Interpreting Parameter Interactions using Global Sensitivity Analysis on a Hillslope-scale Reactive Transport Model
Résumé
Reactive transport models (RTMs) are valuable tools for understanding, interrogating and forecasting subsurface processes, but the complex interplay between intricate flow paths and simultaneous reactions can make them difficult to optimize and interpret. Global sensitivity analysis (GSA) can alleviate these problems in several ways: by identifying the largest sources of uncertainty in model output, by assigning value of information to model input parameters – which, in turn, improves lab and field investigations through targeted data collection – and by clarifying the combined effects of complex parameter interactions. However, GSA is often neglected within the subsurface modeling community due to the high computational costs required to perform many model simulations.
Here, we demonstrate multiple applications of GSA to subsurface modeling by using distance-based generalized sensitivity analysis (DGSA1) to interrogate four separate components of a hillslope-scale RTM. DGSA requires fewer simulations than traditional GSA techniques, while cloud and cluster computing allow us to run 100s of model simulations simultaneously. The 2D model simulates the export of reduced species from pockets of fine-grained, organic-rich sediments embedded in a coarse sand aquifer. Stochastic simulations jointly varied 17 key model input parameters across several orders of magnitude, including the spatial variability of each parameter. DGSA on the model results reveal that the amplitude and variability of exported Fe(II) are most sensitive to the interaction between sand permeability and individual reaction rates. By contrast, the propagation of reducing conditions downgradient from the fine-grained lenses depends most heavily on the rate of dissolved organic carbon and sulfate release from the fines.
1Fenwick, D.; Scheidt, C.; Caers, J. Quantifying Asymmetric Parameter Interactions in Sensitivity Analysis: Application to Reservoir Modeling. Math Geosci 2014, 46 (4), 493–511.