1 INTRODUCTION The estimation of mass discharges (mass/time) from contaminated sites has received increased attention over the past decade. Such estimates are valuable when evaluating the potential risk to downgradient receptors such as water supply wells and surface water bodies, when assessing the efficiency of a site remediation, or when determining the degree of natural attenuation downstream of the source (Basu et al., 2006). Given the many applications of mass discharge estimates, a quantification of the associated uncertainties is important. However, uncertainties related to mass discharge estimates have not yet been given much attention. In the field, the mass discharge migrating from a contaminant source is typically determined across a control plane located downstream of the source and perpendicular to the mean groundwater flow. The uncertainty in a field-estimated mass discharge is highly related to the degree of heterogeneity of the mass flux distribution at this control plane. The more heterogeneous the mass flux distribution is, the finer the monitoring resolution network should be to ensure that the unmeasured areas in the control plane do not influence the estimate significantly (Li et al., 2007; Guilbeault et al., 2005). However, at most non-research field sites the number of monitoring wells is limited. The degree of heterogeneity of the mass flux distribution at the control plane is caused by several factors, where spatially and temporally varying flow conditions and complex contaminant distribution in the source zone are considered most important (Kubert and Finkel, 2006). A quantification of the mass discharge and the associated uncertainty should therefore account for all these factors. However, it is not easy to describe and model the influence of such factors at a specific site, especially when data are sparse. Often the knowledge about e.g. the source and the geological and hydrogeological settings is limited, which makes it very difficult to conceptualize these elements and to incorporate them in a model. The previous studies of mass discharge uncertainty have not taken the influence of different conceptual models into account. We present here a holistic and rigorous approach for quantifying the uncertainty in the mass discharge across a multilevel control plane given sparse field measurements. The new method is based on a Bayesian Monte Carlo modeling approach, where multiple models representing the site conditions are used to address the uncertainty and the ignorance related to the conceptualization. The outcome of the method is an ensemble of model realizations that all honor the observed data at the control plane, through which the mass discharge is to be determined. Uncertainty of mass discharge estimates from contaminated sites using a fully Bayesian framework M. Troldborg 1 , W. Nowak 2 , P. J. Binning 1 , P. L. Bjerg 1 & R. Helmig 2 1. Department of Environmental Engineering, Technical University of Denmark, Miljøvej B113, 2800 Kgs. Lyngby, Denmark 2. Institute of Hydraulic Engineering, Universität Stuttgart, Pfaffenwaldring 61, 70569 Stuttgart, Germany ABSTRACT: Mass discharge estimates are increasingly being used in the management of contaminated sites and uncertainties related to such estimates are therefore of great practical importance. We present here a rigorous approach for quantifying the uncertainty in the mass discharge across a multilevel control plane given sparse field measurements. The method accounts for i) conceptual model uncertainty through Bayesian model averaging, ii) heterogeneity through Bayesian geostatistics with an uncertain geostatistical model, iii) uncertainty in the source characterization and the solute transport parameters and iv) measurement uncertainty. Through Monte Carlo simulation, an ensemble of unconditional steady-state plume realizations is generated numerically. By use of the Kalman Ensemble Generator (similar to the Ensemble Kalman Filter), the parameter field realizations are conditioned on site-specific data. In this way a posterior ensemble of realizations, all honouring the measured data at the control plane, are generated for each of the conceptual models considered. The probability distribution of mass discharge is obtained by combining all ensembles via Bayesian model averaging. Elements of the method are illustrated with results from a synthetic test case.