Center for Groundwater Science Seminar

(Cookies and coffee available at 2:45 pm)

Dr. Yonas Demissie

Civil and Environmental Engineering

University of Illinois at Urbana-Champaign

       It is widely recognized that groundwater modeling results are subject to uncertainty that stem from various sources including natural randomness, uncertain model inputs, parameters, boundary conditions and conceptualizations. Despite this recognition, groundwater models are often calibrated using the least-squares regression techniques, implemented through tools such as PEST or UCODE, which assume that parameter estimation error is the primary source of model uncertainty. Ignoring the remaining sources of uncertainty from calibration process can seriously affect the model predicting ability, potentially yielding biased and misleading results. In particular models forcing terms such as irrigation and the associated recharge from return flow are often uncertain because of lack of reliable water use records for irrigation. The first part of my presentation will focus on demonstrating the effect of irrigation pumping uncertainty on the estimated groundwater model parameter and prediction results using both analytical and experimental groundwater pumping examples. The impact is further evaluated in terms of types of dataset used for calibration. Such analysis is important as groundwater models are often calibrated using data collected during non-irrigation seasons and might not reflect the hydrogeological condition resulting from major stress. We have developed alternative parameter estimation techniques which take into account model input uncertainty, and compared the results with that of the standard least-squares method.

       In the second part of the presentation, I will describe a hybrid or complementary modeling approach to handle some of the systematic errors and uncertainties arising mainly from ignored or misrepresented processes in the groundwater model. The approach involves integration of physically-based groundwater model, such as MODFLOW, with error-correcting data-driven models, such as Decision-Trees. The data-driven models were separately developed to map the MODFLOW head and baseflow prediction errors, which were subsequently used to update the head and baseflow predictions at existing and proposed observation wells. The approach was evaluated using the Republican River Compact Administration (RRCA) groundwater modeling study, which computes the impact of irrigation pumping on baseflow and is used by policy makers to make water allocation decisions. We demonstrated how techniques from artificial intelligence/machine learning can help to reduce the systematic error and uncertainty and there by improve the predicting ability of the calibrated groundwater models. The combined physically-based and data-driven model resulted to more reliable and less biased forecasts and thus to more technically defensible policy decisions.