Model-based decision making: A new perspective

It is understood that models make imperfect predictions, especially when applied to complex, heterogeneous, groundwater systems. However, given recent advances in parameter estimation and uncertainty analysis methodologies, models can be used to both quantify their predictive uncertainty through optimal processing of all available site data and human expertise, and to reduce that uncertainty to its theoretical minimum. In doing this, levels of confidence can be ascribed to the likelihoods that certain events will or will not occur. These "events" can include complex formulations that are pertinent to adaptive environmental management (for example that an unwanted event does not occur without sufficient advanced warning). This can be used as a basis for optimal design of a monitoring network, and/or for assigning thresholds that will trigger revised management plans. A model provides effective decision support when it is designed to explore uncertainty and to test hypotheses pertaining to future environmental behavior under proposed management plans. A hypothesis can be rejected if it has a low likelihood of occurrence, which is defined as when such an event is incompatible with available system hard and soft data. Hard data are composed of empirical point measurements of a system's properties as well as historical measurements of system state; soft data comprise inferred knowledge and expertise based on multi-year site studies or data from other sites. Event likelihood can be tested by explicitly incorporating its occurrence into the calibration dataset before proceeding with model inversion - preferably in a highly parameterized setting where parameter heterogeneity can reflect hydraulic property heterogeneity necessary to achieve the hypothetical event. This gives the calibration process ample opportunity to admit the event by allowing the adjustable parameters the freedom to introduce necessary hydraulic property heterogeneity for event realization. By simultaneously imposing constraints that reflect the most complete understanding of the site (e.g., limits to parameter values), event realization becomes consistent with the best understanding of site condition and hydraulic properties. If the calibration process, formulated in this context as a hypothesis-testing process, demonstrates that an unwanted event cannot happen without requiring unrealistic parameterization or steeply degraded model-to-measurement misfit under calibration conditions, the hypothesized event can be assigned a low probability of occurrence. In this paper, this mode of model usage is demonstrated using a simple calibration and prediction exercise involving tracer transport in a heterogeneous medium. Assessment of predictive likelihood takes into account the imperfect nature of the model as a simulator of real-world environmental processes. It also takes into account expert knowledge of parameter values.