HR: 13:40h
AN: NS13A-01    [Abstracts]
TI: A Comparison Between Deterministic and Stochastic Methods for Inverting Spectral Induced Polarization Data
AU: * Chen, J
EM: jchen@lbl.gov
AF: Lawrence Berkeley National Lab Earth Sciences Division, MS 90-1116 1 Cyclotron Road, Berkeley, CA 94720, United States
AU: Kemna, A
EM: a.kemna@fz-juelich.de
AF: Agrosphere Institute (ICG-IV), Forschungszentrum Julich 52425 Julich, Julich, 52425, Germany
AU: Hubbard, S
EM: sshubbard@lbl.gov
AF: Lawrence Berkeley National Lab Earth Sciences Division, MS 90-1116 1 Cyclotron Road, Berkeley, CA 94720, United States
AB: Cole-Cole model parameters (e.g., chargeability and time constant), extracted from spectral induced polarization (SIP) data, are being increasingly used to characterize subsurface properties. However, fitting Cole-Cole models (especially nested Cole-Cole models) to SIP data is challenging because of nonlinearity and non-uniqueness of the Cole-Cole models. This study compares conventional deterministic approaches (i.e., iterative based estimation methods) with Markov chain Monte Carlo (MCMC) based stochastic approaches for estimating Cole- Cole model parameters. The results of those case studies show that although deterministic methods are able to provide single optimal solutions under certain criteria (e.g., the least squares of misfit) and require minimal computing power, they suffer from two main limitations. The first limitation is that the optimal solutions heavily depend on the choice of the initial values. Different initial values may yield different inversion results, and in many cases, the deterministic methods even cannot converge for the chosen initial values. The second limitation is that those methods provide inadequate or inaccurate information about uncertainty in the estimation. On the contrary, the MCMC-based stochastic approaches are insensitive to the choice of the initial values and can provide extensive information about uncertainty in the estimation. From the drawn large number of samples, we can obtain exhaustive information about unknown parameters, such as the mean, the median, the mode, and even entire probability distribution of each unknown Cole-Cole model parameter. Although MCMC-based stochastic methods typically require that the forward models be run for thousands of times, this is not an issue given the current computer power. Through presentation of extensive synthetic and laboratory case studies, we will illustrate the benefits of the different methods when used individually and in combination with each other.
DE: 0520 Data analysis: algorithms and implementation
DE: 0644 Numerical methods
DE: 1835 Hydrogeophysics
DE: 3260 Inverse theory
DE: 3275 Uncertainty quantification (1873)
SC: Near-Surface Geophysics [NS]
MN: 2007 Fall Meeting