Groundwater Research, Rosbjerg et al. (eds) © 2000 Balkema, Rotterdam, ISBN 90 5809 133 3 Inverse Modeling of Unsaturated Flow Combined with Stochastic Simulation Using Empirical Orthogonal Functions (EOF) N.-O. Kitterød University of Oslo, Department of Geophysics, Norway S. Finsterle University of California, Lawrence Berkeley National Laboratory, Berkeley, California, USA Keywords: inverse modeling, unsaturated flow, EOF simulation ABSTRACT: A Bayesian maximum likelihood method is used to derive optimal parameters for an unsaturated flow problem. Liquid saturation data can be acquired with high spatial density at low cost by indirect methods, making saturation a potentially useful primary observation type for inverse modeling. However, limited sensitivity of saturation to unsaturated flow parameters makes it necessary to also include a priori information into the inversion procedure. The quality of the estimated parameter set is expressed through a covariance matrix. Impacts of parameter uncertainties are evaluated by conditional stochastic simulations, in which not only the most likely parameters are reproduced, but also the cross-correlation structure between the parameters. Ignoring parameter cross-correlations in the simulation procedure lead to Monte Carlo realizations with unlikely parameter combinations. In this paper, we present a conditional simulation method which utilizes the orthogonal functions derived directly from the estimated covariance matrix. 1 MOTIVATION In this study we examine the possibility of using liquid saturation data as primary observations to estimate unsaturated hydraulic properties. A numerical model is developed based on the Richards equation and the van Genuchten constitutive relations. The model is calibrated against liquid saturation data using the maximum likelihood inversion method (Carrera and Neuman, 1986) implemented in iTOUGH2 (Finsterle, 1999). This approach is potentially powerful because of the relative ease with which liquid saturation can be mapped using geophysical and geostatistical methods such as Ground Penetrating Radar (Hubbard et al., 1997) or - as done in this project - a combination of ordinary kriging and high-resolution spatial mapping by the neutron scattering method (Kitterød et al. 1997). These cost-effective methods provide spatially continuous images that also contain important information about the geological architecture. Because liquid saturation is not sufficiently sensitive to all unsaturated flow parameters, it is necessary to include prior information about these parameters to constrain the inversion. The uncertainties in the estimated parameters is evaluated and propagated through a prediction model to assess the uncertainties in the simulated flow fields. Correlations among the estimated parameters must be accounted for to reduce the impact of unlikely parameter combinations on the error analysis. Unlikely parameter combinations are excluded directly during the Monte Carlo simulations using a method known as Empirical Orthogonal Functions (EOF) or Karhunen-Loève simulation (Braud & Obled, 1991, Kitterød and Gottschalk, 1997). This method has been used in the past to generate spatially correlated property fields, but is equally applicable to uncertainty propagation analyses involving correlated parameters, as discussed below. 1 Groundwater Research, Rosbjerg et al. (eds) © 2000 Balkema, Rotterdam, ISBN 90 5809 133 3 2 CONDITIONAL PARAMETER SIMULATION The generation of multiple parameter sets for subsequent Monte Carlo simulations starts with an optimal estimate p, E[p]= p*, where p = pk, k = 1,…,n, and n is number of estimated parameters. The covariance matrix of the estimated parameters, Cpp = E{(p+E[p])(p-E[p])}, can be expressed as (Carrera and Neuman, 1986): ( C pp = s 02 J T C −ZZ1 J ) −1 (1) where s02 is the estimated error variance, J is the Jacobian sensitivity matrix, and Czz is the a priori matrix of measurement errors. Conditional realizations ξ of the stochastic vector p (denoted ξ p) can be generated by applying the Karhunen-Loève theorem (Davenport & Root, 1958): k = 1,..., n (2) ξ p = ξ Φkβk where β k is the eigenvector for the kth parameter derived from: C pp β Tk = µ k β k k = 1,..., n (3) The stochastic coefficient ξ Φ is drawn from a Gaussian probability density function with pk* as mean and the eigenvalue µk equal to the variance of pk, i.e., ξΦk ∈ N(pk*,µk). The expansion is often called double orthogonal because β k β Tj = D k , j = 1,..., n (4) where D has entries dij = 0 if i ≠ j, and E{ξ Φ k ξ Φ j } = δ kj µ k k , j = 1,..., n (5) where δij = 0 if i ≠ j, and 1 if i = j. 3 RESULTS The simulation algorithm is applied to a real case study. The results demonstrate that EOFsimulation narrows the realization outcome compared to standard Monte Carlo simulations where the cross-correlation structure is ignored. References: Braud, I. & Obled, C. 1991. On the use of Empirical Orthogonal Function (EOF) analysis in the simualtion of random fields. Stoch. Hydrol and Hydraul. 5, 125-134 Davenport, W.B. & Root, W.L. 1958. An introduction to the theory of random signalas and noise, McGraw-Hill, New York Finsterle, S. 1999. iTOUGH2 user’s guide. Earth Sciences Division, Lawrence Berkeley National Lab, University of California, LBNL-40040 Carrera, J. & Neuman, S.P. 1986. Estimation of Aquifer parameters under transient and steady state conditions: 1. Maximum likelihood method incororating prior information. Water Res. Res.22(2), 199-210 Hubbard, S.S., Peterson J.E. jr., Majer, E.L., Zawislanski, P.T. & Williams, K.H. 1997. Estimation of permeable pathways and water content using tomographic radar data. The leading edge Nov, 1623-1628 Kitterød, N.-O. & Gottschalk, L. 1997. Simulation of normal distributed fields by Karhunen-Loève expansion in combination with kriging. Stoch. Hydrol and Hydraul. 11, 459-482 Kitterød, N.-O. Langsholt, E., Wong, W.K. & Gottschalk, L. 1997. Geostatistic interpolation of soil moisture. Nordic Hydrology 28(4/5), 307-328 2