A comparison of deterministic and stochastic approaches for regional scale inverse modeling on the Mar del Plata aquifer

• Modeling of flow and transport in real regional aquifers.
• Relative merits of deterministic and stochastic approaches: regional-scale calibration.
• Impact of using T data from large and small-scale tests on the calibration process.

SummaryInversion of the spatial variability of transmissivity (T) in groundwater models can be handled using either stochastic or deterministic (i.e., geology-based zonation) approaches. While stochastic methods predominate in scientific literature, they have never been formally compared to deterministic approaches, preferred by practitioners, for regional aquifer models. We use both approaches to model groundwater flow and solute transport in the Mar del Plata aquifer, where seawater intrusion is a major threat to freshwater resources. The relative performance of the two approaches is evaluated in terms of (i) model fits to head and concentration data (available for nearly a century), (ii) geological plausibility of the estimated T fields, and (iii) their ability to predict transport. We also address the impact of conditioning the estimated fields on T data coming from either pumping tests interpreted with the Theis method or specific capacity values from step-drawdown tests. We find that stochastic models, based upon conditional estimation and simulation techniques, identify some of the geological features (river deposit channels and low transmissivity regions associated to quartzite outcrops) and yield better fits to calibration data than the much simpler geology-based deterministic model, which cannot properly address model structure uncertainty. However, the latter demonstrates much greater robustness for predicting sea water intrusion and for incorporating concentrations as calibration data. We attribute the poor performance, and underestimated uncertainty, of the stochastic simulations to estimation bias introduced by model errors. Qualitative geological information is extremely rich in identifying large-scale variability patterns, which are identified by stochastic models only in data rich areas, and should be explicitly included in the calibration process.