Formatted Title
The More the Merrier: Towards Improved Remedial Outcomes Using Ensemble Methods
Background/Objectives
Expenditures for soil and groundwater cleanup at over 300,000 sites in the United States may exceed $200 billion (not adjusted for inflation) by the year 2033 (USEPA 2004) – not an insignificant amount. Heterogeneous formations often underlie these contaminated sites with distinct geological facies of widely varying hydraulic conductivities. Over time the low-permeability facies in these formations can often behave/act as secondary contaminant sources to adjacent zones of higher permeability via back diffusion, prolonging remedial timeframes. Managers and stakeholders at these sites rely on various tools, including numerical models, to assess the impact of alternative strategies, predict future conditions, and manage remedial costs. Historically, parameters governing groundwater flow and solute transport in these models are fine-tuned in the calibration phase until the model reasonably replicates historical field measurements. These calibrated models are then run in “predictive mode” to forecast future outcomes. However, as Moore and Doherty (2005) demonstrate, a well ‑calibrated model does not necessarily decrease predictive uncertainty. Predictive uncertainty of models is further exacerbated by limited historical data, and incorrect geological conceptual site models (CSMs). If not quantified, these uncertainties can lead to significant increases in remedial timeframes and corresponding life cycle costs and, under certain conditions, unacceptable levels of risk to sensitive receptors or remedy failure.
Approach/Activities
Ensemble smoothers are increasingly used to condition groundwater parameter fields to aquifer state measurements. Rather than calibrating a single model that matches measured data, a host of “likely models” that encompass alternate CSMs, are calibrated, and then used in a predictive mode to bracket predictive uncertainty and quantify certainty of remedial outcomes. The ensemble smoother's documented success and computational efficiency motivated us to explore their use at contaminated sites with discrete geological facies. The Traveling Pilot Point Method (TRIPS) is a recent innovation where pilot points are used to define the geometry of discrete geological facies. This paper demonstrates a framework where the TRIPS method for representing discrete geological facies and ensemble smoothers for conditioning parameter ensembles are used for generating a “multiverse” of calibrated models, and corresponding predictions, thus bracketing predictive uncertainty at contaminated sites, leading to better remedial outcomes.
Results/Lessons Learned
The simple, pragmatic framework developed and presented herein demonstrates that an ensemble-based approach applied in conjunction with data acquisition efforts can ultimately lead to better remedial outcomes. The flow and transport problem analyzed in this paper has several similarities to real-world sites, namely facies with highly contrasting permeabilities (i.e., highly heterogeneous), decade-long solute transit times, and a limited understanding of the subsurface. The approach presented here can readily be applied to large, complex contaminated sites where the associated costs (particularly of ‘being wrong’) far outweigh the increased costs associated with generating an ensemble of likely models. Site managers, regulators, stakeholders, and decision-makers can thus make better informed decisions by evaluating a spectrum of likely outcomes rather than relying on a single, deterministic prediction. The framework includes periodic data collection efforts, which are then used to update the predictions and further reduce uncertainties. Further, data worth analyses which have shown to reduce predictive uncertainty, can be applied to decide the locations of future monitoring wells.