The predictions from models that inform decision making in environmental management and assess the risk of failing to achieve a desired environmental outcome must include estimates of the uncertainties associated with these predictions. These estimates of uncertainty are based in part on expert knowledge as expressed through the construction of the model, its boundary conditions, and its parameterization. This knowledge is typically defined stochastically, as is the nature of expert knowledge in environmental systems. Uncertainty information is also encapsulated in the historical behaviour of the system and is incorporated through the history-matching process. Models more reliably predict the range of future environmental outcomes if they have sufficient “receptacles” for these sources of information.
It follows that a model used in decision-making requires three components: a numerical model, a parameter estimation process wrapped around that model for history matching, and a model predictive uncertainty estimation process wrapped around these components. When considering these components, Kitanidis (2016) suggests a paradigm shift from models as simulators to models as receptors for data important to the decision-making process. He defines a good model as one which quantifies uncertainty and supports a comprehensive risk based evaluation of alternatives. Similarly, Nowak et al (2012) apply statistical methods to test the fitness for purpose of models in respect of decision making and the worth of decision-critical data used to inform these models.
More complex models provide “receptacles” for a greater amount of information. Unfortunately, we also observe that the long run times of complex models preclude running the model the number of times required for proper stochastic analysis. These same run times, accompanied by a tendency for numerical instability, often make the task of history-matching very difficult, if not impossible. We explore the gains and costs of model simplicity in the context of decision fitness for purpose and propose some metrics to be used in this analysis.