Journal Articles Mary C. Hill Comment on "Pilot point methodology for automated calibration of an ensemble of conditionally simulated transmissivity fields, 1, Theory and computational experiments" by B.S. RamaRao, A.M. LaVenue, G. deMarsily, and M.G. Marietta and "Pilot point methodology for automated calibration of an ensemble of conditionally simulated transmissivity fields, 2, Application" by A.M. LaVenue, B.S. RamaRao, G. deMarsily, and M.G. Marietta Richard L. Cooley and Mary C. Hill U.S. Geological Survey 2000, Water Resources Research, vol. 36, no. 9, p. 2795-2797 A method for stochastic modeling of ground-water flow systems using a combination of pilot point parameterization and conditional simulation was presented by RamaRao et al. [1995] and LaVenue et al. [1995]. (We will collectively term these two papers RLMM and term the method developed in RLMM the CS method here.) RLMM [pp. 478-479] state that the CS method is intended to provide a frequency distribution of possible alternative spatial transmissivity (T) distributions that (1) are statistically similar to the observed T distribution, (2) are equally likely given the calibration data, and (3) closely reproduce the measured pressures. The frequency distribution of transmissivities is then used to form frequency distributions of derived functions, such as travel times, which are summarized in the form of uncertainty measures, such as "confidence (or tolerance) intervals", on the derived functions [RLMM, p. 512]. These intervals are interpreted by RLMM as probabilistic intervals on the actual values, which in the case of travel times are values that could occur sometime in the future [RLMM, p. 513]. Cooley [2000] analyzes the RLMM method using linearization and bootstrap theory and concludes that the method can yield accurate uncertainty estimates, but only under some limited circumstances. In this comment we use Cooley's analysis to critique the method. We also identify and discuss some statements made by RLMM about model calibration and their methodology that appear to be misleading. It is unusual to comment on a paper so long after publication. Subsequent work has expanded on the method of RLMM, and the method was used advantageously in the testing documented by Zimmerman et al. (1998), so it is clear that the method has significant strengths. We go back to the 1995 papers for this comment, however, because they display most clearly the methodological difficulties with which we are concerned. Conclusions This comment on the conditional simulation method as presented by RamaRao et al. [1995] and LaVenue et al. [1995] reveals some potentially serious difficulties with the method: The method encourages overparameterization. The effects of the overparameterization are not accurately reflected in the uncertainty analysis. The method could yield poor approximations of prediction intervals because of bootstrap errors and neglecting the effects of unknown errors during the conditional simulations. The presentation in RLMM was not balanced in its comments about alternative parameterization methods. Thus, while RamaRao et al. [1995] and LaVenue et al. [1995] suggest an interesting method with some strengths, the method needs to be considered within the context of its weaknesses and other alternatives. No one method is suitable for all circumstances, and no method is without some weaknesses. Substantial work remains to thoroughly test existing and new methods of both calibrating ground-water models and evaluating the uncertainty of model results and predictions. mchill@usgs.gov Last Modified: August 14, 2000