A Monte Carlo uncertainty analysis of ozone trend predictions in a two-dimensional model

We use Monte Carlo analysis to estimate the uncertainty in predictions of t otal O-3 trends between 1979 and 1995 made by the Goddard Space Flight Cent er (GSFC) two-dimensional (2-D) model of stratospheric photochemistry and d ynamics. The uncertainty is caused by gas phase chemical reaction rates, ph otolysis coefficients, and heterogeneous reaction parameters which are mode l inputs. The uncertainty represents a lower bound to the total model uncer tainty assuming the input parameter uncertainties are characterized correct ly. Each of the Monte Carlo runs was initialized in 1970 and integrated for 26 model years through the end of 1995. This was repeated 419 times using input parameter sets generated by Latin hypercube sampling. The standard de viation (sigma) of the Monte Carlo ensemble of total O-3 trend predictions is used to quantify the model uncertainty. The 34% difference between the m odel trend in globally and annually averaged total O-3 using nominal inputs and atmospheric trends calculated from Nimbus 7 and Meteor 3 total ozone m apping spectrometer (TOMS) version 7 data is less than the 46% calculated l a model uncertainty, so there is no significant difference between the mode led and observed trends. In the northern hemisphere midlatitude spring the modeled and observed total O-3 trends differ by more than la but less than 2 sigma, which we refer to as marginal significance. We perform a multiple linear regression analysis of the runs which suggests that only a few of th e model reactions contribute significantly to the variance in the model pre dictions. The lack of significance in these comparisons suggests that they are of questionable use as guides for continuing model development. Large m odel/measurement differences which are many multiples of the input paramete r uncertainty are seen in the meridional gradients of the trend and the pea k-to-peak variations in the trends over an annual cycle. These discrepancie s unambiguously indicate model formulation problems and provide a measure o f model performance which can be used in attempts to improve such models.