EVALUATING THE PREDICTIVE UNCERTAINTY FOR THE REACTIVE TRANSPORT OF URANIUM IN GROUNDWATER

Reactive transport simulations provide a systematic framework for integrating hydrologic and biogeochemical conceptual process models into a quantitative description of subsurface behaviors. However, subsurface environments are open and complex and subject to multiple interpretations and conceptualizations. The approach of this research is to postulate multiple plausible conceptual models, calibrate each model to observations and then make predictions with the calibrated model ensemble. Parametric and conceptual model uncertainty of the alternative models is jointly evaluated using a maximum likelihood formulation of Bayesian Model Averaging (BMA) (Neuman, 2003). The BMA method is being used with laboratory and field observations to test the hypotheses that model uncertainty dominates parametric uncertainty and that BMA improves predictive performance. The methodology is being applied to the Naturita UMTRA site to investigate how conceptual model uncertainty varies across scales ranging from column tests to the plume scale.

The approach was applied to a set of previously published results on U(VI) transport in columns packed with well-characterized quartz. As in the initial study (Kohler and others, 1996), seven different surface complexation models of varying complexity were calibrated against selected experiments with UCODE_2005 (Poeter and others, 2005) and then used to predict U(VI) transport in four different experiments conducted with different experimental conditions. It was found that, of the seven postulated models, only three had a significant probability. Model probabilities calculated from calibration to different datasets were different. The probability of one conceptual model varied from 22 to 50 percent when calibrated to each individual experiment; however, the model probability was only 16 percent when all three experiments were used simultaneously in the calibration. Model uncertainty significantly exceeded parametric uncertainty even in these well-controlled laboratory experiments and model averaging gave significantly superior predictions relative to any single model. The approach is being applied to results from small-scale tracer tests conducted at the Naturita site. The models for the tracer tests include alternative representations (1) of the subsurface heterogeneity of hydraulic conductivity, (2) of the U(VI) adsorption reactions, (3) of the rate of adsorption and desorption and (4) of geochemical processes affecting the key major ions such as calcium.

Neuman, S.P., 2003. Maximum likelihood Bayesian averaging of alternative conceptual-mathematical models: Stochastic Environmental Research and Risk Assessment: v 17, p 291-305.

Kohler, M., Curtis, G.P., Kent, D.B, and J.A. Davis, 1996, Experimental investigation and modeling of uranium(VI) transport under variable chemical conditions: Water Resources. Research, v. 32, p 3539-3551.

Poeter, E.E., Hill, M.C., Banta, E.R., Mehl, S., Christensen, S., 2005, UCODE_2005 and six other computer codes for universal sensitivity analysis, calibration, and uncertainty evaluation constructed using the JUPITER API: U.S. Geological Survey Techniques and Methods 6-A11, 299 p.