Linear and non-linear response to parameter variations in a mesoscale model

Abstract

Parameter uncertainty in atmospheric model forcing and closure schemes has motivated both parameter estimation with data assimilation and use of pre-specified distributions to simulate model uncertainty in short-range ensemble prediction. This work assesses the potential for parameter estimation and ensemble prediction by analysing 2 months of mesoscale ensemble predictions in which each member uses distinct, and fixed, settings for four model parameters. A space-filling parameter selection design leads to a unique parameter set for each ensemble member. An experiment to test linear scaling between parameter distribution width and ensemble spread shows the lack of a general linear response to parameters. Individual member near-surface spatial means, spatial variances and skill show that perturbed models are typically indistinguishable. Parameter—state rank correlation fields are not statistically significant, although the presence of other sources of noise may mask true correlations. Results suggest that ensemble prediction using perturbed parameters may be a simple complement to more complex model-error simulation methods, but that parameter estimation may prove difficult or costly for real mesoscale numerical weather prediction applications.