Verification-Based Model Tuning

All numerical models (e.g., Numerical Weather Prediction models) have certain parameters within model algorithms which effect forecasts to a different degree, depending on the forecast quantity. The specific values of these model parameters are determined either theoretically using fundamental physics laws but incorporating necessary approximations to reduce computational cost, or empirically using observations from field experiments where observational error introduces uncertainty. In either case, the exact value of the parameter is often unknown a priori, and so their values are usually set to improve forecast quality through some form of forecast verification. Such an approach to model tuning, however, requires knowledge of the observations to which the forecasts must be compared, and therefore, a multitude of highly detailed experimental cases in order to fully resolve parameter values, a data set very difficult to obtain. A knowledge of the relationship between model parameters and forecast quantities, without reference to observations, can not only aid in such an observation-based approach to model tuning, it can also aid in tuning the model parameters according to other criteria that may not be based on observations directly, e.g., a desire to affect the forecasts according to some longterm experience of a forecaster. The main goal of our work has been to develop a framework for representing the complex relationship between model parameters and forecast quantities, without any reference to observations.