An ongoing and interesting question for air-quality models of this nature is the extent to which their predictions are the result of the model architecture (that is, the particular choice of algorithms and operator order that make up the model itself) as opposed to factors such as the inputs to the model (the driving meteorology and the emissions), and the other choices such as the domain size, map projection, and grid spacing employed.
In the current work, predictions from the two models were statistically compared over a relatively high resolution domain (12km horizontal grid spacing), after the implementation of several factors designed to minimize the differences between the models' inputs:
- Both models were run using the same source of meteorological input (the Environment Canada Global Environmental Multiscale (GEM) weather forecast model). - The models were run on the same horizontal domain, map projection (polar stereographic) and grid spacing (12 km). - The models made use of the same emissions database and emissions processing system (2005 US and 2006 Canadian inventories, SMOKE emissions processing system, configured for the SAPRC99 (CMAQ) and ADOM-II (AURAMS) gas-phase mechanisms, respectively. - Biogenic emissions were created using the same emissions source (BEIS3.09 algorithms, speciated for each model).
The models were run for three separate periods, each comprising a month of simulation (excluding spin-up time): July 19 – August 19, 2004 (Summer2004), July 15 – August 25, 2005 (Summer2005), and January 28 – February 28, 2005 (Winter2005). The results of the models were compared to hourly network O3 and PM2.5 data for the same periods. AURAMS performed better than CMAQ for most ozone statistics. For example, during the Summer 2005 simulation, with : ~41800 station-hours compared, AURAMS' mean, minimum, y-intercept, slope, correlation coefficient, mean bias, root mean square error, normalized mean bias and normalized mean error were better than CMAQ; the observed ozone maximum was 100 ppbv, with CMAQ having 100.5 and AURAMS 100.8 ppbv, respectively). However, for PM2.5 (~8650 station-hours), CMAQ's scores for mean, maximum, minimum, y-intercept, mean bias, root-mean-square-error, normalized mean bias, and normalized mean error were better than AURAMS (AURAMS had a more accurate slope and correlation coefficient for PM2.5 than CMAQ).
These results led to a detailed investigation of the causes of the differences between the two models. Several factors were found to have a significant impact on the model results, including: (1) The imposition of a 1 m2s-1 lower limit on diffusivity in CMAQ, which reduced CMAQ's ability to capture nighttime O3 titration events and led to higher positive O3 biases and more negative PM2.5 biases than AURAMS (2) The use of a different vertical discretization and coordinates between the two models. (3) The order of numerical operators within the AURAMS model. (4) The accuracy of the underlying emissions data, which may affect the model results in different ways due to the differences in model architecture noted above. The spatial and temporal allocation factors for the top 20 primary PM2.5 and NO emitters were examined in detail, and a number of errors in the assignment of fields were identified. These resulted in several improvements in the emissions input fields.
Other improvements to the both the models and the emissions fields are underway concurrent with this project, and will be incorporated in a final set of simulations, to be evaluated and included as part of this presentation.
During the course of the comparisons, several improvements were made to the AURAMS model, some of which were subsequently tested in the Environment Canada operational air-quality model GEM-MACH (Global Environmental Multiscale – Modelling Air-quality and Chemistry). The GEM-MACH model makes use of much of AURAMS chemistry architecture, but in an on-line setting. These changes to GEM-MACH and their impact on the predicted forecast fields will be discussed.