INVERSE MODELLING OF UNSATURATED FLOW COMBINED WITH STOCHASTIC SIMULATION USING EMPIRICAL ORTHOGONAL FUNCTIONS (EOF) Stefan Finsterle University of California, Lawrence Berkeley National Laboratory Berkeley, California, USA Nils-Otto Kitterød University of Oslo, Department of Geophysics, Norway e-mail: nilsotto@geofysikk.uio.no Background Input data Oslo Airport is a potential hazard to the unconfined groundwater aquifer at Gardermoen (fig. 1). Biological remediation may prevent serious pollution of the groundwater, but this protection requires that the transport to the groundwater is not too fast. Spatial and temporal variation in unsaturated flow properties however, make short cuircuiting and preferential flow very likely under extreme conditions. • Sedimentological architecture from Ground Penetrating Radar (fig.4 and 5) • Liquid saturation measured by Neutron Scattering and interpolated by kriging (fig.6) • A priori statistical data on flow parameters • Effective infiltration (fig. 7) Top 1 Figure 4. Ground Penetrating Radar 0 lp4 1 utm-N • Use liquid saturation as primary data for Bayesian Maximum 6678500 Likelihood Inversion of unsaturated flow parameters. 6676500 • Simulate parameter uncertainties by Karhunen6674500 Loève expansion. Gardermoen 2 p45 Dip 2 5 0 10m W Ground Penetrating Radar profiles E N 615,700 6,677,740 p42 10 p43 c11 0 c16 foresets sp4 Groundwatertable p41 k20 p43 k26 k10 p45 k28 k2 k4 2 3 k30 k8 k24 k6 p47 6,677,820 2D flow test tr 5 0 10 West-East <m> p35 Figure 6. Liquid saturation, May 11, 1995 6,677,840 Moreppen research site 1.84E-05 N18 Results N32 N12 N34 railway runways kg/s N30 6672500 N36 N38 N40 N42 616 000 N44 N46 N20 8.40E-06 4 • Main character of observed liquid saturation is simulated (fig. 8 and 9) and absolute permeabilities are estimated according to independent observations (fig. 10) • Simulation by EOF reproduce Cpp (fig. 11) • Neglecting Cpp imply unphysical parameter combinations and overestimation of parameter uncertainty (fig. 12) 2 0 25- 26- 27- 28- 29- 3012345678910- 11Apr- Apr- Apr- Apr- Apr- Apr- May- May- May- May- May- May- May- May- May- May- May95 95 95 95 95 95 95 95 95 95 95 95 95 95 95 95 95 -2 -1.16E-05 -4 tim e 618 000 GPR(46) The Forward Flow Model: The numerical code TOUGH2 (Preuss, 1991) is used to solve Richards equation with constitutive relations between pressure p, permeability kr and saturation S according to the van Genuchten model (fig.2): 1 6 -6.60E-06 GPR(44) 1 p S 8 1.34E-05 -1.60E-06 utm-E 1 m e 10 3.40E-06 5m 614 000 Figure 7. Effective Infiltration rate 2.34E-05 N10 500m 12 2.84E-05 Oslo Oslo Figure 1. The Moreppen research site Delta topsets 1 p41 6,677,800 Figure 5. Local Sedimentological Architecture p48 p44 p46 p33 6,677,780 0 615,800 615,750 6,677,760 Moreppen research site Dip 1 3 S Oslo airport Gardermoen Top 2 Flow model 1 n k r S 1 1 S 1 2 e 1 m e m 2 where Se is effective saturation, • Se = (S- Sr)/(1- Sr), • Sr is called residual liquid saturation, • 1/ is called air entry value, and • m=1-1/n where • n is called the pore size distribution index. Figure 8. Reproduction of observed liquid saturation in location c11 and c16 (cf.fig 6) A priori grainsize distribution data courtesy Anne Kristine Søvik Figure 2. Constitutive relations between pressure p, permeability kr and saturation S predef no_lim Inverse modeling: lim where y*j is observation of liquid saturation in space and yi(p) is the forward model response, p={ki, Sri 1/i,ni}, i=1,2,…,number of sedimentological units, in this case equal to 4 (top1, top2, dip1 and dip2) true unknown system response TOUGH2 model 8.00E-01 1.00E-01 6.00E-01 1.00E-02 1/a-vg_n 1/a-K vg_n-vg_n vg_n-K Slr-vg_n Slr-K K-vg_n stopping criteria Hazens m/s Gustafson m/s Top 1 Top 2 Dip 1 Dip 2 1.00E-03 Ks [m/s] 1.00E-04 0.00E+00 1.00E-05 -2.00E-01 1.00E-06 -4.00E-01 1.00E-07 -6.00E-01 1.00E-08 0 0.5 1 parameters calculated system response measured system response minimization algorithm best estimate of model parameters objective function maximumlikelihood theory Uncertainty propagation analysis by Empirical Orthogonal Eigenfunctions 1.00E+00 2.00E-01 K-K corrected parameter estimate prior information 1.00E+00 4.00E-01 Cpp The code iTOUGH2 (Finsterle, 1999) is used. The general inverse modeling procedure is illustrated in fig.3 . In this case the inverse problem is to estimate the parameter p in such a way that the residual vector r is minimized: rj y * j y j (p) Figure 9. Difference between observed and calculated liquid saturation Figure 3. Inverse modeling procedure Figure 11. Cpp reproduced by EOFsimulation 1.5 2 2.5 3 3.5 4 depth [m] Figure 10. Observed and estimated hydraulic conductivities a posteriori error analysis Conclusions: uncertainty propagation analysis • EOF simulation reproduces Cpp, and thereby automatically avoides unlikely parameter combinations Given the covariance matrix Cpp of the best-estimate parameter set p (Carrera and Neuman, 1986): Cpp s02 JT C ZZ1 J • EOF simulation does not rely on second order stationarity Conditional simulation of the parameter set p can be done according to the Proper Orthogonal Decomposition Theorem (Loève, 1977): • Truncation of p reduces the quality of Cpp reproduction 1 p n k 1 k • Geological architecture is critical βk • Liquid saturation data can be used to estimate optimal parameters for flow simulation, but a priori information is necessary where b is the eigenvector derived from: CppβkT kβk k 1,..., n and: E{ k j } kj k k, j 1,...,n where is the eigenvalue, and ij = 0 if i j, else 1. Figure 12. Improved simulation by EOF • Non-steady infiltration improves parameter estimation mm/d The purpose of this study is to: • Estimate sedimentological architecture and liquid saturation by Ground Penetrating Radar. p47