Can more data make a model worse? A transport experiment in an exhaustively sampled sandstone slab

An advantage of numerical models over analytic methods is their ability to include field measurements, such as hydraulic conductivity (K) at measured locations. Poor model fit with observed conservative transport is typically attributed to lack of K measurements. Previous research on a 30 x 30 x 2.2 cm slab of Massillon sandstone has shown that despite including the finest K measurements (every 0.33 cm), traditional numerical models fail to capture significant details of observed solute transport. The average K values still cannot adequately represent the sub-measurement scale heterogeneities within the sandstone. The resulting spatially averaged K field violates the assumptions of the local advection-dispersion equation (ADE). We present a new method which uses statistically similar Monte Carlo realizations within typical numerical models that can reproduce of the early breakthrough and late tails seen in natural aquifers. Using MODFLOW and a particle tracking code, ensemble results using a local ADE and no conditioning points capture the observed tracer transport better than models using the 8,649 K measurements directly. The results imply the need for a non-local transport model which can account for the sub-grid heterogeneities, while still honoring the K measurements. Increasing model discretization using an anisotropic fractal scaling model while including all K measurements may also help improve model fit with the experimental results.