Reproducing tailing in breakthrough curves: Are statistical models equally representative and predictive?

Breakthrough curves (BTCs) observed during tracer tests in highly heterogeneous aquifers display strong tailing. Power laws are popular models for both the empirical fitting of these curves, and the prediction of transport using upscaling models based on best-fitted estimated parameters (e.g. the power law slope or exponent). The predictive capacity of power law based upscaling models can be however questioned due to the difficulties to link model parameters with the aquifers' physical properties. This work analyzes two aspects that can limit the use of power laws as effective predictive tools: (a) the implication of statistical subsampling, which often renders power laws undistinguishable from other heavily tailed distributions, such as the logarithmic (LOG); (b) the difficulties to reconcile fitting parameters obtained from models with different formulations, such as the presence of a late-time cutoff in the power law model. Two rigorous and systematic stochastic analyses, one based on benchmark distributions and the other on BTCs obtained from transport simulations, are considered. It is found that a power law model without cutoff (PL) results in best-fitted exponents (aPL) falling in the range of typical experimental values reported in the literature (1.5 < alpha(PL) < 4). The PL exponent tends to lower values as the tailing becomes heavier. Strong fluctuations occur when the number of samples is limited, due to the effects of subsampling. On the other hand, when the power law model embeds a cutoff (PLCO), the best-fitted exponent (aCO) is insensitive to the degree of tailing and to the effects of subsampling and tends to a constant alpha(CO) approximate to 1. In the PLCO model, the cutoff rate (lambda) is the parameter that fully reproduces the persistence of the tailing and is shown to be inversely correlated to the LOG scale parameter (i.e. with the skewness of the distribution). The theoretical results are consistent with the fitting analysis of a tracer test performed during the MADE-5 experiment. It is shown that a simple mechanistic upscaling model based on the PLCO formulation is able to predict the ensemble of BTCs from the stochastic transport simulations without the need of any fitted parameters. The model embeds the constant alpha(CO) = 1 and relies on a stratified description of the transport mechanisms to estimate lambda. The PL fails to reproduce the ensemble of BTCs at late time, while the LOG model provides consistent results as the PLCO model, however without a clear mechanistic link between physical properties and model parameters. It is concluded that, while all parametric models may work equally well (or equally wrong) for the empirical fitting of the experimental BTCs tails due to the effects of subsampling, for predictive purposes this is not true. A careful selection of the proper heavily tailed models and corresponding parameters is required to ensure physically-based transport predictions.