Impact of numerical artifact of the forward model in the inverse solution of density‐dependent flow problem

Abstract

Abstract We investigate the effect of computational error on the inversion of a density‐dependent flow and transport model, using SEAWAT and UCODE‐2005 in an inverse identification of hydraulic conductivity and dispersivity using head and concentration data from a 2‐D laboratory experiment. We investigated inversions using three different solution schemes including variation of number of particles and time step length, in terms of the three aspects: the shape and smoothness of the objective function surface, the consequent impacts to the optimization, and the resulting Pareto analyses. This study demonstrates that the inversion is very sensitive to the choice of the forward model solution scheme. In particular, standard finite difference methods provide the smoothest objective function surface; however, this is obtained at the cost of numerical artifacts that can lead to erroneous warping of the objective function surface. Total variation diminishing (TVD) schemes limit these impacts at the cost of more computation time, while the hybrid method of characteristics (HMOC) approach with increased particle numbers and/or reduced time step gives both smoothed and accurate objective function surface. Use of the most accurate methods (TVD and HMOC) did lead to successful inversion of the two parameters; however, with distinct results for Pareto analyses. These results illuminate the sensitivity of the inversion to a number of aspects of the forward solution of the density‐driven flow problem and reveal that parameter values may result that are erroneous but that counteract numerical errors in the solution.