Sequential Bayesian Geostatistical Inversion and Evaluation of Combined Data Worth for Aquifer Characterization at the Hanford 300 Area

Subsurface characterization for predicting groundwater flow and contaminant transport requires us to integrate large and diverse datasets in a consistent manner, and quantify the associated uncertainty. In this study, we sequentially assimilated multiple types of datasets for characterizing a three-dimensional heterogeneous hydraulic conductivity field at the Hanford 300 Area. The datasets included constant-rate injection tests, electromagnetic borehole flowmeter tests, lithology profile and tracer tests. We used the method of anchored distributions (MAD), which is a modular-structured Bayesian geostatistical inversion method. MAD has two major advantages over the other inversion methods. First, it can directly infer a joint distribution of parameters, which can be used as an input in stochastic simulations for prediction. In MAD, in addition to typical geostatistical structural parameters, the parameter vector includes multiple point values of the heterogeneous field, called anchors, which capture local trends and reduce uncertainty in the prediction. Second, MAD allows us to integrate the datasets sequentially in a Bayesian framework such that it updates the posterior distribution, as a new dataset is included. The sequential assimilation can decrease computational burden significantly. We applied MAD to assimilate different combinations of the datasets, and then compared the inversion results. For the injection and tracer test assimilation, we calculated temporal moments of pressure build-up and breakthrough curves, respectively, to reduce the data dimension. A massive parallel flow and transport code PFLOTRAN is used for simulating the tracer test. For comparison, we used different metrics based on the breakthrough curves not used in the inversion, such as mean arrival time, peak concentration and early arrival time. This comparison intends to yield the combined data worth, i.e. which combination of the datasets is the most effective for a certain metric, which will be useful for guiding the further characterization effort at the site and also the future characterization projects at the other sites.