Towards the Diagnosis and Simulation of Discrepancies in Dynamical Models

Models are frequently used to make predictions in regions where experimental testing is difficult. This often involves extrapolating to regions far from where the model was validated. In this paper an example is shown where, despite using a Bayesian analysis to quantify parameter estimation uncertainties, such an extrapolation performs poorly. It is then demonstrated that, in the presence of measurement noise, treating a system’s parameters as being time-variant (even if this is not believed to be true) can reveal fundamental flaws in a model. Finally, existing methods which can be used to quantify model error—the inevitable discrepancies that arise because of approximations made during model development—are extended towards dynamical systems.