Investigation of numerical error sources in coupled model predictions of atmospheric transport and dispersion

Numerical modeling is one of the chief means by which scientists study complex multi-disciplinary environmental systems. However, most models are largely “stovepipe” systems designed to predict only a very limited part of the earth-ocean-atmosphere-sun system. Even with modern technological tools, such as the Earth System Modeling Framework, to facilitate computationally efficient coupling of numerical models, there remain many unanswered questions about potential sources of error arising from coupling disparate component models. Time and space scales crucial to phenomena represented in models for different media are often very different, leading to potential sampling and interpolation errors. Moreover, there is considerable opportunity to introduce errors when each model requires information about similar parameterized variables associated with their interfaces (e.g., surface flux calculations in ocean and atmosphere models).

To study these possibilities we focus on a fairly simple coupled system. The MM5 mesoscale numerical weather prediction (NWP) model has been coupled to the SCIPUFF atmospheric transport and dispersion (AT&D) model. Using a case from the Cross Appalachian Tracer Experiment (CAPTEX), we investigate the role of temporal frequency in MM5 output on the accuracy of SCIPUFF's predicted plume concentrations versus observations of an inert tracer gas measured downwind by a surface monitoring array over a distance of ~1200 km. Next, we examine concentration errors resulting from horizontal and vertical spatial interpolations between the native MM5 grid and that of SCIPUFF. Finally, we study the role of better physical coupling by modifying SCIPUFF to use more of the non-state variables predicted by MM5 for surface fluxes and turbulence processes in the atmospheric boundary layer. Each experiment uses the same 36-h MM5 4-km simulation of the CAPTEX case, obtained by dynamic analysis using four-dimensional data assimilation, to drive SCIPUFF.

Experimental evidence demonstrates that substantial errors can occur in predicted surface concentrations for all three elements of the coupling between MM5 and SCIPUFF. In this case, the greatest errors were found to be related to coupling of the physics between the two models.