A COMPARISON OF MODEL CALIBRATION METHODS FOR GROUNDWATER MODELS

The calibration very large and complex groundwater models is becoming common as a means to help address issues of reliability and uncertainty. Models with many parameters might require thousands of model runs to achieve an acceptable calibration. In addition, larger basin scale models often take hours to run. Obviously, the efficiency of the calibration method can be crucial to practical calibration of these large models. Model-independent estimation packages such as PEST (Doherty, 2005) that are based on the Gauss-Newton-Levenberg-Marquardt (GNLM) method provide inverse modeling capabilities with considerable flexibility in choosing parameters and observations. However, when dealing with highly nonlinear problems, they may converge to a local, rather than global minimum. Vrugt at al. (2003a, b, c) presented a global optimization algorithm, the improved Shuffled Complex Evolution Metropolis method (SCEM and MOSCEM), for single- and multi-objective calibration of hydrologic models. The SCEM algorithm is a general purpose optimization algorithm that uses adaptive Markov Chain Monte Carlo (MCMC) sampling to provide an efficient search of the parameter space. In this study, we compare the local and global methods on several different synthetic groundwater models ranging from a layered basin model to a complex unconfined model. Algorithms are compared on a basis of computational efficiency and robustness of the solution.