Sensitivity and Uncertainty Analysis of Physical Parameterization and Initial Conditions on Meteorological Variables and CO2 Mixing Ratios

Atmospheric inversions are used to assess biosphere-atmosphere CO2 surface exchanges. However, large uncertainties exist among inverse flux estimates independent of the spatial scales. Atmospheric transport model errors are one of the main contributors to the uncertainty affecting CO2 inverse flux estimates, but have not been quantified thoroughly. In the present work, atmospheric transport errors are evaluated over the Mid-Continental Intensive domain with an ensemble of simulations created with the Weather Research and Forecasting (WRF) mesoscale model using different physical parameterizations (e.g., planetary boundary layer (PBL) schemes and land surface models (LSMs), cumulus parameterizations and microphysics parameterizations). Modeled meteorological variables and atmospheric CO2 mixing ratios were compared to observations (e.g., radiosondes, AmeriFlux and CO2 mixing ratio towers) during the summer of 2008. Meteorological model-data mismatch of wind speed, wind direction and PBL height are used to identify systematically biased members, calibrate our ensemble and verify that we generated an ensemble that represents the transport errors over the region. We explore the sensitivity of atmospheric conditions to the choice of physical parameterizations to identify which govern resulting CO2 mixing ratios. Results show that the performance of the different model configurations is highly variable over the region, limiting the selection of an optimal configuration. However, we were able to identify the different physical schemes that were contributing to the systematic errors across the region. Sensitivity tests indicate that all physical schemes except but microphysics influence model variability in atmospheric CO2. Using the flatness of the rank histogram as a metric we were able to find a calibrated ensemble that represents transport errors for all three meteorological variables. Finally, we present an assessment of spatial error correlations from the calibrated ensemble to generate error covariances from filtered model error structures.