SUBSURFACE PARAMETER UNCERTAINTY QUANTIFICATION USING ENVIRONMENTAL TRACERS AND SURROGATE MODELING METHODS

Accurate forecasts of groundwater flow and subsurface chemical transport are essential to groundwater resource management and remediation efforts. Calibration of numerical subsurface flow and transport model parameters - for example, the spatially heterogeneous permeability and porosity fields - suffer from non-uniqueness that leads to uncertainties in subsurface system identification and subsequent forecasts. For computationally expensive reactive transport models, robust characterization of these parameter and predictive uncertainties are often intractable. In this work, we investigate surrogate modeling methods to perform subsurface parameter estimation and uncertainty quantification of reactive transport models that simulate environmental tracer transport at the hillslope scale. We train both Gaussian Process Regression and Deep Neural Network surrogate models with a synthetically generated groundwater flow and environmental tracer concentration data set, which includes dissolved SF₆ and CFC's, tritium, and ⁴He. We use Markov chain Monte Carlo calibration to estimate the permeability and porosity parameter fields and their associated uncertainties using both the high-fidelity reactive transport model and the trained surrogate models. We compare posterior predicted fields from both the reactive transport and surrogate models to assess the ability of the surrogate models to accurately replicate environmental tracer reactive transport results and parameter uncertainties. Through this comparison, we asses the ability of combining high-performance reactive transport modeling with surrogate modeling techniques to facilitate permeability and porosity field estimation and uncertainty quantification in groundwater modeling. These results can be used to help improve hydrogeologists' ability to assimilate environmental tracer information into numerical subsurface calibration procedures and improve subsurface system forecast accuracy.