USE OF STOCHASTIC METHODS TO IMPROVE GROUND-WATER MODEL CALIBRATION AND EVALUATE MODEL-PREDICTION UNCERTAINTY, CAMP EDWARDS, CAPE COD, MASSACHUSETTS

Live-fire training activities and munitions disposal at Camp Edwards on the Massachusetts Military Reservation (MMR), Cape Cod, Massachusetts has resulted in the release of contaminants into the underlying sand and gravel aquifer, and there is concern that migration of contaminants from the installation could adversely affect water-supply resources. Contaminant sources at the site generally are not well characterized and large, readily mapped plumes generally do not occur. As a result, there is a reliance on numerical models to characterize ground-water flow and advective transport at the site. Deterministic models calibrated by trial-and-error methods by different groups reasonably match observed data; however, the models often yield different predictions of advective transport and do not provide measures of uncertainty for the model predictions.

Inverse (stochastic) calibration techniques using MODFLOW-2000 were applied at the site to 1) quantify formal sensitivities of observations to input parameters, 2) estimate the optimal parameter values that yield the best statistical fit to observed data, and 3) evaluate uncertainties associated with predictions of advective transport. The results indicate that formal sensitivity and parameter-estimation techniques were effective at 1) enhancing understanding of model sensitivities and potential biases, 2) improving model calibration to heads, flows, and plume paths, and 3) quantifying uncertainties for predictions of advective transport. Absolute mean residuals for heads, a traditional measure of model fit, improved from 1.73 to 1.41 feet. Calibration to measured stream flows also improved. The use of inverse calibration techniques allowed for a reasonable fit to an observed contaminant plume that had not been simulated properly by the deterministic model.

Inverse methods can be used to develop 70-, 95-, and 99.7-percent linear confidence intervals on predictions of forward and reverse particle tracks as well as contributing areas to production wells. The confidence intervals can be readily illustrated as geographically mapped quantities that can be understood by a lay audience. The results suggest that the use of inverse methods provide decision makers with more accurate model predictions as well as useful measures of prediction uncertainties.