Radioactive Contaminanant Transport in Subsurface Porous Environment

An Jin and Shoou-Yuh Chang

Abstract The deterministic model is developed based on the discrete advection-dispersion-adsorption equations with radioactive decay mechanism. The data assimilation scheme is designed to utilize the Kalman Filter (KF) to incorporate the knowledge of uncertainties in both the model and the measurement. Spatially correlated regional noise structures are proposed and integrated into a data assimilation scheme. This model demonstrates that the data assimilation scheme reduces the uncertainty and predicts more accurately than a deterministic model. Through absorbing information from observation, the predictive plumes of radioactive contamination from the assimilation system can follow the change of a randomized irregular plume shape in the real world more closely than a non-assimilation deterministic model.

Introduction
Radioactive agent has its decay chain. To simulate the complete transport process of the radioactive agent and the corresponding decay products in the subsurface environment, this study introduced the retardation factor to simplify the interactive effects between the porous materials and the dispersion, advection and decay, leading no explicit absorption term in the transport equation. Subsurface properties uctuate in space and time (Heuvelink and Webster, 2001). The subsurface environment is more heterogeneous and anisotropic compared to the atmosphere and surface water environment. Due to natural intrinsic uncertainties as well as the infeasible measurability, the prediction is inevitably uncertain and some random variables must be introduced to represent the unpredictable variation. Therefore, stochastic data assimilation such as Kalman lter (KF) can be integrated with deterministic models to give better predictions. Spatially independent Gaussian noise is easy to obtain and is widely assumed in many areas for examining KF properties. However,
A. Jin (B) North Carolina A&T State University, Greensboro, North Carolina 27411, USA

G.A. Uzochukwu et al. (eds), Proceedings of the 2007 National Conference on Environmental Science and Technology, DOI 10.1007/978-0-387-88483-7 24, C Springer Science+Business Media, LLC 2009

181

182

A. Jin and S.-Y. Chang

for subsurface modeling, it is more reasonable to consider the data and noise as regional i.e. spatially correlated instead of spatially independent Gaussian (Conwell et al., 1997; Goovaerts, 1999; Brooker, 2001; Ferreyra et al., 2002). In this study, a radioactive contaminant transport model with KF is constructed. Then a random eld simulation scheme is utilized to examine the impact of spatially correlated noise on the effectiveness of the data assimilation results.

Approach Radioactive Contaminant Plume in Porous Subsurface Environment

For radioactive contaminants transport in subsurface environment, contributions by nuclear actions of the radioactive agent and sorption caused by porous medium must be taken into considerations. This radioactive decay is a rst-order kinetic process with decay rate k. By adding absorption and decay mechanisms to the general advection-dispersion equation, we obtained C =D t 2C x2 V C x S t kC

where C is the concentration of contaminant (mass of solute per unit mass of solution), Dx and Dy are the longitudinal and transverse dispersion coefcients, respectively, V is the average linear pore liquid velocity, is bulk density of the porous medium, is the porosity, S is the contaminant concentration in the sorbed phase and k is the rate of the radioactive decay. In practice, we can simplify the equation by assuming linear partitioning between the C and S and introducing retardation factor R: Dx C = t R 2C x2 + Dy R 2C y2 V R C x 0.693 C R L hal f (1)

where Lhalf is the half-life of the radioactive decay and R is the retardation factor R =1+ KP

where we assumed Kp is linear partition coefcient.

Stochastic Markov Chain

Stochastic Markov chain is used for simulating the complicated real-world transport processes which usually have unknown or uncertain sources, inaccurate

Radioactive Contaminanant Transport in Subsurface Porous Environment

183

transport parameters or numeric errors. The stochastic dynamics can be represented as deterministic dynamics of transport and a random noise term Xk+1 = Axk + pk , k = 0, 1, 2, 3, . . . (2)

where xk and xk+1 are the subsurface state variables at time k and k+1, respectively. A is the nite-difference operator that advances the xk by one time step. pk is the model system error, the difference between the model prediction and the optimal estimate of the true value. The model error vector, pk , is assumed to have covariance matrix Qk and zero mean. Observation zk is described by data conversion matrix H and error ok . Similarly, ok is assumed to be the error of observations that has covariance matrix Rk and zero mean: Zk = HxT k + ok (3)

Optimal Estimation and Data Assimilation with KF

Considering a stochastic process dened by Equations (2) and (3) above, the Kalman optimal estimator (Kalman and Bucy, 1961) is Xk+1 ( + ) = xk+1 () + Kk+1 (zk+1 Hxk+1 ()) (4)

where xk+1 (+) indicates the estimated value after the KF adjustment, and xk+1 () the value before the KF adjustment, i.e. the predicted value from the model. The matrix Kk+1 is determined by Kk+1 = Pk+1 ()HT (HPk+1 ()HT + Rk+1 )1 (5)

where Pk+1 is the optimal estimate error covariance matrix, MT and M1 denote the transpose and inverse of matrix M, respectively. Pk+1 can be obtained by recursive calculating Pk+1 (+) = Pk+1 ()Pk+1 () HT [H Pk+1 () HT + R]1 H Pk+1 () and Pk+1 () = APk (+)AT + Qk (6)

(7)

184

A. Jin and S.-Y. Chang

Regional Noise Field Simulation

System noise is usually denoted as a vector pk where pk is q-by-1 column vector. The statistical structure of Gaussian noise with time independence is usually described as E{ pk } = 0 and E{ pk pl } = Q r kl , kl = 0 if k = l, kl = 1 if k = l (9) (8)

where Q is q-by-q symmetric and positive semi-denite. Q is a diagonal matrix if noise is spatially independent Gaussian. Since the noise will be integrated into KF, the most convenient way is to create the noise that satises given error covariance structures or given semi-variogram structures. Let pk (t) be a 2-D random noise state vector, t G s 2 , with given covariance matrix Q(h) that represents a certain spatial dependence of the pk (t) Q i j = Q(t j ti ) = Q(h i j ) (10)

where t is space location index, Gs is grid of system noise injection, hij is distance between the space location ti and tj . Assuming that the spatially-dependent pk (t) has a reasonable decomposition with components of a spatially-independent unit Gaussian vector gk (t) and a linear transformation L pk (t) = Lgk (t) (11)

with E(gg T ) = I (Chang and Jin, 2005), the linear transformation matrix L can be solved as the lower triangular of Cholesky decomposition of Q if Q is positive denite. i.e. Q = L LT (12)

Experiments
The model grid Gm is dened on a 2-D plane domain. Gm has nx grid points in x direction and ny grid points in y direction. In the numerical scheme, we set nx = 20, ny = 11, dx = 1.5 m, dy = 1.5 m, dt = .20 day. We assumed that horizontal dispersion Dx = 1.5 m2 /day, Dy = .25 m2 /day, porosity = .40, velocity = 1.5 m/day along the x direction and retardation = 1.525. The initial condition is an instantaneous contaminant source injected into a location with coordinates (5,5). Initial contamination background is set to 10 ppm.

Radioactive Contaminanant Transport in Subsurface Porous Environment

185

Since the transport process and its predictions are actually occurring in a geographically specic subsurface area, it is obvious that to utilize a regional error frame to simulate the uncertainty is more reasonable than to use a geo-independent error frame. Root Mean Squared Error (RMSE) was chosen as the error parameter to evaluate the KF effectiveness on the whole domain. RMSE(k) is dened as 1 Ni j 1

RMSE (k) =

[C E (i, j, k) C(i, j, k)]2

i, j

(13)

Results Prediction with Half-life 216 Day and Inaccurate Hydrologic Parameters
In practice, the sampling places and times are restricted and description of the hydrologic parameters can never be considered complete and accurate as required by subsurface transport models. More effort in continuously pursuing completeness and accuracy in the labs and sites may reduce the errors caused by these hydrological parameters, but this is usually very costly due to highly heterogeneous subsurface environment. Applying KF data assimilation scheme can provide another cost-effective way to improve accuracy by merging nite observations. In this experiment, we simulate the scenario that has one inaccurate hydrologic parameter in the prediction model and test the impact of regional noise R. For simplicity, the average linear pore liquid velocity, V, is assumed inaccurate in the prediction model. V is increased by 20% when the true elds were simulated. However, a deterministic model with the uncertain input still employs the given V in its prediction. To reect the regional nature of the pollutant in the subsurface environment, observation error matrix R is assumed to satisfy R[i,j] = E{o[i]o[j]} = c2 exp(hij /a), where hij is grid distance between grid point i and j, c and a are positive parameters that need to be specied in a particular model. The regional noise eld simulation scheme has been developed and examined by verifying experimental semivariogram of the generated noise elds. The prediction from the numerical model that has no data assimilation scheme is shown in Fig. 1. Note that the pollutant plume in the prediction eld moves behind the plume of the true eld due to the inaccurate V used in prediction model. Both Figs. 2a and 2b give KF ltered results of contaminant plume at 6.4 day. With the data assimilation scheme, the qualities of model predictions are improved. However, by comparing the two ltered results, the KF using regional noise structure (Fig. 2b) is able to follow the irregular shapes of true plume more accurately due to a more effective data assimilation process. In Fig. 2b, the prediction

186 Fig. 1 Prediction without KF Data Assimilation (Half-life = 216 day)

A. Jin and S.-Y. Chang

contours in the signicant pollution areas (with pollutant concentration greater than 100 ppm) follow the true contours more closely than what has shown in Fig. 2a, KF using Spatially Independence Gaussian(SIG) noise structure. The correction effect of different KFs can be seen more clearly in Fig. 3, in which the comparison of the average prediction errors (RMSE) for the three schemes, the deterministic model scheme without the KF lter, the model with the KF lter, and the KF model with the regional error structure are calculated at different time steps. As shown in Fig. 3, by applying the regional noise KF data assimilation scheme,

Fig. 2a Predictions improved by KF with SIG covariance R

Radioactive Contaminanant Transport in Subsurface Porous Environment Fig. 2b Predictions by KF with regional covariance R

187

RMSE is converged to 25 ppm at 6.4 day. This is a 37% decrease compared to the estimate from the model without KF and about 29% less compared to the model with KF using SIG noise structure. This correction process of KF involved in prediction is very effective in correcting the error that is caused by inaccurate hydrologic parameter V. Even without knowing which parameter is inaccurate, this correction process occurred by sampling from only nine sites and applying appropriate obser-

RMSE on each time step 40

35

30 RMSE (ppm)

25

20
non-KF prediction

15

KF prediction with SIG covariance KF prediction with regional covariance

10

10

15

20

25

30

35

Time Step k (t = 0.2 day)

Fig. 3 Prediction Error (RMSE) (halftime = 216 day)

188

A. Jin and S.-Y. Chang

vation error covariance matrix R in KF. As a comparison, Fig. 3 suggests that the stochastic model with KF data assimilation is more accurate than the non-KF deterministic model; moreover, the accuracy can be greatly improved by selecting the correct noise statistic structure in the KF scheme.

Conclusions
A hydrodynamic model is derived to simulate the radioactive contaminant in a subsurface porous environment. Errors arising from numerical models and from the uncertain geo-physical parameters can result in unavoidable prediction deviations. Test cases have examined the adaptive performance in the cases that having an inaccurate hydrologic parameter. The results indicate that the prediction error (RMSE) of the KF scheme using a regional noise covariance matrix is smaller than that from the models prediction without KF or that from the KF using a SIG noise matrix. The results from test cases also show the potential applications of a regional noise structure in environmental transport modeling. Applying a regional noise structure in KF data assimilation improved the KF correction process for inaccurate hydrologic parameters.

References
Brooker, P.I. (2001) Modelling spatial variability using soil proles in the Riverland of South Australia. Environmental International, 27(2001), 121126. Chang, S.-Y. and Jin, A. (2005). Kalman ltering with regional noise to improve accuracy of contaminant transport models. Journal of Environmental Engineering, 131(6), 971982. Conwell, P.M., Silliman, S.E. and Zheng, L. (1997). Design of a piezometer network for estimation of the variogram of the hydraulic gradient: The role of the instrument. Water Resources Research, 33(11), 24892492. Ferreyra, R.A., Apezteguia, H.P., Sereno, R. and Jones, J.W. (2002) Reduction of soil sampling density using scaled semivariograms and simulated annealing. Geoderma, 110(2002), 265289. Goovaerts, P. (1999) Geostatistics in soil science: state-of-the-art and perspectives. Geoderma, 89(1999), 145. Heuvelink, G.B.M. and Webster, R. (2001) Modelling soil variation: past, present, and future. Geoderma, 100(2001), 269301. Kalman, R.E., and Bucy, R.S. (1961) New results in linear ltering and prediction theory. Trans. ASME, Series D, Journal of Basic Engineering, 83, 95108.