recently advanced, e.g. [3], the Biot equations of consolidation are still used today
in a great variety of ﬁelds, ranging from reservoir engineering to biomechanics. For example, poro-elasticity is the basic
theory to predict the compaction of a producing hydrocarbon reservoir and the related hazards, including land subsidence
and borehole damages [4–6]. Several environmental issues connected with groundwater withdrawal, e.g. [7], or the safe
long-term disposal of wastes in the subsurface, e.g. [8], can be addressed with the aid of poro-elastic models. In biomechanics the poro-elastic theory is used to describe tumor-induced stresses in the brain [9] and the bone deformation under
Despite the intensive research in the area, the numerical solution to the Biot partial differential equations is still a challenging task for a number of reasons. First, the fully coupled approach leads to algebraic systems which can be quite difﬁcult
to solve, with the number of unknowns easily growing up to several hundreds of thousands in real 3-D applications. Second,
the coefﬁcient matrix resulting from the numerical discretization can be severely ill-conditioned especially at the early stage
of the process [11], hence advanced preconditioners and solvers are needed to handle the fully coupled system [12,13]. Another approach relies on the iteratively coupled scheme which solves separately in a staggered way the mechanical and the
ﬂuid ﬂow models until convergence. If sufﬁciently tight convergence criteria are prescribed, the iteratively coupled solution
can be as accurate as the fully coupled one, but the number of required steps might be prohibitively high. Moreover, convergence of the iterative procedure is not always guaranteed depending on the ﬂuid and soil mechanical properties [14].

* Corresponding author. Tel.: +39 049 8271332; fax: +39 049 8271333.
URL: http://www.dmsa.unipd.it/~ferronat (M. Ferronato).
doi:10.1016/j.jcp.2010.03.018

tÞ u nn 5 4P ‘i ðxÞuz. i ¼ 1.i ðtÞ 2 ð6Þ i¼1 The ﬂuid pore pressure pðx.i . tÞ > > : v ðx. . with Bu the strain elastic matrix) and rtot is the total stress vector (¼ ½rx. respectively. pðtÞ and qðtÞ whose components are the nodal displacements ux. A weak form of the coupled system (1)–(3) is obtained using the classical approaches followed in both elasto-statics and ﬂuid-dynamics. The unknown functions u ^ ðx. Ferronato et al. n the outer normal to C and x the position vector in In Eqs. (4) and (5) CD [ CN ¼ Cp [ Cq ¼ C. the function wk gives rise to a unitary ﬂux through the kth face and a null ﬂux through all other faces. cyz . . Continuity of wk is ensured only along the normal direction relative to the kth face. .tot . tÞ > pðx. . Variational formulation The partial differential system (1)–(3) with boundary conditions (4) and initial conditions (5) is solved numerically using ^ . cxy . are the discrete unknowns of the variational problem. the integral form is found by minimizing the total potential energy in the domain X with the aid of the virtual work principle: Z X v . .i ðtÞ 7 u 7 ¼ Nu ðxÞuðtÞ 7 6 i¼1 7 6 ^ z ðx. according to the classical Voigt notation [28]. z .M.tot . uz. 2. tÞnðxÞ ¼ tN ðx. As for Eq. . tÞ u > > > <r  tot ðx. . .i ðtÞ 3 7 3 6 i¼1 7 6 ^ x ðx. nn . i ¼ 1. respectively. we set: pðx. The medium displacement u imated in the space generated by the continuous piecewise linear polynomials ‘i . while the right-hand sides are known functions. tÞ u 7 6P nn 6 6 7 ^ ðx.T rtot dX ¼ Z ^ v . p and v are found in an appropriate space such that the integral form of a variational formulation. 0Þ ¼ u ^ 0 ðxÞ u ð5Þ pðx. . respectively. Eq. rz. The vectors uðtÞ. . j ¼ 1. ne . nn . sxy. . tÞ ’ ne X T hj ðxÞpj ðtÞ ¼ h ðxÞpðtÞ ð7Þ wk ðxÞqk ðtÞ ¼ WðxÞqðtÞ ð8Þ j¼1 vðx. y .T b dX þ u X Z ^ v .i . tÞ and Darcy’s velocity v ðx.T tN dC u ð11Þ CN where. tÞ 5 ’ 6 ‘i ðxÞuy. (10) implies that wk is not zero over the tetrahedrons sharing the kth face and is discontinuous across it. The apex v denotes the virtual variables. Denoting by ne and nf the number of elements and faces.1. .tot . syz. .tot . tÞ  nðxÞ ¼ qN ðx.tot . tÞ 4815 over CD over CN ð4Þ over Cp over Cq and initial conditions:  ^ ðx. uy. r R3 . 0Þ ¼ p0 ðxÞ  tot is the total stress tensor.i . tÞ ¼ pD ðx. The ± sign in (10) identiﬁes a conventional face orientation such that wk points outward the element T ðjÞ with the smallest index j. Recalling Terzaghi’s principle: . sxz. (1). Because of this orientation. tÞ are represented in the piecewise constant polynomial and lowest order Raviart–Thomas space. / Journal of Computational Physics 229 (2010) 4813–4830 8 ^ ðx. the elemental pressures pj . ry. . . tÞ ¼ 4 u ^y ðx. . with nn the number of FE nodes in X: 2P nn ‘i ðxÞux. tÞ is approx(1)–(3) is satisﬁed for all test functions belonging to the corresponding space. k ¼ 1. and the face normal ﬂuxes qk . cxz T ¼ Bu u. nf .  is the strain vector (¼ ½x . thus satisfying the LBB condition that ensures the well-posedness of the discrete problem [27]. tÞ ’ nf X k¼1 where hj are element-wise constant functions: ( hj ðxÞ ¼ 1 x 2 T ðjÞ 0 ð9Þ x 2 X  T ðjÞ and wk are vectorial functions in R3 associated to the kth face belonging to the jth tetrahedron T ðjÞ : ( wk ðxÞ ¼ ðxxk Þ x 2 T ðjÞ  3jVðT ðjÞ Þj ð10Þ x 2 X  T ðjÞ 0 with V the volume of T ðjÞ and xk the position vector of the node opposite to the kth face in T ðjÞ .tot T ). tÞ ¼ u ^ D ðx.

k ¼ 1. (20) and (21) is numerically integrated in time by a ﬁnite difference scheme. . with De the elastic constant matrix) and i the Kronecker delta in vectorial form. . Unlike standard FEs. Numerical implementation The system of differential-algebraic equations (14). syz . . 2. Using the approximations (6)–(8) gives: Z hdivv dX þ X Z Z h X  1 v dX þ WT j X Z @/bp dX þ @t Z h X W T rp dX ¼ 0 ^ @ adivu dX ¼ @t Z hf dX ð18Þ X ð19Þ X Assuming that /. (27) h is a scalar value comprised between 0 and 1. The discrete solution scheme thus obtained is the following: . ry . In Eq. whereas the Neumann boundary conditions are to be imposed in a strong way. the semi-discrete mixed FE expression of Eqs. (2) and (3) is obtained by a standard Galerkin approach.2. the Dirichlet boundary conditions are the natural conditions for Eq. b and a are independent of time and using a weak form for the last integral in Eq. and differentiating with respect to the virtual displacements. Consider any time-dependent function to vary linearly in time between t and t þ Dt. The integral form of Eqs. nf . (1): Ku  Q p ¼ f 1 ð14Þ where Z K¼ BTu De Bu dX X Z Q¼ aBTu ihT dX X f1 ¼ ð15Þ Z X ð16Þ Z NTu b dX þ CN NTu tN dC ð17Þ The Dirichlet boundary conditions along CD are prescribed in a strong way. (2) and (3) read: BT q þ Pp_ þ Q T u_ ¼ f 2 ð20Þ Aq  Bp ¼ f 3 ð21Þ where Z A¼  1 W dX WTj ð22Þ xhT dX ð23Þ X B¼ Z X P¼ Z T ð24Þ hf dX ð25Þ /bhh dX X f2 ¼ Z X f3 ¼  Z pD W T n dC ð26Þ Cp The components of x in (23) are equal to divðwk Þ. Eq. . Ferronato et al. rz . (21) and are therefore prescribed in a weak way. sxz T ¼ De . and approximate any timederivative at the intermediate instant s: s ¼ hðt þ DtÞ þ ð1  hÞt ð27Þ by a ﬁrst-order incremental ratio. / Journal of Computational Physics 229 (2010) 4813–4830 rtot ¼ r  api ð12Þ where r is the effective stress vector (¼ ½rx . sxy . (19). (11) yields: Z  Z Z Z BTu De Bu dX u  aBTu ip dX ¼ NTu b dX þ X X X NTu tN dC CN ð13Þ Introducing the pore pressure approximation (7) into (13) produces the ﬁnal discrete form of Eq.4816 M.

/ Journal of Computational Physics 229 (2010) 4813–4830 4817 h i   t tþDt h KutþDt  Q ptþDt ¼ ð1  hÞ Qpt  Kut þ f 1 þ hf 1 ð28Þ h i Q ut þ Ppt Q utþDt þ PptþDt t tþDt þ hBT qtþDt ¼ ð1  hÞ f 2  BT qt þ þ hf 2 Dt Dt h i ð29Þ T T h½AqtþDt  BptþDt  ¼ ð1  hÞ Bpt  Aqt þ t f3 þ tþDt hf 3 ð30Þ where the initial values used for q are consistent with the initial pressure space distribution p0 ðxÞ. a speciﬁc block version of a preconditioned Krylov subspace method is to be implemented. while A. (29) by Dt and Eq. The numerical solution at time t þ Dt can be therefore computed by solving a linear algebraic system: AztþDt ¼ f t ð31Þ where 2 QT P ðpÞ f ðuÞ f ðqÞ ptþDt 3 2 f ðpÞ 3 7 tþDt 6 tþDt 7 t 6 ðuÞ 7 ¼ 4u 5 f ¼ 4f 5 0 5 z ðqÞ qtþDt cB 0 cA f h i t tþDt ¼ ðDt  cÞ f 2  BT qt þ Q T ut þ Ppt þ cf 2 h i t tþDt ¼ w Kut  Q pt  f 1  f 1 h i t tþDt ¼ ðDt  cÞ Aqt  Bpt  f 3  cf 3 6 A¼4Q f 2 3 cBT K ð32Þ ð33Þ ð34Þ ð35Þ The matrix A in (31) has size ne þ 3nn þ nf and is sparse. is generally not convenient from a computational point of view. (28) by h. however. (38) and (39) could be computed exactly as P is diagonal. as both K and A are SPD and c is positive. Now consider the following A decomposition:  A¼ I 0 HP 1 I  P 0 0 S " I P 1 HT 0 I # ð38Þ where S is the SPD Schur complement: " 1 T S ¼ C þ HP H ¼ K þ QP 1 Q T cBP1 Q T cQP1 BT cA þ c2 BP1 BT # ð39Þ Eqs. A changes at each step because Dt. Suitable solvers for (31) belong to the class of the iterative projection-type Krylov subspace methods properly preconditioned. Write A as a 2  2 block matrix: " A¼ P HT H C # ð36Þ with:  H¼  Q  C¼ cB  K 0 0 cA ð37Þ Note that C is a block-diagonal symmetric positive deﬁnite (SPD) matrix. The explicit construction of A. hence c. Therefore.13]. B. we develop a variant of the block constraint approach successfully applied to standard FE consolidation models [12. K. however S turns out to be much less sparse than C. multiply Eq. P and Q can be computed just once at the beginning of the simulation as they do not depend on t. Ferronato et al. A simple approximation relies on prescribing the dropped S to preserve the same block-diagonal structure as C: " S’e S¼ e S1 0 0 e S2 # " ¼ K þ QP 1 Q T 0 0 cA þ c2 BP1 BT # ð40Þ The blocks e S 1 and e S 2 cannot be efﬁciently inverted.M. so an additional approximation is performed by replacing them with their incomplete Cholesky factorizations: " e S’ LS1 LTS1 0 0 LS2 LTS2 # ¼ LS LTS ð41Þ . symmetric and indeﬁnite. can be generally increased as the consolidation proceeds. As far as the preconditioner is concerned. In fact. Set c ¼ hDt and w ¼ ð1  hÞ=h. hence it is convenient to drop some of its blocks. divide Eq. (30) by Dt.

s1 ¼ P 1 y s1 z 1  s1 8. Terzaghi’s problem [1] consists of a ﬂuid-saturated column of height L with a constant loading P L on top (Fig. respectively. The repeated solution of Eq. Model veriﬁcation The model is veriﬁed against both Terzaghi’s (1-D) and Mandel’s (2-D) consolidation problems. the analytical solution reads [31. tÞ ¼ 4 p p0 1 X m¼0 " #   1 ð2m þ 1Þ2 p2 ct ð2m þ 1Þpz sin exp 2 2m þ 1 2L 4L ( uðz. 3. 1. Assuming the z-axis positive downward. cM ¼ ðk þ 2lÞ1 the vertical uniaxial compressibility. it is compared to a standard FE model. (38) in place of the exact Schur complement S. (31) starting from the initial conditions (5) gives the discrete ^ . Finally.4818 M. s2 ¼ LT S z2 6. Ferronato et al. two realistic engineering applications are brieﬂy addressed. / Journal of Computational Physics 229 (2010) 4813–4830 e S is used in Eq.32]: pðz.g. A homogeneous sandy column with unit section and L ¼ 15 m is simulated. with the relevant hydraulic and mechanical properties given in Table 2. The load is applied instantaneously at time t ¼ 0 yielding a non-zero initial overpressure p0 ðzÞ and a corresponding settlement u0 ðzÞ. the model is veriﬁed against 1-D and 2-D analytical solutions. and c ¼ k=½qgðM 1 þ a2 cM Þ the consolidation coefﬁcient. 2. 3. [29. The basement is ﬁxed. K u ¼ k þ 2l=3 þ a2 M the undrained bulk modulus.1. . The prescribed distributed load P L is 104 Pa. Then. The preconditioner (42) is used to accelerate the SQMR solver [26] which has proved a robust and efﬁcient algorithm for sparse symmetric indeﬁnite problems. (38) and accounting for (41) yields: " M1 ¼ U 1 L1 ¼ I P 1 HT LT S 0 LT S #" # P1 0 1 L1 S HP L1 S ð42Þ with M1 the block constraint preconditioner (CP). (6)–(8). vectors u. z1 ¼ P 1 r1 t ¼ Hz1 t t  r2 z2 ¼ L1 S t 5. thus providing a factored approximation of A that can be used as preconditioner. First. 1). y ¼ H T s2 7. The time integration is performed with a ﬁrst-order implicit scheme ðh ¼ 1Þ and a constant time step Dt ¼ 0:1 s. The column is discretized into regular tetrahedrals with a characteristic element size h ¼ 0:5 m (Fig. A good matching be- Table 1 CP application to a vector. 1). Drainage is allowed for through the upper moving boundary only. The resulting complete CP-SQMR algorithm for the solution of equations (31) is provided in Appendix A. The simulation proceeds until steady state conditions are attained. The algorithm for applying M1 to a vector r: s ¼ M1 r )  s1 s2 "  ¼ I P1 HT LT S 0 LT S #" P1 0 1 L1 S HP L1 S # r1 r2  ð43Þ is provided in Table 1. p and v through Eqs. Inverting Eq.30]. cbr the solid grain compressibility. e. 4. hence the approximate u 3. tÞ ¼ cM p0 ðL  zÞ  1 8L X p2 m¼0 1 ð2m þ 1Þ2 exp " # ð2m þ 1Þ2 p2 ct 4L2 ð44Þ  ) ð2m þ 1Þpz þ u0 cos 2L ð45Þ where aM PL K u þ 4l=3 1 u0 ðzÞ ¼ PL ðL  zÞ K u þ 4l=3 p0 ðzÞ ¼ ð46Þ ð47Þ with M ¼ ½/b þ ða  /Þcbr 1 the Biot modulus. p and q at the selected values of time. Numerical results The stability and accuracy of the 3-D mixed FE consolidation model previously described is investigated with the aid of a few examples.

0 0. In Terzaghi’s and Mandel’s consolidation tests the sample is made from sand only. The outer surfaces ðx ¼ aÞ are traction-free and drained. Mandel’s problem [33] consists of a poroelastic slab sandwiched between two rigid. a force per unit length in the y-direction. Porous medium ksand (m/s) 105 kclay (m/s) 108 0. as is shown in Figs. respectively.5 t = 3600 s 15.5 m θ = 1.0 0 2000 4000 6000 8000 10000 Pressure [Pa] Fig.5 t = 1800 s 10.0 a tween the analytical and the numerical solution is obtained for both pore pressure and vertical displacement. is suddenly applied at t ¼ 0. The slab is inﬁnitely long in the y-direction with a 2a  2b wide cross-section. The load 2F. 1.0 Analitycal solution Numerical solution 2.0 12. 2 and 3. 4. frictionless.e. i. both loaded by a constant vertical force as shown in Fig. namely [34]: Δ t = 0. Ferronato et al. zÞ. / Journal of Computational Physics 229 (2010) 4813–4830 Fig.5 t = 60 s z [m] 5.0 ðx. zÞ and settlement ux.0 t = 600 s 7. .4819 M. and impermeable plates. zÞ and uz. causing both an instantaneous overpressure p0 ðx. Table 2 Hydraulic and mechanical parameters of the porous medium used for the model veriﬁcation.1 s h = 0. 2.375 / b ðMPa1 Þ k (MPa) l (MPa) 4:4  104 40 40 1.0 ðx. Terzaghi’s consolidation test: analytical and numerical solutions for the pore pressure. Sketch of the setup for Terzaghi’s consolidation test: (a) 1-D physical problem and (b) numerical discretization.

5 15. x ¼ a=4.50e-04 5.0 0.00e+00 2.0 0. Fig.8 0.5 10. . 3. p/p0 1.0e-02 1.0e-01 dimensionless time. and x ¼ a=2.0 0. / Journal of Computational Physics 229 (2010) 4813–4830 Δ t = 0. 1. Terzaghi’s consolidation test: analytical and numerical solutions for the vertical displacement.4 0.00e-04 7.6 Analytical solution at x = 0 Analytical solution at x = a/4 Analytical solution at x = a/2 Numerical solution 0.0 t = 60 s 2.1 s h = 0.5 t = 600 s 5. Ferronato et al. Mandel’s problem: pore pressure history at x ¼ 0.5 m θ = 1.0 1.0 t = 3600 s z [m] t = 1800 s 7.0e+00 2 Fig.25e-03 1. Sketch of the setup for Mandel’s problem: (a) 2-D physical problem after [34] and (b) numerical discretization. 4.00e-03 1.50e-04 1.0 Analytical solution Numerical solution 12.2 0.50e-03 u [m] Fig.2 normalized pressure.4820 M.0e-03 1. 5. ct/a 1.

1 s Δ t = 0. respectively. zÞ ¼ 2l a Fð1  mu Þ y uz.025 s 1. 6. Mandel’s problem: pore pressure proﬁle along the x-axis at various times.4821 M. x/a Fig.2 0 0 0.5 s Analytical solution at t = 5 s Analytical solution at t = 10 s Numerical solution 0. 0 10 .0 pressure error tan an ¼  -4 10 -5 10 10 -2 10 -1 h .2 0. spacing h. 7.4 0.6 0.05 s Analytical solution at t = 0. zÞ ¼  2l a p0 ðx. p/p0 1 Analytical solution at t = 0. tÞ ¼ 2p0 1 X n¼1 m the undrained and drained Pois-  2  sin an an x an ct cos  cos an exp a a2 an  sin an cos an ð51Þ where an are the positive roots of the nonlinear equation: m a mu  m n ð52Þ 1.8 0. z. Ferronato et al.4 0.element size Fig. zÞ ¼ ð48Þ ð49Þ ð50Þ with B ¼ aM=K u Skempton’s coefﬁcient.2 normalized pressure. Terzaghi’s consolidation test: L1 pressure error norm vs.8 0.0 ðx.05 s Δ t = 0. The analytical solution for the pore pressure reads [34]: pðx.6 1 normalized distance. -2 10 Δ t = 0.0 -3 10 1. / Journal of Computational Physics 229 (2010) 4813–4830 1 Bð1 þ mu ÞF 3a F mu x ux.0 ðx. and mu ¼ ½3m þ aBð1  2mÞ=½3  aBð1  2mÞ and son’s ratio.

025 s 1. The pore pressure vs. CP-SQMR convergence proﬁles varying h with Dt ¼ 0:1 s (left) and varying Dt with h ¼ 0:5 m (right). 8. Notice the peculiar non-monotonic pressure behaviour. ranging from h ¼ 1 m to 0. The 2-D problem is solved using a 3-D cube as domain with appropriate boundary conditions to warrant a plane strain condition along the y-direction [35].0 -7 10 -8 10 -2 -1 10 0 10 10 h .0 -6 10 2. i. time and space is provided in Figs. the so-called Mandel–Cryer effect. Ferronato et al. In Fig.element size Fig. then convergence tends to deteriorate as h decreases. / Journal of Computational Physics 229 (2010) 4813–4830 A homogeneous and isotropic slab is simulated using a ¼ b ¼ 1 m and the properties given in Table 2. A regular tetrahedral grid and an implicit ðh ¼ 1Þ integration in time are used. with h ¼ 0:0625 m and Dt ¼ 0:01 s. This agrees with the theoretical results provided in [23] for a 1-D formulation. 5 and 6. The convergence properties of the 3D mixed FE model have been investigated by comparing the analytical solutions (44) and (45) of Terzaghi’s consolidation problem at t ¼ 60 s with the numerical solution obtained on progressively reﬁned computational grids. F is set equal to 104 N=m. The ﬂattening portion of each proﬁle is less pronounced as Dt is reduced.5000 m h = 0. 90 100 110 120 . The domain symmetry allows for limiting the study to a quarter only of the x–z plane (Fig.2500 m h = 0. 7 the L1 -norm of the pressure error is plotted vs. with the lack of convergence for any given Dt due to the time dis- -5 10 displacement error Δt = 0. Δ t = 0.0625 m.1 s Δt = 0.5 m 4 relative residual norm 1×10 -7 -8 -9 -10 0 10 20 30 40 50 60 70 80 iteration Fig. respectively.0625 m 2 1×10 1×10 h = 0. and shows a good agreement between the analytical and the numerical solution. An initially linear behaviour can be noticed.05 s Δt = 0. 4). h in a double log–log plot using a constant Dt value for each proﬁle.e.1 s 1×10 1×10 1×10 3 relative residual norm 1×10 1×10 1×10 0 -1 -3 -4 1×10 1×10 1 -5 1×10 1×10 1×10 1×10 1×10 1×10 -8 1×10 -9 100 150 iteration 200 250 300 1×10 0 -1 -3 -5 -6 1×10 50 -2 Δ t = 10 s -1 Δ t = 10 s 0 Δ t = 10 s 1 Δ t = 10 s 1 -4 1×10 -7 0 2 -2 1×10 1×10 -6 1×10 3 1×10 -2 1×10 1×10 1×10 h = 0. spacing h.4822 M.1250 m h = 0. 9. Terzaghi’s consolidation test: L2 displacement error norm vs. with p rising above the initial p0 at small time values because of the slab contraction at the drained faces inducing an additional build-up in the interior.

/ Journal of Computational Physics 229 (2010) 4813–4830 cretization error. Fig. 0. For example. 11.5 t = 3600 s 10.5 Reference solution FE Numerical solution steady state z [m] 5. . Ferronato et al. 8).0 t = 600 s 12.5 Reference solution Mixed FE Numerical solution steady state z [m] 5. Heterogeneous consolidation test: reference and mixed FE numerical solutions for the pore pressure. the theoretical quadratic convergence [23] is progressively lost as h decreases while Dt is kept constant.0 t = 600 s 12. Consistent with the pore pressure L1 -norm proﬁle. 10. This may suggest that neglecting the off-diagonal blocks of the Schur complement is the most signiﬁcant approximation introduced in the CP algorithm.0 2.5 t = 60 s 15. We observe on passing that using the S 2 in Eq. (41). at the cost of a much larger CPU time spent for the CP computation and application. 0. Heterogeneous consolidation test: reference and standard FE numerical solutions for the pore pressure.5 t = 3600 s 10. The different iteration count indicates that the conditioning of A gets worse as h decreases and Dt increases.0 2. 9 shows the relative residual norm: t rr ¼ kf  AztþDt k2 ð53Þ t kf k2 vs.5 15.0 -10000 t = 60 s -8000 -6000 -4000 -2000 0 2000 Pressure variation [Pa] Fig. in the problem with h ¼ 0:5 m and Dt ¼ 0:1 s convergence is achieved in 88 iterations instead of 113. does not provide a signifexact factorizations of e S 1 and e icant solver acceleration. Similar remarks hold true for the displacements using the L2 -norm of the error (Fig. the number of iterations required by the CP-SQMR algorithm of Appendix A to solve the linear system (31).4823 M.0 7.0 7. which is consistent with the numerical behaviour typically encountered in mixed FE models of subsurface ﬂow [36].0 -10000 -8000 -4000 -6000 -2000 0 Pressure variation [Pa] Fig. in place of the incomplete Cholesky decompositions.

2. The computational grid is the same as in Fig. Fig.0 Reference solution FE Numerical solution Mixed FE Numerical solution 12. / Journal of Computational Physics 229 (2010) 4813–4830 3. 10. with no oscillations at the sand–clay interface.0 7.4824 M.0 2. Heterogeneous consolidation test: reference and numerical solutions for the vertical displacement. 11 provides the pore pressure solution simulated by a standard FE model based on the same mesh and equal order pressure and displacement approximations. By contrast. For the sake of a comparison.0 2. Drainage is allowed for also on the bottom boundary where a pore pressure variation of 9810 Pa is prescribed. The porous column of Fig. the vertical displacement is stable and accurate in both the FE and mixed FE models (Fig. Comparison with standard FEs The stability of the numerical solution obtained with the Mixed FE model is investigated in a heterogeneous test problem. 0. The simulated pore pressure at different times is shown in Fig. The clay permeability is 1000 times smaller than sand (Table 2) while the elastic properties are the same. The numerical solution is close to the reference one and appears to be quite stable.5 t = 3600 s 10. The reference solution has been obtained by progressively reﬁning both the tetrahedral mesh and the time step until convergence. 0. The top boundary is now traction-free.5 10.00e-04 8.0 -10000 t = 60 s -8000 -6000 -4000 -2000 0 2000 Pressure variation [Pa] Fig.00e-04 6.5 t = 3600 s t = 60 s t = 600 s 5.00e-04 4. 1 is now assumed to consist of 10 m of clay on top and 5 m of sand on bottom.0 0. . Heterogeneous consolidation test: reference and standard FE numerical solutions for the pore pressure on a regularly reﬁned tetrahedral mesh so that the total number of unknowns equates that of the mixed FE model. 13. 1.5 Reference solution FE Numerical solution steady state z [m] 5. Ferronato et al. standard FEs exhibit pronounced initial pressure oscillations at the sand–clay interface. 12).00e-03 u [m] Fig. 12. Even in this simple problem.5 15.00e+00 2. The result used as a reference is obtained with h ¼ 0:03125 m and Dt ¼ 0:01 s.00e-04 1.0 z [m] steady state 7.5 15.0 t = 600 s 12.

819 1. respectively.597 1. time due to the application of a surface load. Axonometric view of the FE grid used in the realistic applications.732 1. kxx ¼ kyy (m/s) kzz =kxx / b ðMPa1 Þ k (MPa) l (MPa) a Clay Silt Sand 107 1 0.227 0. To have an idea. respectively. . 15. with an overall system Table 3 Hydro-geological properties of the shallow sediments in the upper Adriatic basin used in the realistic model applications.3 4:32  104 1. pressure and velocity.3 104 0. while the Mixed FE model totals 837.0 4:32  104 2. / Journal of Computational Physics 229 (2010) 4813–4830 The stable pore pressure solution provided by the Mixed FE model is obtained at the cost of a larger number of unknowns. 720 and 1688 unknowns for displacement.0 4:32  104 1.1 0. 14.3 106 1 0. Ferronato et al.4825 M.0 Fig. the standard FE model has 837 and 279 displacement and pressure unknowns. Fig.819 1.227 0. Test case 1: pore pressure variation vs.

4826 M. The following additional boundary conditions apply: the outer boundary is ﬁxed and drained.38].05 m. The load is set equal to 8 kN/m2 and is representative of an artiﬁcial gravel mound similar to that built at the inlets of the Venice 0 day 1 day 2 day 3 day 10 day 20 Δ z [m] -0. 14) with zero ﬂux and horizontal displacement prescribed on the inner boundaries. 3. 356 nodes. The axial symmetry of the model geometry allows for the discretization of one fourth only of the overall porous volume (Fig.04 -0. 14. 368 faces with an overall model size equal to 253. 13). To have practically insigniﬁcant oscillations (650 Pa) the mesh size h at the sand–clay interface must be 60. time.3.37. oscillations in the pore pressure solution do not vanish. Fig. Italy. / Journal of Computational Physics 229 (2010) 4813–4830 size equal to 3245. Test case 2: pore pressure variation 20 days after the start of pumping. a vertically regularly reﬁned grid is used totaling nn ¼ 13.03 -0. 3. if the tetrahedral mesh is regularly reﬁned so that the standard FE system size is three times larger. Ferronato et al.02 -0. the bottom is ﬁxed and impervious. However. 100 .1. As shown in Fig. Test case 1: vertical land displacement vs. although they are less pronounced than before (Fig. 16. the top is traction-free and drained. 080 elements and nf ¼ 143. ne ¼ 70. silty and clayey layers down to 50 m depth is simulated.3.05 0 25 50 75 r [m] Fig. 17. Realistic applications The 3-D mixed FE model has been experimented with in two realistic applications addressing the consolidation of a shallow formation in the geological basin underlying the Venice lagoon.516. A cylindrical stratiﬁed porous volume made of a sequence of alternating sandy. The hydro-geological properties are summarized in Table 3 and are representative of a shallow sedimentary sequence in the upper Adriatic basin [7.01 -0. Test case 1: surface load A uniform surface load distributed over a circular area centered on the domain top with a 10-m radius is applied.

Test case 2: vertical land displacement vs. / Journal of Computational Physics 229 (2010) 4813–4830 0. time on a horizontal plane located in the middle of the upper clay layer.05 -0.00 day 1 day 2 day 3 day 10 day 20 Δ z [m] -0.20 0 50 100 150 200 250 r [m] Fig. Fig. . Ferronato et al. Test case 2: pore pressure variation vs.4827 M.10 -0. time. 18. 19.15 -0.