Factor de Impacto:
4.911
Factor de Impacto de 5 años:
3.179
SJR:
1.008
SNIP:
0.983
CiteScore™:
5.2
ISSN Imprimir: 21525080
Acceso abierto
Volumes:

International Journal for Uncertainty Quantification
DOI: 10.1615/Int.J.UncertaintyQuantification.v1.i1.40
pages 4976 ASSIMILATION OF COARSESCALE DATA USING THE ENSEMBLE KALMAN FILTERAbstract Reservoir data is usually scale dependent and exhibits multiscale features. In this paper we use the ensemble Kalman filter (EnKF) to integrate data at different spatial scales for estimating reservoir finescale characteristics. Relationships between the various scales is modeled via upscaling techniques. We propose two versions of the EnKF to assimilate the multiscale data, (i) where all the data are assimilated together and (ii) the data are assimilated sequentially in batches. Ensemble members obtained after assimilating one set of data are used as a prior to assimilate the next set of data. Both of these versions are easily implementable with any other upscaling which links the fine to the coarse scales. The numerical results with different methods are presented in a twin experiment setup using a twodimensional, twophase (oil and water) flow model. Results are shown with coarsescale permeability and coarsescale saturation data. They indicate that additional data provides better finescale estimates and fractional flow predictions. We observed that the two versions of the EnKF differed in their estimates when coarsescale permeability is provided, whereas their results are similar when coarsescale saturation is used. This behavior is thought to be due to the nonlinearity of the upscaling operator in the case of the former data. We also tested our procedures with various precisions of the coarsescale data to account for the inexact relationship between the fine and coarse scale data. As expected, the results show that higher precision in the coarsescale data yielded improved estimates. With better coarsescale modeling and inversion techniques as more data at multiple coarse scales is made available, the proposed modification to the EnKF could be relevant in future studies. KEYWORDS: Kalman filter, reservoir engineering, spatial uncertainty, multiscale estimation, parameter estimation
1. IntroductionBroadly speaking, the measured data used for description of reservoir porosity and permeability characterization consist of static and dynamic data. Static data such as well logs and core samples can resolve heterogeneity at a scale of a few inches or feet with high reliability. However, dynamic data such as fractional flow or water cut (neglecting any preexisting mobile water in the reservoir, this could be defined as the ratio of the injection fluid to the total fluid produced at the production wells), pressure transient, and tracer test data typically scan the length scales comparable to the interwell distances. Additional dynamic data such as timelapse seismic images [1] can provide improved spatial sampling but at a lower precision. The ensemble Kalman filter (EnKF) is now being used in a number of studies for reservoir history matching. Some of the recent studies are listed in Evensen [2]; also see Nævdal et al. [3], Wen and Chen [4], Gu and Oliver [5], and Jafarpour and McLaughlin [6]. In general, reservoir data is often scaledependent and exhibits multiscale features, and integration of additional multiscale data could further reduce the uncertainty (see Lee et al. [7], Efendiev et al. [8, 9] and references therein). Also, it is important to resolve finescale heterogeneity for various purposes such as enhanced oil recovery, environmental remediation, etc. With that perspective, integration of data at coarse and fine scales is an important objective. Computationally efficient assimilation of multiscale data using EnKF to estimate finescale fields for subsurface characterization is the main topic of this study. The main reason we used EnKF in this paper is because it requires fewer ensemble members than the particle filters (where, rather than updating the ensemble members model state, we update the probability assigned to each ensemble member based on model data misfit), e.g., see [10] and references therein for further details. In this paper, apart from the watercut data, we consider two kinds of coarsescale measured data as well. The coarsescale data are assumed to be permeability and/or saturation at some specified level of precision. The unknown variables (permeability, at the fine scale), are estimated using a modification to the EnKF algorithm, linking the data at different scales via upscaling (from the finest to the coarsest scales). The main idea behind upscaling is to obtain an effective coarsescale permeability which yields the same average response as that of the underlying finescale field, locally. First we consider coarsescale permeability data, which could be obtained either from geological considerations or coarsescale inversion of dynamic, fractional flow data on a coarse grid [7, 9] or also using Markov Chain Monte Carlo (MCMC) techniques [8]. This coarsescale, static data could be viewed as prior information regarding the permeability or in other words, a constraint which is to be satisfied up to the prescribed variance while obtaining the finescale estimates in every data assimilation cycle using the EnKF. Upscaling methods relate the solution at the fine scale to the coarse scale; therefore, in the Kalman filtering context, it amounts to modeling a nonlinear observation operator. In this paper we study two ways to assimilate the coarsescale data using the EnKF. The standard EnKF [2] could be used for assimilating all the available data in one step, or alternatively, the measured data could be used in batches. For example, the estimate with one data becomes a prior while assimilating the other measured data; further details are given in Section 3. The second kind of coarsescale observed data we consider is dynamic and is motivated based on the increasing availability of timelapse seismic images (or 4d seismic data). Integration of inverted 4d seismic data (at fine scale) using the EnKF has been addressed in Dong et al. [11] and Skjervheim et al. [12]. In this article we consider the seismic data, not to correspond to the finest scale but to a coarse scale, since timelapse seismic data typically have a lower spatial resolution compared to the finescale geologic models [13]. Since the timelapse seismic data is collected only at specific time intervals, we used coarsescale fluid saturation as measured data to be available at a prescribed level of precision (which accounts for the inaccuracies involved in inversion of 4d seismic data) and only for certain assimilation cycles. Therefore, unlike the coarsescale static permeability data considered earlier, the coarsescale saturation data is assimilated only in certain assimilation cycles (see Section 4.3 for details). Following is the plan of this paper. For the paper to be selfcontained and for notational clarity, we briefly review the governing equations and sequential data assimilation using the EnKF in Section 2. This is followed by a description of the EnKF for assimilation of coarsescale data in Section 3. For our numerical results in Section 4, we consider a fivespot pattern, with the injection well placed in the middle of a rectangular domain and four production wells located at the vertices of the rectangle. A reference case is used to provide true data, which is randomly perturbed to obtain synthetic measurements in a twin experiment setup. After presenting the assimilation results with both coarsescale permeability and saturation data, we conclude with some directions for future work in Section 5. 2. Preliminaries2.1 FineScale ModelIn this paper we consider twophase flow in a subsurface formation under the assumption that the displacement is dominated by viscous effects. For simplicity, we neglect the effects of gravity, compressibility, and capillary pressure, although our proposed approach is independent of the choice of physical mechanisms. Also, porosity is considered constant. The two phases are referred to as water and oil, designated by subscripts w and o, respectively. We write Darcy’s law for each phase as follows:
The above descriptions are henceforth referred to as the finescale model of the twophase flow problem. Here κ_{f} is the (finescale) permeability of the medium, λ(S) is the total mobility, μ_{j} denotes phase viscosity, pr is the pressure, h is the source term, and ϕ and S denote porosity and water saturation (volume fraction), respectively. 2.2 Sequential Estimation Using EnKFUsing dynamic measured data such as water cut, we can sequentially estimate the unknown parameters (permeabi lity, porosity, etc.) and state variables such as pressure, water saturation (twophase flow) and production data at well locations using the EnKF [3, 5, 6, 14]. The combined stateparameter to be estimated is given by = [ln(κ_{f}),pr,S, W_{c}]^{T}, where ln(·) is natural logarithm of the permeability field, W_{c} denotes water cut, and porosity is assumed to be known. The EnKF introduced by Evensen [15] is a sequential Monte Carlo method where an ensemble of model states evolves in statespace, with the mean as the best estimate and spread of the ensemble as the error covariance, as summarized in the following steps. Each of the ensemble members is forecasted independently,^{1}
where F[·] is the forecast operator [finescale model Eqs. (1) and (2b)], superscript (i) denotes the ith ensemble member; from this point we on drop the time subscript. The ensemble mean and covariance are defined as
where A = (b^{(1)},b^{(2)},...,b^{(Ne)}), b^{(i)} = Ψ^{(i)}  Ψ, and N_{e} is the number of ensemble members. In a twin experiment, the observed water cut W_{c}^{o} is related to the truth via W_{c}^{o} = H[ Ψ^{t}], where H[ Ψ^{t}] is the true water cut. For each ensemble member, we randomly perturb W_{c}^{o} to generate observational samples,
where ν^{(i)} simulates observational error sampling, obtained as independent and identically distributed (iid) samples [16] from a normal distribution with zero mean and variance R. We note that if only the watercut data is being measured, the mapping from model to observational space H is trivially equal to [0 0 0 I], since Ψ = [ln(κ),pr,S,W_{c}]^{T} . The forecasted ensemble Eq. (3) is updated by assimilating the observed data,
where K is the Kalman gain, given by Computationally efficient implementation of the EnKF is discussed, for example, in [2] and [17] (note that the ensemble error covariance, before or after assimilation, is not explicitly computed and we instead use the ensemble members for obtaining the covariance information). We use the above set of assimilated ensemble states, {^{(i)}}_{i=1}^{Ne}, in the finescale simulation model Eq. (3) for prediction until the next set of observational data is available. 3. Coarsescale data assmilationThe EnKF presented so far used only the dynamic production data (water cut) y with error ν = y  H[ Ψ^{t}], ν ~ (0,R) to update the ensemble Eq. (6). In addition to y, we now consider another independently measured data with independent errors z, which is static and is on a coarser scale compared to the finescale variables in Eqs. (1) and (2b). We assume that the corresponding measurement error is given by ω = z  U[ Ψ^{t}], with zero mean and Q covariance, ω ~ (0,Q), and U is a mapping of finescale variables Ψ to coarsescale data, z, i.e., U : Ψ z. Then the likelihood of z is given by
If this static data z corresponds to coarsescale permeability data (as considered in [9] and [7]), then U = [0 0 0], where : κ_{f} κ_{c} is a nonlinear mapping that maps the finescale permeability field (κ_{f}) to coarsescale field (κ_{c}) via an upscaling procedure (e.g., Durlofsky [18] and Durlofsky [19]). (Details are provided in Section 3.3.) Alternatively, if z corresponds to coarsescale saturation inverted from 4d seismic data (as mentioned in the Introduction), then U = [0 0 0], such that is a mapping of finescale saturation S_{f} to coarsescale saturation S_{c} = S_{f}. (Here we consider a simple volume averaging for ; further details are provided in Section 4.3). Now, our goal is to obtain an estimate which is based on both watercut and available coarsescale data. The likelihood of watercut data y is given by
The probability distribution function (pdf) of the predicted ensemble,
where Ψ and P^{f} are the predicted ensemble mean and covariance, respectively Eqs. (4a) and (4b). Then, using Bayes theorem, we obtain
Based on the above equation, following the (*) term, all the available data (z and y) could be assimilated in one step (details follow in Section 3.1), whereas based on the (†) term, the measured data y and z can be assimilated in a sequential manner. First assimilate the fractional flow (y) to obtain an ensemble conditioned on y, i.e., p( Ψy), which could then be used to assimilate the coarsescale data z (further explained in the following Section 3.2). 3.1 CoarseScale Data Assmilation: In OneStepAll the available data, y and z, could be assimilated in one assimilation step by a modification to the modeltoobservation space operator, H. If coarsescale permeability data at a single coarsescale is to be assimilated, H = [0 0 I]. Alternatively, if coarsescale saturation data is available, H = [0 0 I]. The finescale estimated ensemble is obtained as in Section 2.2, with Eq. (6) modified to account for the additional coarsescale data,
where . From now we on refer to this onestep assimilation of coarsescale data procedure as reg EnKF. A consequence of the particular form of H is that it introduces contributions for crosscorrelations between upscaled variables and production data. From a computational point of view, it has been known that EnKF is not very efficient for assimilation of large amounts of data [20], which could arise in our case, in complex threedimensional cases, and also if data at multiple coarse scales is to be assimilated. In such a situation, different kinds of data could be assimilated in batches [21], which is described in the Section 3.2. 3.2 CoarseScale Data Assmilation: In a SequenceWe obtain an intermediate ensemble by assimilating y, denoted by {^{(i)}}_{i=1}^{Ne},
as discussed in Section 2.2. This intermediate ensemble and likelihood in Eq. (7) can then be combined [† term in Eq. (10)] to obtain the final estimate {^{(i)}}_{i=1}^{Ne},
Therefore, in a leastsquared sense, the final estimate maximizes the posterior pdf p( Ψz,y) and corresponds to the minimum of = _{z} + _{y} + _{f}. See Appendix A for further details (where we show that the solution ^{(i)} corresponds to the minimum of , for any i^{th} ensemble member). If coarsescale data is available at only one coarse scale, then the finescale estimated ensemble is obtained by first assimilating production data followed by assimilation of the coarsescale data,
^{f} is approximated using the intermediate ensemble ^{(i)}; henceforth we refer to this sequential, coarsescale EnKF data assimilation procedure as csEnKF. Note that data at multiple coarse scales can be sequentially assimilated by suitable repetition of Eq. (14b), with corresponding upscaling operators. For the coarsescale saturation data, which may be available at only certain times, for only those assimilation cycles is Eq. (14b) applicable, whereas in the case of permeability data, considering it to be prior information regarding the finescale permeability, it is always to be honored; hence, both of the above steps (14a) and (14b) are to be always applied. The csEnKF algorithm is detailed in Appendix B, and a flow chart is given in Fig. 1. Implementation of this algorithm entails upscaling of each ensemble member at every assimilation step, i.e., N_{e} times the upscaling operator U[·] needs to be applied. In addition, if the dimension of the coarsescale grid is N_{c} = n_{c} × n_{c}, then we need to perform an Singular Value Decomposition (SVD) of a rectangular matrix of size N_{c} × N_{e}. Hence, the total computation expense involves N_{e} upscales and SVD of the N_{c} × N_{e} matrix. Note that a similar upscaling is involved in the case of the reg EnKF, but the size of the matrix to compute SVD is now (N_{c} + N_{wc}) × N_{e}, where N_{wc} is the dimension of the watercut data. In addition, if there are a number of coarse scales then the size of the matrix whose SVD is to be computed will grow for the reg EnKF, since all the data is assimilated in one step. Whereas for the csEnKF, coarsescale data is assimilated in a sequence, the estimate from one coarse scale being used as a prior for the next scale, the matrix size to compute SVD is always N_{c} × N_{e}. (See also remark 3 of the coarsescale EnKF algorithm in Appendix B.) FIG. 1: Flowchart for assimilation of coarsescale data using the EnKF. For nonlinear upscaling operators, such as , the final estimates from the reg EnKF would be different from those obtained using the csEnKF. When both the coarsescale and watercut data are assimilated together as in the above reg EnKF, it would imply fitting a multivariate normal likelihood to the different measured data together, whereas when the different kinds of data are assimilated one after another, as in the csEnKF, we fit each data separately, with a different pdf. (For further details on this topic, please see Dance [21] and references therein; also see Section 5.) 3.3 Upscaling MethodsIn brief, the main idea behind upscaling of absolute finescale permeability is to obtain effective coarsescale permeability for each coarsegrid block.Upscaling techniques in conjunction with the upscaling of absolute permeability have been used in groundwater applications (see, e.g., [19, 22, 23]). The link between the coarse and the finescale permeability fields is usually nontrivial, because one needs to take into account the effects of all the scales present at the fine level. In the past simple arithmetic, harmonics, or power averages have been used to link properties at various scales. These averages can be reasonable for low heterogeneities or for volumetric properties such as porosity. For permeabilities, simple averaging can lead to inaccurate and misleading results. In this paper we use the flowbased upscaling methods. Consider the finescale permeability that is defined on a domain with underlying fine grid as shown in Fig. 2. On the same plot a coarsescale partitioning of the domain is also illustrated. To calculate the coarsescale permeability field at this coarse level, we need to determine it for each coarse block, Ω_{c}. The coarsescale permeability is computed so that it delivers the same average response as that of the underlying finescale problem, locally. The calculation of the coarsescale permeability based on local solutions is schematically shown in Fig. 2. For each coarse domain Ω_{c} we solve the local problems
with some coarsescale boundary conditions. An example of such boundary conditions is given by ϕ_{j} = 1 and ϕ_{j} = 0 on the opposite sides along the direction e_{j} and no flow boundary conditions on all other sides, alternatively, ϕ_{j} = x_{j} on ∂Ω_{c}. For these boundary conditions the coarsescale permeability (κ_{c}) is given by
where ϕ_{j} is the solution of Eq. (15) with prescribed boundary conditions. Various boundary conditions such as periodic, Dirichlet, etc. can have some influence on the accuracy of the calculations. These issues have been discussed, e.g., in [24]. In particular, for determining the coarsescale permeability field one can choose local domains that are larger than target coarse block, Ω_{c}, in Eq. (15). Furthermore Eq. (16) is used in the domain Ω_{c}, where ϕ_{j} are computed in the larger domains with correct scaling (see [24]). This way one reduces the effects of the artificial boundary conditions imposed on Ω_{c} (for details see [24]). FIG. 2: Schematic illustration of upscaling (not to scale). Bold lines indicate a coarsescale partitioning, while thin lines show a finescale partitioning within coarsegrid cells. In this paper we upscaled a 50 × 50 fine grid to a 5 × 5 coarse grid. We denote by the local operator that maps the local finescale permeability field κ_{f} onto κ_{c}, defined on the coarse grid as in the above Eq. (16). For our computations we assume
where ϵ are some random fluctuations that represent inaccuracies in the coarsescale permeability. One source of these fluctuations is the errors associated with solving inverse problems on the coarse grid. The other source of the inaccuracies include the fact that the inversion on the coarse grid does not take into account the adequate form of the coarsescale models. Indeed, the inversion on the coarse grid for flow problems often involves the same flow equations as the underlying fine ones, for example, the same relative permeabilities are used for the coarsescale problems as those for the finescale problems or the effects of macrodispersion are neglected. It is known that the flow equations at the coarse level may have a different form than the underlying finescale equations [19, 2527]. In general, this form depends on the detailed nature of the heterogeneities, which are very difficult to obtain in solving inverse problems. In our paper we use Gaussian errors in Eq. (17) and consider the impact of coarsescale data precision (i.e., nature of ϵ) by varying the variance of ϵ (see Section 4.2 for more details). 4. Numerical resultsFor our numerical tests, we use a 50 × 50 fine grid (dimensionless domain size 50 × 50) and two kinds of coarsescale data in a twin experiment setup. First we consider coarsescale permeability, which in reality, could be obtained by coarsescale inversion of fractional flow data on a coarse grid [9, 28]. In this study we upscaled the reference finescale permeability (described below) to a 5 × 5 grid to obtain a coarsescale permeability using flow based upscaling (Section 3.3). This coarsescale field could be thought of as static data, which is to be honored as constraint (up to the prescribed measurement data variance) in Eq. (7); hence, we need to always assimilate it in every assimilation cycle. In reality we never know the reference field; therefore, this experimental setting is unrealistic. However, it allows us to compare and contrast a variety of test cases. For the second set of results, a coarsescale saturation is used which in practice could be obtained from inversion of 4d seismic measurements (see Section 1). Here, the coarsegrid saturation was obtained by volume averaging of true finescale saturation at some specific observation times (further details are given in Section 4.3). Therefore, unlike coarsescale permeability, static data constraint, which is to be always satisfied, the coarsescale saturation data is assumed to be available at only a few observation times. Following the flowchart in Fig. 1, we always have coarsescale data if it is coarse permeability, and only at those few observation times for coarsescale saturation data. An initial ensemble with different permeability realizations was generated using the sequential Gaussian simulation (Deutsch and Journel [29]). We specified a Gaussian variogram model with a correlation length of 20 gridblocks in the x direction and 5 gridblocks in the y direction. One of the realizations is used as the “true” field (shown in Fig. 3) and was removed from the ensemble. Porosity (ϕ) is assumed to be equal to 0.15 for all grid blocks. For simplicity, relative permeabilities k_{rj} are assumed to be linear functions of water saturation (S): k_{rw}(S) = S,k_{ro}(S) = 1  S. One injection well at the center of the field (injection rate: 71.4 m^{3}/day) and four producing wells at the four corners (all with equal rate of 17.85 m^{3}/day) were considered. The finescale model Eqs. (1)(2b) are solved with no flow boundary conditions, zero initial water saturation, and by discretizing the transport equation using the firstorder upwind finite volume method. In the top panel of Fig. 4 we provide the predicted fractional flow for 256 initial ensemble members along with the true fractional flow (obtained from the true permeability field). FIG. 3: Schematic illustration of upscaling (not to scale). Bold lines indicate a coarsescale partitioning, while thin lines show a finescale partitioning within coarsegrid cells. In this paper we upscaled a 50 × 50 fine grid to a 5 × 5 coarse grid. FIG. 4: Schematic illustration of upscaling (not to scale). Bold lines indicate a coarsescale partitioning, while thin lines show a finescale partitioning within coarsegrid cells. In this paper we upscaled a 50 × 50 fine grid to a 5 × 5 coarse grid. To compare and contrast our results using coarsescale data and different versions of EnKF, we use the following mean L_{2}norm error. Since we know the true (fine and coarsescale) field for our synthetic problem, denoting the true permeability field by κ^{true}, the error for any ensemble member is given by Consider the L_{2} norm of the error for each member, e^{(i)}_{2} = Σ_{j}[e^{(i)}_{j}]_{2}, by which we define the mean L_{2} error as
so that e gives us an indication of the distance of the entire ensemble from the true solution κ^{true}. Since after assimilating any observation we updated all the ensemble members, we can monitor the variation of e over the time of assimilation; the success of assimilation can therefore be related to the decrease in e. 4.1 EnKF with Water Cut Data OnlyWe start with a presentation of results obtained with assimilation of watercut data only. Next we discuss results with coarsescale data. The watercut data from the reference field is assumed to be available every 200 days, with mean zero and standard deviation of 0.01 (therefore R^{1/2} = 0.01I_{4}, where I_{4} is the unit matrix of size 4 × 4, since there are four producing wells). The observed data is assumed to be available up to 2400 days; hence, we performed assimilation between 200 and 2400 days. A prediction beyond the interval of data assimilation, up to 4000 days, is also provided. The choice of ensemble size (N_{e}) is very important for successful data assimilation using EnKF. This is because a finite size ensemble prediction is used to estimate the prior error covariance P^{f} Eq. (4b). For small sample sizes, sampling errors in the covariance estimates result in insufficient variance for P^{f}, so that observations which lie outside the small ensemble spread are completely ignored [17, 30]. (We are trying to sample a covariance matrix for unknown variables: ln(κ),pr,S,W_{c}, i.e., an unknown of size 3 × 2500 plus four fractional flow data in this case, using sample sizes that are far lesser, resulting in severely reduced rank covariance matrices.) Different approaches such as covariance inflation and localization have been proposed to alleviate this problem of ensemble inbreeding, which is discussed elsewhere (see [3134] and references therein for further details). An ensemble with sufficiently large number of members needs to be selected so that the assimilation system would not severely suffer from the abovedescribed problem.^{2} Here we present our data assimilation results with N_{e} = 256, and in Sections 4.2 and 4.3 we briefly discuss some important characteristics of the error covariance matrix such as variance and eigenspectrum in the context of coarsescale data assimilation. The issue of coarsescale data assimilation with smaller ensemble sizes would be tackled in the future. In the bottom panel of Fig. 4 we plot the ensemble and true watercut data after assimilation of only watercut data with the EnKF. Comparing with the initial forecast (top panel), we observe that the assimilated ensemble better envelopes the true data. Also, the ensemble mean saturation field after 500, 1000, 2000, 3000, and 4000 days of simulation better compares with the true saturation than with no assimilation in Fig. 7. The final permeability field after assimilation for the ensemble mean and a few members is compared with the true field in Fig. 9. Note that the central, southeastnorthwest channel is prominent, but the features at the southwest and northeast corners are not well captured, which is reflected in the plot of mean saturation (Fig. 7), where many finescale features present in the true saturation field are not present in the ensemble mean. Therefore, assimilation of only watercut data helps in identifying some of the important features. 4.2 EnKF with WaterCut and CoarseScale Permeability DataIn addition to watercut production data, the coarsescale permeability data, as described in Section 3.3 was used as additional measured data. Flowbased upscaling of the reference permeability field was used as a proxy for the inverted coarse field. Following our previous notation, this coarsescale permeability data is denoted by z Eq. (7). The mapping between state variables (at finescale) and observations (at coarsescale) as given by U = [0 0 0], , denotes flowbased upscaling. For the reg EnKF, H = [0 0 I] in Eq. (11) of Section 3.1. Exactly as in the previous section, we prescribed the same frequency (of availability) and precision R for the watercut data. Since we use coarsescale permeability as additional data, it is to be assimilated whenever we assimilate watercut data. For 5 × 5 coarsescale data with mean zero and variance, Q = qI_{25} (we present results with q = 4,2,1, and 0.1), so that we can consider the impact of coarsescale data precision. In the left panel of Fig. 5 we plot the variation of mean L_{2} error e Eq. (18) with observation time at the coarse scale for different values of q and using reg EnKF as well as csEnKF. In the right panel of the same figure we show the correlation between coarsescale ensemble mean and true fields for q = 1. The values of correlation coefficients for different values of q are provided in Table 1. Note that as the precision of coarsescale data is increased, i.e., for smaller value of variance, we observe a larger decrease in coarsescale mean L_{2} error and higher correlation with true coarsescale field. This would be expected because smaller variance Q implies more strict coarsescale data constraint in Eq. (7) and hence, the coarsescale data is more accurately assimilated as it is made more precise. The watercut data prediction using the final permeability field after assimilation for different coarsescale data precisions is plotted in Fig. 6. (The nature of results with q = 4 is similar to those with q = 2,1,0.5; thus, we drop it.) Notice the improved fit of ensemble prediction to the true data for more precise coarsescale data and also when compared to the assimilation of only water cut in Fig. 4, which is a consequence of the additional coarsescale data being available. However, the watercut prediction with the csEnKF compares better with the truth than that with the reg EnKFfor higher values of q = 2 and 1; with q = 0.5, the reg EnKF prediction is highly improved. In Fig. 7 we compare the ensemble mean with the true saturation. Once again, as the coarsescale data constraint is more precisely imposed, the ensemble mean saturation captures most of the features in the true field. Also, notice that the reg EnKF saturation prediction improves more markedly as q is lowered when compared to the csEnKF, which could explain the better watercut fit in Fig. 6 for q = 0.5 with the reg EnKF. FIG. 5: Schematic illustration of upscaling (not to scale). Bold lines indicate a coarsescale partitioning, while thin lines show a finescale partitioning within coarsegrid cells. In this paper we upscaled a 50 × 50 fine grid to a 5 × 5 coarse grid. TABLE 1: Correlation coefficient between ens. mean permeability ln (κ) and true permeability ln(κ^{true}), at coarse as well as fine scales for different precisions q of coarsescale permeability data. The coarse scale is denoted with subscript c and finescale with f. Results with both reg EnKF and csEnKF are given. For only watercut data assimilation, corr [ln(κ_{f}),ln(κ_{f}^{true})] = 0.3074.
FIG. 6: Same as in Fig. 4, but for assimilation of both coarsescale permeability and watercut data. The coarsescale data precision is varied, q = 2,1,0.5. Results with q = 4 had a trend similar to the ones plotted and hence, are not shown. FIG. 7: Plot of the evolution of saturation field for the truth, ensemble mean: initial forecast (no assimilation), assimilation of only water cut, and assimilation of both water cut and coarsescale permeability data with various precisions. We denote only water cut with “Wc”, and watercut and coarsescale permeability data with “Wc+Kc”. In the left panel of Fig. 8 we plot the finescale mean L_{2} error for the permeability field with different values of q, and in the right panel of the same figure we plotted the correlation between the ensemble mean and the true finescale permeability for q = 1. The correlation coefficients are given in Table 1. (For assimilation of only watercut data, we obtained a correlation coefficient equal to 0.3074.) Though the mean L_{2} error is lower with the csEnKF, a slightly higher correlation is obtained with the reg EnKF. We observe that higher precision, i.e., lower q, does not necessarily imply highest correlation, whereas we obtained a lower mean L_{2} error. The final permeability field after assimilation for the ensemble mean and a few members is shown in Fig. 9. We note that the lowpermeability region at the northeast and high permeability at the southwest corners are well captured. Also, the ensemble mean permeability with the csEnKF has some features not very well present, in that with the reg EnKF, particularly with q = 2,1, e.g., the lowpermeability region to the left of the central southwestnortheast channel. As we noted in Section 3.1, the estimates obtained using the reg EnKF are expected to be different from those obtained with the csEnKF. We observed that for higher values of the coarsescale data variance (q = 2 and 1), the csEnKF yielded better water cut, ensemble mean saturations, and permeability estimates when compared to the reg EnKF, whereas with q = 0.5, the results are similar (but not exactly same) with both versions of the EnKF. These results indicate that an optimal value for the coarsescale data variance is important, particularly for the different versions of the EnKF (reg EnKF or csEnKF), and it could be obtained by a prior calculation of the uncertainty in the coarsescale data which can be addressed in a future study. Also, more complex, studies in coarsescale data at more than one scale and for threedimensional models are needed to understand the merits and demerits of each version of the EnKF. FIG. 8: Same as in Fig. 5, but for the fine scale. FIG. 9: Plot of the finescale permeability field for the truth, ensemble mean: initial forecast (no assimilation), assimilation of only watercut, and assimilation of both watercut and coarsescale permeability data with various precisions in the top row. For the second column onward we show selected ensemble members: 50,100,150, and 200, respectively. We denote only water cut with “Wc”, and watercut and coarsescale permeability data with “Wc+Kc”. Regardless of the version of EnKF being used, after every assimilation cycle we can obtain an estimate of the analysis error covariance matrix for the ensemble finescale permeability. Using the finescale assimilated permeability fields κ_{f}^{(i)} we define the following ensemble mean permeability and error covariance,
Minimization of variance of the covariance matrix P_{κf,κf} is desired in various Kalman filtering applications [2] as it provides a measure of uncertainity. In the top row (left panel) of Fig. 10 we plotted the normalized variance of P_{κf,κf} during assimilation with various precisions of coarsescale data and for both reg EnKF and csEnKF, and also for assimilation of only fractional flow data (we normalized using the variance from the initial ensemble). We observed that a higher reduction in variance is obtained as coarsescale data precision is increased, and the csEnKF obtains more reduction than the reg EnKF. This trend is seen even for an ensemble of much larger size, for example, with 1000 members as shown in the bottom row (left panel) of Fig. 10. FIG. 10: Top row (left panel): Normalized variance of the error covariance matrix (Section 4.2) within assimilation window (normalized with respect to variance before assimilation, using the initial ensemble), for assimilation of only watercut, and assimilation of both water cut and coarsescale permeability data with various precisions. Right panel: Eigenspectrum of the error covariance matrix for the last assimilation cycle after 2400 days with 256 ensemble members. Bottom row: Same as top row but with 1000 ensemble members. Another important property of interest, particularly for ensemble data assimilation is the eigenspectrum of P_{κf,κf}, which is plotted in the right panels of Fig. 10; the top row represents 256 ensemble members, whereas the bottom row represents 1000 members. This eigenspectrum is from the last assimilation cycle at the end of 2400 days; for other assimilation cycles, the spectrum was similar. A smooth distribution of eigenvalues is desired, and indeed, we observe that, but the tails of the spectrum are steep for only watercut data assimilation and csEnKF. As the coarsescale data precision is improved, the trailing eigenspectrum gets smoothed out, which may indicate that the eigenvectors associated with the small eigenvalues are able to resolve some smallscale correlation features. Also, the larger variance associated with the leading eigenvectors as obtained for only watercut data assimilation seems not to be the case with coarsescale data assimilation (some of the desirable properties regarding eigenspectrum of covariance matrices are given, e.g., by Hamill et al. [32]). We note that for an ensemble size of N_{e} = 1000, the spectra are almost identical for both versions of the EnKF. In Fig. 11 we compare sidebyside the spectra with N_{e} = 256 for the reg EnKF and csEnKF, with the 1000 ensemble member eigenspectrum. (Since the spectra are similar for the 1000 ensemble member case, we plot and compared only with the reg EnKF.) As noted in Fig. 10, with higher precision of coarsescale data, the tail of the spectrum gets smoothened, but also the magnitude of trailing eigenvalues is decreasing, which suggests that for smaller ensemble sizes there could be issues with loss of rank of the error covariance matrix. This problem seems to be slightly more aggrevated for the csEnKF. When compared to the 1000 member case, the leading eigenvalues seem to be slightly larger, as observed in [32]. In any case, in this study we did not apply any covariance inflation or localization (and always used an ensemble of size 256, which seems to preserve the rank of ensemble up to about 245 eigenvalues; see Fig. 11). These topics in the context of coarsescale data assimilation will be covered in a future study. FIG. 11: Comparison of the eigenspectrum of the error covariance matrix for 256 and 1000 ensemble members for different precisions (q) of the coarsescale data. As in Fig. 10, the spectrum corresponds to the last assimilation cycle. Plot of the spectrum for csEnKF and reg EnKF for 1000 ensemble members was very similar (bottom row of Fig. 10) and hence, is not plotted. 4.3 EnKF with WaterCut and CoarseScale Saturation DataAs mentioned in the Introduction, by coarsescale inversion of 4dseismic data we could obtain dynamic data such as coarsescale pressure and saturation. In this section we attempt to assimilate such a coarsescale saturation in a twin experiment along with the fractional flow data using the EnKF algorithms discussed in Sections 3.2 and 3.1. To this end, the saturation obtained by using the reference permeability is saved at three different times: 200,1200, and 2400 days, which respectively correspond to the start, middle, and end of the time window of data assimilation. This saved finescale saturation field is then upscaled (see Section 3.3) by volume averaging to a 5 × 5 coarsescale grid and used as observed coarsescale saturation data. If we denote the volume averaging by operator , acting on finescale saturation S_{f}, to give coarsescale saturation S_{c} = S_{f}, then the mapping between state variables at fine scale and measured data at coarsescale is given by U = [0 0 0]. Therefore for the reg EnKF (Section 3.1), the measured data is related to the finescale variables via H = [0 0 I] in Eq. (11). For the csEnKFwe use above U operator to compute the misfit: zU[ ] in Eq. (14b) (i.e., steps 2.1 and 2.4 of the csEnKF algorithm in Appendix B). Unlike the coarsescale permeability data which was taken into account at every assimilation step, by construction, the coarsescale saturation data is available only at a few assimilation steps, in this particular case, assimilation after 200,1200, and 2400 days. To be consistent with our previous results, the frequency (of availability) and precision R for the watercut data has been kept the same. For the coarsescale saturation data we prescribed zero mean and variance, Q = q_{s}I_{25}, with q_{s} = 0.1,0.01, such that the precision is varied from low to high. Since the saturation ranges between 0 and 1, and the fractional flow data is usually more accurately measured than 4dseismic data, we picked q_{s} to be always larger than the variance in fractional flow data. In the left panel of Fig. 12 we plotted the variation of mean L_{2} error for the coarsescale saturation (while assimilating) vs observation time. (For our test case, we had assumed zero initial water saturation; therefore, the water saturation increases in time and hence, the inherent, increasing trend in this figure.) Note that whenever the coarsescale saturation is assimilated (200,1200, and 2400 days) the error decreases for both values of q_{s} considered. The watercut data prediction using the assimilated ensemble members is given in Fig. 13; the fit of ensemble watercut with the truth gets better as the coarsescale saturation is prescibed higher precision. FIG. 12: Assimilation of coarsescale saturation and water cut. Left panel: Variation of mean L_{2} norm error for coarsescale saturation, at different precisions of coarsescale data q_{s}, for both reg EnKF and csEnKF with coarsescale saturation data. Right panel: Same as in left, but for the finescale permeability field. FIG. 13: Similar to Fig. 6, but assimilation of watercut and coarsescale saturation at different precisions. We denote only water cut with “Wc”, and watercut and coarsescale saturation data with “Wc+Sc”. We discuss the finescale results starting with finescale saturation and then the finescale permeability. The ensemble mean saturation is compared to the true field at certain times in Fig. 14. By assimilating the coarsescale saturation data we are able to capture many of the subtle features that are present in the true saturation field, such as the fingers that develop off the center toward the northeast corner and sharp contrast between different levels of saturation, throughout the entire time interval (up to 4000 days) considered. Regarding the finescale permeability, a comparison of the mean L_{2} error is shown in the right panel of Fig. 12. Note the marked reduction in error when coarsescale saturation data is assimilated. Also, the correlation of the ensemble mean permeability with the truth is improved as q_{s} is decreased, as shown in Table 2. The finescale permeability fields for a few ensemble members and the mean are shown in Fig. 15. Based on the above results, we observed that unlike in the case of coarsescale permeability data assimilation, coarsescale saturation data assimilation does not yield significant reduction in error (comparing the left panel in Fig. 8 and the right panel in Fig. 12); neither are the correlations of finescale ensemble mean with the truth (second column of Tables 1 and 2). The results (ensemble mean saturation and permeability) are improved when compared to assimilation of only watercut data, particularly with q_{s} = 0.01, but again, not as much improved as with coarsescale permeability data assimilation. This could be anticipated, since the finescale permeability is more correlated to coarsescale permeability than to the coarsescale saturation. Regarding the two versions of the EnKF considered for assimilating the coarsescale saturation data, results are very similar to each other, even those for the variance of error covariance matrix P_{κf,κf} and its eigenspectrum at the end of 2400 days (Fig. 16). Note the steepness of the trailing eigenvalues with q_{s} = 0.1 (for both reg EnKF and csEnKF), which is similar to the assimilation of only fractional flow data; however, for q_{s} = 0.01, this undesirable effect has been smoothned out. This is similar to the result obtained with coarsescale permeabilty data assimilation in Fig. 10. Our observation that both versions of EnKF performed similarly could be due to the linearity of the finescale to coarsescale saturation mapping ; however, identical results would not be possible due to the different analysis equations and sampling of errors used in Eqs. (14b) and (11). FIG. 14: Assimilation of coarsescale saturation and water cut. Left panel: Variation of mean L_{2} norm error for coarsescale saturation, at different precisions of coarsescale data q_{s}, for both reg EnKF and csEnKF with coarsescale saturation data. Right panel: Same as in left, but for the finescale permeability field. TABLE 2: Same as Table 1, but for assimilation of coarsescale saturation data.
FIG. 15: Same as in Fig. 9, but for the assimilation of coarsescale saturation and water cut. FIG. 16: Same as in Fig. 10 (top row), but for the assimilation of coarsescale saturation and water cut. 5. ConclusionsThe EnKF is increasingly being used for subsurface characterization in various geological and groundwater applications to identify finescale state and parameters. Recently, dynamic data other than production data has been considered in the EnKF context [11, 12]; nevertheless, the observed data to be assimilated was assumed to be at the finest scale. For a number of reasons, it is widely recognized that usage of additional multiscale data could further reduce the uncertainty at the fine scale. Also, it is often important to preserve largescale features of the permeability field. These are coarsescale features that can typically represent connectivity of the media. For example, facies consisting of high permeable regions are important for flow and transport. On the other hand, these features alone are not sufficient to match watercut data. By performing inversion on the coarse grid, we can capture these largescale features with more accuracy and certainty, and in turn, the coarsescale inverted permeability could be used as a prior constraint in EnKF data assimilation of the finescale fields. Here we proposed assimilation of coarsescale data along with watercut production data. We showed that the modifications to the EnKF for multiscale data assimilation are completely recursive and easily implementable. The relation between fine and coarse scales was modeled via flowbased upscaling, which could be thought of as a nonlinear observation operator linking the coarsescale data to the unknown finescale variables. In addition, the proposed methodology could be used in any other sequential data assimilation method, as well to assimilate data at multiple coarse scales. Two versions of EnKF were suggested: (i) reg EnKF, where all the data (coarsescale and watercut) were assimilated together, and (ii) csEnKF, where the data were assimilated sequentially in batches. Ensemble members obtained after assimilating watercut data are used as a prior to assimilate coarsescale data. Though in our current paper we used only one coarse scale, the proposed method can be easily implemented to integrate as many coarse scales as required by the available data. Also the methodology is independent of the upscaling operator. The assimilation setup was tested and compared for a twodimensional synthetic 50 × 50 heterogenous true field. Two kinds of coarsescale data were considered. In the first implementation, coarsescale permeability data was considered and in the second, coarsescale saturation. In our twin experiment setup, both of these data were derived from the reference field and, in both cases a 5 × 5 coarse grid was used. The coarsescale data variance was varied from low to high in order to study its impact on finescale assimilated fields and watercut predictions. In all cases we observed that the assimilated, ensemble mean coarsescale (permeability/saturation) field for all variances was highly correlated to the true coarsescale field. In addition, lower variance in the coarsescale data yielded higher correlation. The watercut data was better honored, for higher precision of coarse data. When assimilating coarsescale permeability we observed that the csEnKF gave better fit with the true saturation, watercut, and finescale permeability than the reg EnKF. It also yielded less error in an averaged L_{2} norm error taken with regard to the reference field, whereas both versions of EnKF performed similarly when assimilating coarsescale saturation. As shown in Appendix A, for a linear observation operator, assimilation of coarsescale data in batches (i.e., as in csEnKF) or in shot (as in reg EnKF) would yield the same estimate. Then the cause for the difference in the performance of the two versions of EnKF for assimilation of coarsescale permeability could be either due to ensemble size or linearity/nonlinearity of the upscaling (observation) operator. This issue has been outlined by Dance [21] and references therein. As far as the number of ensembles is concerned, we observed that even with a larger ensemble size of 1000 members, we noticed different performance of the two versions of EnKF (Fig. 10, reduction in normalized variance of P_{κf,κf} while having similar ranks in terms of the eigenspectrum). Certainly further study with different upscaling operators and coarsescale data at multiple levels would be needed to study this aspect of the two versions of EnKF considered here. Over all, inclusion of coarsescale data replicated many of the subtle features present in the finescale permeability and saturation fields which were not present after assimilating only watercut data. As our results indicate that the inclusion of coarsescale data enhances identification of the multiscale reservoir characteristics, it is important to study methods to obtain coarsescale data as well as its precision. In a realistic scenario, coarsescale inversion [7, 28], and in the future perhaps with more computing resources, MCMC methods [8], could be used to obtain such data. The coarsescale saturation obtained using inversion has been shown to yield improved estimates in a threedimensional reservoir case by Devegowda et al. [35]. Our current and future work is directed toward obtaining coarsescale data with higher precision and its assimilation using ensembles of smaller size for complex threedimensional cases. AcknowledgmentsThe work of Akhil DattaGupta and Yalchin Efendiev is partially supported by DOE (DEFG0300ER15034), NSF CMG 0724704, and KAUST award number KUSC101604. References1. Lumley, D. E., Timelapse seismic reservoir monitoring, Geophysics, 66:5053, 2001. 2. Evensen, G., Data Assimilation: The Ensemble Kalman Filter, Springer, 2006. 3. Nævdal, G., Johnson, L., Aanonsen, S., and Vefring, E., Reservoir monitoring and continuous model updating using ensemble Kalman filter, SPE J., 10:6674, 2005. 4. Wen, X.H. and Chen, W. H., Realtime reservoir model updating using ensemble Kalman filter, SPE Reservoir Simulation Symposium, SPE 92991, 2005. 5. Gu, Y. Q. and Oliver, D. S., The ensemble kalman filter for continuous updating of reservoir simulation models, J. Energy Resour. Technol.Trans. ASME, 128:7987, 2006. 6. Jafarpour, B. and McLaughlin, D. B., History matching with an ensemble Kalman filter and discrete cosine parametrization, Comput. Geosci., 12:227244, 2008. 7. Lee, S. H., Malallah, A., DattaGupta, A., and Higdon, D., Multiscale data integration using Markov random fields, SPE Reservoir Eval. Eng., 5(1):6878, 2002. 8. Efendiev, Y., DattaGupta, A., Ginting, V., Ma, X., and Mallick, B., An efficient twostage Markov chain Monte Carlo method for dynamic data integration, Water Resour. Res., 41:W12423, 2005. 9. Efendiev, Y., DattaGupta, A., Osako, I., and Mallick, B., Multiscale data integration using coarsescale models, Adv. Water Resour., 28:303314, 2005. 10. Liu, Y. and Gupta, H. V., Uncertainty in hydrologic modeling: Toward an integrated data assimilation framework, Water Resour. Res., 43:W07401, 2007. 11. Dong, Y., Gu, Y., and Oliver, D. S., Sequential assimilation of 4D seismic data for reservoir description using the ensemble Kalman filter, J. Pet. Sci. Eng., 53:8399, 2006. 12. Skjervheim, J.A., Evensen, G., Aanonsen, S., Ruud, B., and Johansen, T., Incorporating 4D seismic data in reservoir simulation models using an ensemble Kalman filter, SPE J., 12:282292, 2007. 13. Gosselin, O., Aanonsen, S. I., Aavatsmark, I., Cominelli, A., Gonard, R., Kolasinski, M., Ferdinandi, F., Kovacic, L., and Neylon, K., History matching using timelapse seismic (HUTS), SPE Annual Technical Conference and Exhibition, Denver, SPE 84464MS, Oct 2003. 14. Gu, Y. and Oliver, D. S., History Matching of the PUNQS3 reservoir model using the ensemble Kalman filter, SPE J., 10:217224, 2005. 15. Evensen, G., Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast error statistics, J. Geophys. Rese., 99(10):14310, 162, 1994. 16. Burgers, G., van Leeuwen, P. J., and Evensen, G., Analysis scheme in the ensemble Kalman filter, Mon. Weather Rev., 126(6):17191724, 1998. 17. Chen, Y. and Zhang, D., Data assimilation for transient flow in geologic formations via ensemble Kalman filter, Adv. Water Resour., 29:11071122, 2006. 18. Durlofsky, L. J., Numerical calculation of equivalent grid block permeability tensors for heterogeneous porous media, Water Resour. Res., 27:699708, 1991. 19. Durlofsky, L. J., Coarse scale models of two phase flow in heterogeneous reservoirs: Volume averaged equations and their relationship to the existing upscaling techniques, Comput. Geosci., 2:7392, 1998. 20. Lawniczak, W., Hanea, R., Hemmink, A., and McLaughlin, D., Multiscale ensemble filtering for reservoir engineering applications, Comput. Geosci., 13:245254, 2009. 21. Dance, S. L., Issues in high resolution limited area data assimilation for qualitative precipitation forecasting, Physica D, 196:127, 2004. 22. Durlofsky, L. J., Behrens, R. A., Jones, R. C., and Bernath, A., Scale up of heterogeneous threedimensional reservoir descriptions, SPE Paper 30709, 1996. 23. Durlofsky, L. J., Jones, R. C., and Milliken, W. J., A nonuniform coarsening approach for the scale up of displacement processes in heterogeneous media, Adv. Water Resour., 20:335347, 1997. 24. Wu, X. H., Efendiev, Y. R., and Hou, T. Y., Analysis of upscaling absolute permeability, Discrete Contin. Dyn. Syst., Ser. B, 2:185204, 2002. 25. Efendiev, Y. R. and Durlofsky, L. J., Numerical modeling of subgrid heterogeneity in two phase flow simulations, Water Resour. Res., 38(8):W1128, 2002. 26. Efendiev, Y. and Durlofsky, L., Generalized convectiondiffusion model for subgrid transport in porous media, SIAM Multiscale Model. Simul., 1:504526, 2003. 27. Efendiev, Y. and Durlofsky, L., Accurate subgrid models for twophase flow in heterogeneous reservoirs, SPE J., 9:219226, 2004. 28. Yoon, S., Malallah, A. H., DattaGupta, A., Vasco, D. W., and Behrens, R. A., A multiscale approach to productiondata integration using streamline models, SPE J., 6:182192, 2001. 29. Deutsch, C. V. and Journel, A., GSLIB Geostatistical Software Library and User‘s Guide, Oxford Univ. Press, 1992. 30. Hendricks Franssen, H. J. and Kinzelbach, W., Realtime groundwater flow modeling with the Ensemble Kalman Filter: Joint estimation of states and parameters and the filter inbreeding problem, Water Resour. Res., 44:W09408, 2008. 31. Anderson, J. L., Exploring the need for localization in ensemble data assimilation using a hierarchical ensemble filter, Physica D, 230(12):99111, 2007. 32. Hamill. T., Whitaker, J., and Synder, C., Distancedependent filtering of background error covariance estimates in ensemble Kalman filter, Mont. Weather Rev., 129:27762790, 2001. 33. ArroyoNegrete, E., Devegowda, D., DattaGupta, A., and Choe, J., Streamlineassisted ensemble Kalman filter for rapid and continuous reservoir model updating, SPE Reservoir Eval. Eng., 11(6):10461060, 2008. 34. Sun, A. Y., Morrisa, A., and Mohantya S., Comparison of deterministic ensemble Kalman filters for assimilating hydrogeological data, Adv. Water Resour., 32(2):280292, 2009. 35. Devegowda, D., Akella, S., DattaGupta, A., and Efendiev, Y., Interpretation of partitioning interwell tracer tests using EnKF with coarsescale constraints, SPE Reservoir Simulation Symposium, SPE 119125MS, The Woodlands, TX, 24 Feb 2009. APPENDIX A. TwoStep CoarseScale Constrained Kalman Filter EstimateFrom Section 3, and For notational simplicity we denote μ _{Ψ} as μ and denote P^{f} by B. Step 1 (minimize _{f} + _{y}): First we minimize the sum, _{1} = _{f} + _{y}. The gradient^{3} of the above quadratic cost functional with respect to (w.r.t.) Ψ is given by Then the minimizer of _{1} satisfies (we assume H to be linear) By rearranging the above equation we get
Note that the Hessian of _{1} w.r.t. Ψ is given by B^{1} + H^{T} R^{1}H, and for linear quadratic cost functionals, the Hessian inverse is equal to the error covariance matrix. Therefore, the error covariance matrix for is given by
Step 2 (minimize _{g} + _{z}): We use , in Therefore, the minimum of _{g} + _{z} satisfies Using Eqs. (21) and (20) we can rewrite the above as It is trivial to show that also satisfies Therefore, the twostep method to obtain the final estimate gives the same results as a oneshot approach of minimizing _{f} + _{y} + _{z}. APPENDIX B. The CoarseScale EnKF AlgorithmAlgorithm 1. Coarsescale EnKF algorithm Run the simulation model up to a particular observation time for the entire ensemble to get predicted samples: {Ψ^{(i)}}_{i=1}^{Ne}, A = (Ψ^{(1)}, Ψ^{(2)},...,Ψ^{(Ne)}).
Remark 1: Remark 2: Note that the above algorithm is independent of the choice of upscaling procedure, and we can use the same algorithm for different kinds of coarsescale observed data (if available). Remark 3: ^{1} In this work we focus primarily on assimilation of coarsescale data using the EnKF, its feasibility, and impact on finescale estimates with different kinds of coarsescale data (Section 3); hence, we neglect modeling errors, which will be addressed in the future. ^{2} Our choice of N_{e} = 256 was based on observing the eigenspectrum and variance, discussed in Sections 4.2 and 4.3, by comparing the results for the 256 ensemble with a 1000sized ensemble; the 256sized ensemble did not suffer from the insufficient variance problem discussed above. ^{3} We note in passing that B and R are covariance matrices and are positive definite by construction and hence, for our derivation purposes, are formally invertible. ^{4} As noted in [2], this approach of accounting for the nonlinear observations operator U works well as long as U is weakly nonlinear and a monotonic function of model variables. 
Portal Digitalde  Biblioteca Digital  eLibros  Revistas  Referencias y Libros de Ponencias  Colecciones 