Advances in Water Resources 29 (2006) 18661875 www.elsevier.

com/locate/advwatres

Modeling groundwater velocity uncertainty in nonstationary composite porous media

Chuen-Fa Ni *, Shu-Guang Li
Department of Civil and Environmental Engineering, Michigan State University, East Lansing, MI 48824, United States Received 12 September 2005; received in revised form 13 January 2006; accepted 13 January 2006 Available online 2 March 2006

Abstract Despite the intensive research over the past decades in the eld of stochastic subsurface hydrology, our ability to analyze and model heterogeneous groundwater systems remains limited. Most existing theories are either too restrictive to handle practical complexity or too expensive to be applied to realistic problem sizes. In this paper we present approximate, closed-form equations that allow modeling 2D nonstationary ows in statistically inhomogeneous aquifers, including composite aquifers containing multiple zones characterized by dierent statistical models. The composite representation has the eect of decreasing the variance of deviations from the mean, relaxing the limitation of the small-perturbation assumption. The simple formulas are illustrated with a number of examples and compared with a corresponding rst-order nonstationary numerical analysis and Monte Carlo simulation. The results show that, despite the gross simplications, the closed-form equations are robust and able to capture complex variance dynamics, reproducing surprisingly well the rst-order numerical solutions and the Monte Carlo simulation even in highly nonstationary, variable situations. 2006 Elsevier Ltd. All rights reserved.
Keywords: Stochastic modeling; Spectral method; Perturbation method; Heterogeneity; Uncertainty modeling; Composite porous media; Monte Carlo simulation

1. Introduction Probabilistic theories of subsurface ow and transport have had a signicant impact on the way we think about uncertainty and heterogeneity. However, they have not had much impact on the way that predictions are generated and reported in practical groundwater modeling studies [39]. One major reason for this signicant gap lies in that most existing stochastic theories require that the aquifer of interest be statistically homogeneous, the hydraulic gradient statistically uniform, and the deviation from the uniform mean small (e.g., [19,31]). It is becoming increasingly clear that if stochastic modeling is to become a viable practical tool, it must be made much more general and exible. Specically, a stochastic model must be able to easily incorporate site-specic aqui*

Corresponding author. E-mail address: nichuenf@egr.msu.edu (C.-F. Ni).

fer structure before it can be routinely applied in practice. It must allow modeling exible zonations, layering, and general trending as most real-world aquifers exhibit not only random variability but also systematic structural variations and the statistics characterizing aquifer heterogeneity can vary (e.g., from region to region and layer to layer) in response to systematic changes in the distribution of aquifer materials. The research by Graham and Mclaughlin [5], Li and McLaughlin [13], and Neuman and Orr [20] represents rst steps in this direction. Loaiciga et al. [16], Gelhar [4], Li and McLaughlin [14], Rubin [25], Indelman and Rubin [9], Indelman and Zlotnik [10], and Ni and Li [21] later investigated ows in heterogeneous trending media and presented explicit, special analytical solutions in illustration of the eect of nonstationarity. More recently, Winter and Tartakovsky [32], Lu and Zhang [17], Rubin [26], and Guadagnini et al. [7] derived closed-form solution for ow in special composite media that consist of multiple zones

0309-1708/$ - see front matter 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.advwatres.2006.01.003

C.-F. Ni, S.-G. Li / Advances in Water Resources 29 (2006) 18661875

1867

with nonstationary uctuation statistics. These illustrative solutions provided useful insight into how large-scale nonuniformity and zonations interact with small-scale heterogeneity, although they must be generalized before they can actually be used in practice. There are a number of numerical approaches that can be used to analyze statistically nonuniform ow and transport in more general, statistically inhomogeneous aquifers. These include, for example, Monte Carlo methods (e.g., [28,29,20,15]) and perturbation techniques, such as the moment equation methods (e.g., [5,38,33,17,34,37]) and the so called rst-order second-moment methods based on Taylors expansions (e.g., [30,35,2]). All of these methods, however, are computationally demanding when applied to ow and transport problems of realistic size, mostly because they all require solving large numbers of partial dierential equations on very ne spatial discretizations in order to predict the impact of the small-scale heterogeneous dynamics (e.g., [5,4,15,11,12]). In this paper we present a general, approximate methodology for modeling two-dimensional groundwater systems in complex, composite media. In particular, we apply the popular spectral technique [4] to derive a set of closed-form formulas for predicting nonstationary velocity variances in aquifers that consist of complex zonations characterized by dierent statistical models. These formulas provide nonstationary predictive capabilities of a stochastic numerical model while retaining the convenience of analytical solutions. We demonstrate the eectiveness of the simple closed-form formulas using a number of examples involving complex material zonations and nonstationary conductivity statistics. 2. Problem formulation We can make our presentation and discussion more specic by considering a relatively simple problem: steady-state simulation of hydraulic heads and groundwater velocities in a saturated region. Here we assume that the region is composed of multiple subregions containing systematically dierent materials characterized by dierent statistical models. We further assume that uctuation in each subregion is locally isotropic and statistically uniform but their statistics can vary from region to region and the trends vary more smoothly than the uctuations. In practice, the trend can be estimated from a sample conductivity function by applying a low-pass or moving window lter [24]. Model-based groundwater velocity predictions form the basis for travel time, capture zone, wellhead protection, and solute transport analyses. Uncertainties in these velocities can be substantial if the hydraulic conductivity is heterogeneous, even when the piezometric head is smooth. In practical groundwater modeling applications it is important to have some feeling for the magnitude of velocity uncertainty. The random log conductivity uctuation is approximately related to the uctuations in piezometric head and

Darcy velocity by the following rst-order, mean-removed ow equations: o 2 h0 oh0 of 0 li x J i x ; x 2 D; oxi oxi oxi ox   i oh0 u0i K g x J i xf 0 ; x 2 D; oxi h0 x 0; x 2 CD ; rh0 x nx 0; x 2 CN ;  where J i x oh=oxi is the deterministic mean head gra dient, li x of =oxi is the mean log conductivity gradient, and Kg(x) is the geometric mean conductivity. Both li, Kg are in general spatially variable and may be discontinuous across internal subregion boundaries. These equations are written in Cartesian coordinates, with the vector location symbolized by x and summation implied over repeated indices. The point values of the log hydraulic conductivity uctuation f 0 (x), head uctuation h 0 (x), and Darcy velocity uctuation u0i x; i 1; 2; 3, are dened throughout the domain D. A homogeneous condition is dened on the specied head boundary CD and the specied ux boundary CN. Note that we consider f is solely the source of uncertainty applied in the aquifer system and is locally (within an individual subregion) stationary and globally (across dierent regions) nonstationary. 3. Spectral solution in composite media Spectral methods oer a particularly convenient way to derive groundwater velocity statistics from linearized uctuation equations such as (1) and (2) [1,4,1315,11,12]. Invoking the spectral representation in each subregion for the locally stationary log conductivity perturbation gives Z 1 0 f x expk i xi dZ f k; x; 3
1

1 2

where 11=2 , ki is component i of the wave number k, and dZf(k, x) is the random Fourier increment of f 0 (x), evaluated at k. The x dependence in the spectral amplitude reects the fact that it is in general globally nonstationary and can vary from region to region. The Fourier representation can be viewed as the continuous version of a Fourier series expansion of f(x). The random Fourier increment at a particular wave number is analogous to the random amplitude of one of the terms in the Fourier integral. Stationary Fourier increments within each region satisfy the following orthogonality property [23,4]: dZ f k; x dZ k0 ; x S ff k; xdk k0 dk dk0 ; f 4

where the asterisk superscript represents the complex conjugate, d() is the Dirac delta function, and S(k, x) is the spectral density function of the log hydraulic conductivity [23,4]. Papoulis [22] shows that the output (e.g., h 0 , u0i ) of linear transformations such as (1) and (2) are stationary only if

1868

C.-F. Ni, S.-G. Li / Advances in Water Resources 29 (2006) 18661875

the input (e.g., f 0 ) is stationary and the transformations are spatially invariant. In the problem of interest here, spatial invariance implies that the uctuation Eqs. (1) and (2) should have constant coecients with the boundaries suciently distant having no eect on velocity uctuations in the region of interest. Such spatial invariance requirement is clearly not met because, for heterogeneous composite media, the coecients li(x) and Ji(x) may both vary with x (e.g., [5,16,27,20,8,14,32,7]). Like many investigators [1,18,4], here we seek approximate solutions to (1) and (2). Taking advantage of the scale disparity between the uctuation process and the mean process, we further assume that the input log-conductivity, the dependent head, and velocity output uctuation be stationary in a local sense (and away from the immediate proximity of boundaries), so that they also have an approximate spectral representation dened in terms of a locationdependent spectral amplitude Z 1 0 h x expk i xi dZ h k; x; 1 Z 1 expk i xi dZ uj k; x: 5 u0j x
1

Integrating in the wave number domain, one obtains the following integral expression for evaluating the velocity variance: r2i x C ui lr2 xK 2 xJ 2 x; u f g where ! kik1 C ui l 1 2 k k j lj x 1 1 ! kik1 1 2 sff k; x dk 1 dk 2 : k k j lj x Z
1

10

Note s is dimensionless spectral density function, sff S ff =r2 . The symbols r21 and r22 represent respectively f u u the longitudinal and transverse velocity variances. 3.1. Approximate spectral method To obtain explicit results, one must in general evaluate the associated double integrals numerically. In most cases, these evaluations can be quite dicult. Most prior research focuses on some very special cases for which (10) can be reduced to a form that allows exact, closed-form integration [3,4,36,26]. In this paper, we evaluate (10) approximately under more general conditions. Our approximate solution is based on the following observations: Trends in conductivity inuence the variance dynamics through C ui l and Kg(x)J(x) (see (9)). It is the evaluation of C ui for a general, multi-dimensional trend distribution li(x) that is dicult. More specically, it is the presence of the k j lj x term in (10) that makes the integration analytically intractable. It is, however, predominantly Kg(x)J(x) that controls the nonstationary spatial variance dynamics. For most trending situations, change in the mean log conductivity over the characteristic length of small-scale heterogeneity (a correlation scale k) is small (or l2 k2 ( 1) since the mean conductivity is expected to j be much smoother than the uctuation [4]. To enable variance modeling under general, complex conditions, we propose approximating C ui l via the following Taylors expansion-based expression: C ui l C ui 0 oC ui 0 lj % C ui 0: olj 11

This local homogeneity assumption implies that we regard li(x) and Ji(x) in (1) and (2) as varying slowly in space relative to the characteristic scale of the uctuation, that is, that they do not change signicantly over distance corresponding to the correlation scale of h. Note that l1 will i be a typical length scale for change in the mean gradient, so that the notion of local statistical homogeneity will be meaningful when the product of li and the correlation scale is small relative to 1 [4]. By using the local spectral representations (5) in (1), the spectral amplitude for head in two-dimensional problems is dZ h k; x k i J i x dZ f k; x; k 2 k j lj x k2 k2 k2: 1 2 6

Substituting (5) and (6) in (2), we obtain the following spectral amplitude of Darcy velocity: ! kikj dZ ui k; x K g xJ j x dij 2 dZ f k; x; k k j lj x 7 where dij is the Kronecker delta function. Now, if the x1 coordinate is selected to be aligned with the local mean hydraulic gradient (J1(x) = J(x), J2(x) = 0), then the velocity spectral density functions in each material zone becomes S ui uj k; x K 2 xJ 2 x g    kik1 kjk1 1 2 1 2 k k m lm x k k m lm x S ff k; x: 8

Essentially we suggest that the small, but hard-to-evaluate contribution to variance nonstationarity from C ui be ignored relative to the much more important contribution from Kg(x)J(x). This assumption may seem quite crude but proves to be highly eective and makes general, approximate variance modeling in nonstationary media possible. Previous studies investigating the eects of trending are all based on full integration of (9) and (10) or full solution of (1) that is intractable unless the trends are

C.-F. Ni, S.-G. Li / Advances in Water Resources 29 (2006) 18661875

1869

assumed to be of special forms [9]. These highly restrictive assumptions severely limit the practical utilities of the results. Substituting (11) into (9), we obtain r2i x r2 xK 2 xJ 2 x u f g 2 Z 1 Z 1  kik1 1 2 sff k; x dk 1 dk 2 : k 1 1

12

Eq. (12) can be easily integrated in the polar coordinate system. For the statistically isotropic case, the result is the following simple, explicit expressions, independent of the specic form of the ln K spectrum: r21 x 0:375r2 xK 2 xJ 2 x; u f g r22 x u 0:125r2 xK 2 xJ 2 x: f g 13 14

Note these expressions are of the same form as those derived by many previous researchers (e.g., [3,4,6]) for statistically uniform ow, except that Kg and J are now allowed to vary over space as a function of the trending and r2 can f vary from zone to zone. In the following section, we illustrate how these simple analytical expressions can be used to quantify robustly and quite accurately groundwater velocity uncertainty in complex, nonstationary hydraulic conductivity elds. 4. Illustration examples To systematically test the eectiveness, robustness, and generality of the closed-form solutions, we consider a range of nonstationary, composite media congurations and use dierent statistical models to represent the small-scale heterogeneity. Our rst example considers a relatively simple situation in which the overall domain of interest contains two embedded, rectangular areas of distinctly dierent materials. Our second example is patterned after a real situation involving more complex zonations of irregular shapes delineated based on a map of surcial deposits. Our third example considers fan-shaped deposits of water-transported material (alluvium) that forms at the
Table 1 Parameter denitions for three examples Example 1 ln K correlation structure ln K variance ln K correlation scale k (m) Mean conductivity Kg (m/day) Domain length (m) Grid number (ASM and NSM) Grid number (MCS) Global recharge (m/day) West boundary condition East boundary condition North boundary condition South boundary condition Exponential/hole 0.5, 1.0, and 2.0 0.5, 1.0, and 2.0 5, 25, and 125 (see Fig. 2a) 80 80 81 81 256 256 No recharge Const. Head 100 m Const. Head 99.2 m No ow No ow

base of fractured mountain blocks where there is a marked break in slope. The alluvial fan is coarse-grained and very permeable at the mouth and becomes gradually nergrained towards the edge as it meets with a large surface water body. In all three examples, the overall conductivity distributions are strongly nonstationary both in the mean and uctuation. The mean conductivity varies within the zones (Example 3) and between them (Examples 13) and these large-scale nonstationarities are represented as deterministic trends. The uctuation statistics in each zone are characterized by an independent statistical model with a zone-dependent variance and correlation scale. It is our opinion that if a methodology is able to predict the groundwater velocity uncertainty in such dierent composite media congurations, it should be able to handle perhaps most situations that can be realistically represented using eld data in real-world groundwater modeling. Table 1 presents detailed information dening the aquifer parameters, boundary conditions, and other inputs used in the three examples. Fig. 1ac present realizations of the conductivity distributions along the domain center line for Example 13, respectively. Fig. 2ac illustrate the distributions of the mean conductivity, prescribed stresses, and corresponding steady-state mean head. The strong trends and irregular zonations in the mean log hydraulic conductivities yield nonuniform ow patterns which in turn cause strong nonstationarity in the velocity variances. To compute the velocity variances and demonstrate their accuracy, we follow the following three step procedure. First, we solve the mean deterministic groundwater ow equation without accounting for the small-scale heterogeneity. We then use the computed head to evaluate the mean hydraulic gradient and substitute it into (13) and (14) to obtain the local variance values. Finally, we compare the closed-form solutions with the corresponding numerical solutions obtained from the rst-order nonstationary spectral method [13,14,11,12] and Monte Carlo simulation (based on 10,000 realizations). The nonstationary spectral method is based on a generalized, nonstationary spectral

Example 2 Exponential/hole 0.5, 0.8, 1.0, and 1.5 0.5, 0.8, 1.0, and 1.5 5, 25, 50, and 100 (see Fig. 2b) 160 160 161 161 512 512 No recharge Const. Head 100 m Const. Head 98.4 m No ow No ow

Example 3 Exponential/hole 1.0 and 2.0 5.0 and 10.0 2.0 and kriged K eld (see Fig. 2c) 2000 1500 251 186 1024 768 0.002 No ow No ow No ow No ow

ASM: approximate spectral method; NSM: nonstationary spectral method; MCS: Monte Carlo simulation.

1870
8.0 7.0 6.0 5.0

C.-F. Ni, S.-G. Li / Advances in Water Resources 29 (2006) 18661875

Mean Mean + fluctuations

ln k

4.0 3.0 2.0 1.0 0.0 0 40 60 80

(a)
8.0 7.0 6.0 5.0

20

x 1 (m)

Mean Mean + fluctuations

ln k

4.0 3.0 2.0 1.0 0.0 0 20 40 60 80 100 120 140 160

(b)
6.0 5.0 4.0 3.0

x1 (m)

Mean Mean + fluctuations

ln k

2.0 1.0 0.0 -1.0 -2.0 0 500 1000 1500 2000

(c)

x 1 (m)

Fig. 1. A realization of nonstationary conductivity distribution plotted along the domain centerline in the horizontal direction for (a) Example 1, (b) Example 2, (c) Example 3.

representation theorem and solves numerically the full version of the linearized perturbation equation. Without having to make the local stationarity assumption and dropping the secondary terms (in the head perturbation equation), the nonstationary spectral approach allows accommodating exactly boundary conditions, spatially variable mean gradients, and other sources of nonstationarity [1315,11,12]. 5. Results and discussion Figs. 35 present the comparative results for all three examples. Fig. 3 presents contour maps showing the complex, nonuniform distributions of the predicted velocity standard deviations. The results are obtained using the approximate spectral method (left column), the numerical

nonstationary spectral method (middle column), and the Monte Carlo simulation (right column) based on the simple exponential covariance model. Fig. 4 shows the same results in a prole along the domain centerline. The results clearly show that, despite the simplications and the strong multi-dimensional medium nonstationarities, the simple closed-form solutions reproduce well the corresponding rst-order, nonstationary spectral solutions and allow capturing both the spatial structure (see Fig. 3) and the magnitude (see Fig. 4) of the highly nonstationary, complex uncertainty distributions. The closed-form predictions of the velocity uncertainty also match very well with those obtained from the Monte Carlo simulation for three examples. Fig. 5 presents similar comparisons between the closedform and numerical solutions based on an alternative ln K

C.-F. Ni, S.-G. Li / Advances in Water Resources 29 (2006) 18661875

1871

Fig. 2. Spatial distribution of the geometric mean conductivity and the corresponding mean head distribution for (a) Example 1, (b) Example 2, (c) Example 3.

covariance modelthe hole exponential covariance model. Although the hole-type model looks similar to the simple exponential model in the physical space domain, it is very dierent in the frequency domain at low wave numbers. Gelhar [4] suggested that a hole-exponential model may represent the eld data better than the simple exponential model. The results show that the closed-form solutions based on this hole model also match very well for all three examples with the corresponding numerical perturbation and Monte Carlo solutions. The surprisingly robust performance of the closed-form solutions for various highly nonstationary situations involving dierent media zonations, mean conductivity distributions, ow and uncertainty patterns, boundary conditions, and conductivity covariance models suggests that the seemingly crude simplifying assumptions and empirical observations made in deriving the analytical formulas are highly eective. These simple assumptions can indeed capture the dominant factors controlling the spatial uncertainty distributions and make it possible to model velocity uncertainty in complex, nonstationary groundwater systems. The closed-form formulas do introduce errors near media interfaces as they make no explicit use of boundary conditions. The solutions become inaccurate near the zone

discontinuities, especially at the prescribed head boundaries (see Figs. 35). However, these errors are limited to approximately within 2 or 3 log-conductivity correlation scales of the discontinuities. The inaccurate regions are considered localized since for stochastic modeling to apply the domain size must be much larger than the size of heterogeneity. We feel these errors are especially easy to accept considering what we gain from the simplifying assumptionsthe enormous exibility and eciency in our ability to model complex, composite systems. It should also be pointed out that our ability to model composite media has the eect of decreasing signicantly the variance of deviations from the mean. This is why our rst-order solutions match so well with the Monte Carlo simulations even under highly variable conditions. Large-scale changes in log conductivity that increase the variance around a constant mean are now treated as nonstationarities (e.g., trends) since the method does not require that the log conductivity mean be constant. Therefore, the small-perturbation assumption becomes much less limiting an assumption as in other stationarity-based perturbation methods. The methods requirement that all uncertainty must ultimately be related to stationary random uctuations in each

1872

C.-F. Ni, S.-G. Li / Advances in Water Resources 29 (2006) 18661875

Fig. 3. Spatial distribution of the predicted velocity variances using the closed-form formulas, nonstationary spectral method, and Monte Carlo simulation based on the simple exponential covariance model: (a) Example 1, (b) Example 2, (c) Example 3.

medium component (zone) may still seem to be a signicant limitation. In reality, this requirement reects a very important tacit assumption of stochastic groundwater hydrology. To be specic, suppose we wish to know how predicted velocity uncertainty is aected by spatial variations in hydraulic conductivity. A number of probabilistic

methods, including the one described here may be used to infer the (possibly nonstationary) ensemble statistics of velocity from the ensemble statistics of hydraulic conductivity. In order to apply any stochastic methods to practical problems, we need to obtain zone specic estimates of the conductivity statistics which form the basis for our stochas-

C.-F. Ni, S.-G. Li / Advances in Water Resources 29 (2006) 18661875

1873

Fig. 3 (continued)

0.4 0.3

Approximate Spectral Method Nonstationary Spectral Method Monte Carlo Method Nr=10000

0.6 0.5 0.4

Approximate Spectral Method Nonstationary Spectral Method Monte Carlo Method Nr=10000

u1

0.2 0.1

u1

0.3 0.2 0.1

0.0 0.4 0.3

Approximate Spectral Method Nonstationary Spectral Method Monte Carlo Method Nr=10000

0.0 0.4 0.3 Approximate Spectral Method Nonstationary Spectral Method Monte Carlo Method Nr=10000

u2

u2

0.2 0.1 0.0

0.2 0.1 0.0

(a)

20

40

60 0.12 0.10 0.08

80

x1 (m)

(b)

20

40

60

80

100

120

140

160

x1 (m)

Approximate Spectral Method Nonstationary Spectral Method Monte Carlo Simulation Nr=10000

u1

0.06 0.04 0.02 0.00 0.12 0.10 0.08 Approximate Spectral Method Nonstationary Spectral Method Monte Carlo Simulation Nr=10000

u2

0.06 0.04 0.02 0.00 0 500 1000 1500 2000

(c)

x1 (m)

Fig. 4. Centerline prole of the predicted velocity standard deviations using the closed-form formulas, numerical spectral method, and Monte Carlo simulation based on the simple exponential covariance model: (a) Example 1, (b) Example 2, (c) Example 3.

1874
0.4 0.3

C.-F. Ni, S.-G. Li / Advances in Water Resources 29 (2006) 18661875

0.6 0.5 0.4

Approximate Spectral Method Nonstationary Spectral Method Monte Carlo Method Nr=10000

Approximate Spectral Method Nonstationary Spectral Method Monte Carlo Method Nr=10000

u1

0.2 0.1 0.0 0.4 0.3

Approximate Spectral Method Nonstationary Spectral Method Monte Carlo Method Nr=10000

u1

0.3 0.2 0.1 0.0 0.4 0.3 Approximate Spectral Method Nonstationary Spectral Method Monte Carlo Method Nr=10000

u2

u2

0.2 0.1 0.0

0.2 0.1 0.0

20

40

60

80

20

40

60

80

100

120

140

160

(a)

x1 (m)
0.12 0.10 0.08

(b)
Approximate Spectral Method Nonstationary Spectral Method Monte Carlo Simulation Nr=10000

x1 (m)

u1

0.06 0.04 0.02 0.00 0.12 0.10 0.08 Approximate Spectral Method Nonstationary Spectral Method Monte Carlo Simulation Nr=10000

u2

0.06 0.04 0.02 0.00 0 500 1000 1500 2000

(c)

x1 (m)

Fig. 5. Centerline prole of the predicted velocity standard deviations using the closed-form formulas, numerical spectral method, and Monte Carlo simulation based on a hole-type covariance model: (a) Example 1, (b) Example 2, (c) Example 3.

tic analysis. In practice, these statistics are typically derived from a limited number of eld measurements available in a zone. In most situations the only type of parameter nonstationarity that we can hope to infer from eld data is a large-scale trend in the local mean (both within the zones or between them). The closed-form formulas can accommodate such trends since they require that the mean removed parameter within a zone be stationary. This suggests that the methods requirement that uncertainty be related to stationary random parameters within a zone is really not a practical limitation. We need to do this any way if we want our composite media analysis to rely on data gathered in the eld. 6. Conclusions In this paper we have developed and demonstrated closed-form formulas to predict velocity variances for

two-dimensional ow in complex composite media. Three examples were used to illustrate the approximate methodology. The results reveal that, despite the gross simplications, the analytical expressions are highly eective and robust and reproduce surprisingly well the solutions of corresponding rst-order and Monte Carlo predictions. The results also show how the nonuniform log conductivity structure changes signicantly the spatial distributions of the velocity variances. In summary, the approximate spectral method makes it possible to model the velocity uncertainty in complex composite media. The analysis represents a step closer to our ultimate goal to include a systematic uncertainty analysis as a part of routine groundwater modeling. We are currently in the process of extending the approximate methodology to model velocity variances and velocity covariances for general unsteady ow in conned/unconned and anisotropic aquifers in the presence of a variety

C.-F. Ni, S.-G. Li / Advances in Water Resources 29 (2006) 18661875

1875

of complex sources and sinks (e.g., streams, lakes, recharge, and wells). References
[1] Bakr AA, Gelhar LW, Gutjahr AL, MacMillan JR. Stochastic analysis of spatial variability in subsurface ows: 1. A comparison of one- and three-dimensional ows. Water Resour Res 1978;14(2): 26371. [2] Cirpka OA, Nowak W. First-order variance of travel time in nonstationary formations. Water Resour Res 2004;40(3). doi:10.1029/2003WR002851. [3] Dagan G. Flow and transport in porous formations. Berlin: Springer-Verlag; 1989. [4] Gelhar LW. Stochastic subsurface hydrology. Englewood Clis, NJ: Prentice-Hall; 1993. [5] Graham W, Mclaughlin D. Stochastic analysis of non-stationary subsurface solute transport. 1. Unconditional moments. Water Resour Res 1989;25(2):21532. [6] Graham WD, Tankersley C. Optimal estimation of spatially-variable recharge and transmissivity elds under steady-state groundwater ow: 1. Theory. J Hydrol 1994;157(14):24766. [7] Guadagnini A, Tartakovsky DM, Winter CL. Random domain decomposition for ow in heterogeneous stratied aquifers. Stoch Environ Res Risk Assess 2003;17:394407. [8] Indelman P, Abramovich B. Nonlocal properties of nonuniform averaged ows in heterogeneous media. Water Resour Res 1994;30(12):338594. [9] Indelman P, Rubin Y. Flow in heterogeneous media displaying a linear trend in the log conductivity. Water Resour Res 1995;31(5): 125765. [10] Indelman P, Zlotnik V. Average steady nonuniform ow in stratied formation. Water Resour Res 1997;33(5):92734. [11] Li SG, Liao HS, Ni CF. A computationally practical approach for modeling complex mean ows in mildly heterogeneous media. Water Resour Res 2004;40(12). [12] Li SG, Liao HS, Ni CF. Stochastic modeling of complex nonstationary groundwater systems. Adv Water Resour 2004;27(11):1087104. [13] Li SG, McLaughlin DB. A non-stationary spectral method for solving stochastic groundwater problems: Unconditional analysis. Water Resour Res 1991;27(7):1598605. [14] Li SG, McLaughlin DB. Using the nonstationary spectral method to analyze ow through heterogeneous trending media. Water Resour Res 1995;31(3):54151. [15] Li SG, Mclaughlin DB, Liao HS. A computationally practical method for stochastic groundwater modeling. Adv Water Resour 2003;26(11):113748. [16] Loaiciga HA, Leipnik RB, Hubak PF, Marino MA. Stochastic groundwater ow analysis in heterogeneous hydraulic conductivity elds. Math Geol 1993;25(2):16176. [17] Lu ZM, Zhang DX. On stochastic modeling of ow in multimodal heterogeneous formations. Water Resour Res 2002;38:1190. doi:10.1029/2001WR001026. [18] Mizell SA, Gutjahr AL, Gelhar LW. Stochastic analysis of spatial variability in two-dimensional steady groundwater ow assuming

[19] [20]

[21]

[22] [23] [24] [25] [26] [27] [28]

[29]

[30]

[31] [32]

[33]

[34]

[35]

[36] [37]

[38]

[39]

stationary and nonstationary heads. Water Resour Res 1982;18(4):105367. Neuman SP. Stochastic groundwater in practice. Stoch Environ Res Risk Assess 2004;18(4):26870. Neuman SP, Orr S. Prediction of steady state ow in nonuniform geologic media by conditional moments: Exact nonlocal formalism, eective conductivities, and weak approximation. Water Resour Res 1993;29(2):34164. Ni CF, Li SG. Simple closed form formulas for predicting groundwater ow model uncertainty in complex, hetereogeneous trending media. Water Resour Res 2005;41(11). Papoulis A. Probability, random variables, and stochastic processes. New York, NY: McGraw-Hill; 1984. Priestley MB. Spectral analysis and time series. Multivariate series, prediction and control, Part B, vol. 2. Academic Press; 1981. Rajaram H, Mclaughlin D, Identication of large-scale spatial trends in hydrologic data. 1990. Rubin Y. Flow and transport in binormal heterogeneous formations. Water Resour Res 1995;31(10):24618. Rubin Y. Applied stochastic hydrogeology. Oxford University Press; 2003. Rubin Y, Bellin A. The eects of recharge on ow nonuniformity and macrodispersion. Water Resour Res 1994;30(4):93948. Smith L, Freeze RA. Stochastic analysis of steady state groundwater ow in a bounded domain. 1: One-dimensional simulations. Water Resour Res 1979;15(3):5218. Smith L, Freeze RA. Stochastic analysis of steady state groundwater ow in a bounded domain: 2. Two-dimensional simulations. Water Resour Res 1979;15(6):154359. Sun NZ, Yeh WWG. A stochastic inverse solution for transient groundwater ow: parameter identication and reliability analysis. Water Resour Res 1992;28(12):326980. Winter CL. Stochastic hydrology: practical alternatives exist. Stoch Environ Res Risk Asses 2004;18(4):2713. Winter CL, Tartakovsky DM. Groundwater ow in heterogeneous composite aquifers. Water Resour Res 2002;38(8). doi:10.1029/ 2001WR000450. Winter CL, Tartakovsky DM, Guadagnini A. Numerical solutions of moment equations for ow in heterogeneous composite aquifer. Water Resour Res 2002;38(5). doi:10.1029/2001WR000222. Winter CL, Tartakovsky DM, Guadagnini A. Moment dierential equations for ow in highly heterogeneous porous media. Surveys Geophys 2003;24:81106. Yeh T-C, Jin MH, Hanna S. An iterative stochastic inverse method: conditional eective transmissivity and hydraulic head elds. Water Resour Res 1996;32(1):8592. Zhang D. Stochastic methods for ow in porous media, coping with uncertainties. New York, USA: Academic Press; 2002. Zhang D, Lu Z. An ecient, high-order perturbation approach for ow in a random porous media via karhunen-loeve and polynomial expansions. J Comput Phys 2004(194):77394. Zhang DX, Winter CL. Moment-equation approach to single phase uid ow in heterogeneous reservoirs. SPE J 1999;4(2): 11827. Zhang YK, Zhang D. Forum: The state of stochastic hydrology. Stoch Environ Res Risk Assess 2004;18(4):265.