Application of global methods for sensitivity analysis and uncertainty assessment of atmospheric chemistry transport models

Abstract

Atmospheric concentrations of air pollutants remain high, and air quality is still an issue that requires attention in many countries including the UK. Atmospheric chemistry transport models (ACTMs) are widely used to provide scientific support for policy development in relation to the mitigation of the detrimental effects of air pollution on human health and ecosystems. Hence it is important to assess the level of uncertainty associated with model predictions. In this work, the application of global sensitivity analysis and uncertainty assessment methods is investigated for two ACTMs of different complexity: the Fine Resolution Atmospheric Multi-pollutant Exchange (FRAME) model and the UK regional application of the European Monitoring and Evaluation Programme (EMEP4UK) model. For both models, the uncertainties in the outputs resulting from uncertainties in the model input emissions are quantified and apportioned. Additionally, the overall model response to variations in the input emissions within a ± 40 % range from the baseline is investigated as this range of variation is typically used for future scenario simulations. FRAME is a Lagrangian ACTM with 5 km x 5 km horizontal resolution over the UK domain that is used to estimate annual average concentrations and deposition of sulphur and nitrogen species. In the model, air columns with 33 vertical layers of varying thickness (from 1 m at the surface to 100 m at the top of the mixing layer) move from the boundary of the domain along straight-line trajectories with different starting angles at a 1° resolution. The model utilises annually averaged meteorology to define the column trajectories and rainfall. The chemical scheme includes gaseous-and aqueous-phase reactions. FRAME supplies Source-Receptor Relationship (SRR) matrices for the UK Integrated Assessment Model, which directly underpins UK air pollution control policies. EMEP4UK is a 3-D Eulerian model with a horizontal resolution of 5 km × 5 km over the British Isles and 20 vertical levels, extending from the ground to 100 hPa, which is also extensively used to inform UK air quality assessment. The chemical scheme implemented in the model is EmChem09 with the MARS equilibrium module for gas-aerosol partitioning of secondary inorganic aerosol. In addition to the pollutants modelled by FRAME, EMEP4UK is capable of modelling ozone (O3) and speciated particulate matter (PM2.5 and PM10) concentrations, all at hourly temporal resolutions. In this study, the uncertainty ranges for the input emissions from UK anthropogenic land-based sources were assigned according to the data provided by the UK National Atmospheric Emissions Inventory. For the FRAME model, the uncertainties in the outputs were propagated from the uncertainties in the emissions of SO2, NOx, and NH3. For the EMEP4UK model, an increased number of input variables was used; the emissions of NOx, SO2, NH3, VOC, and primary PM2.5 were split into 13 model inputs based on the contributions from different emission source sectors. The optimised Latin hypercube sampling design was used for both models to construct model runs that covered a chosen range of input emission perturbations. The FRAME model was investigated using several regression techniques. The response of the model to emission perturbations within a ± 40% range from the baseline value was found to be substantially linear. Surface concentrations of SO2, NOx, and NH3 together with the deposition of S and N were found to be predominantly sensitive to the emissions of the respective pollutant, while sensitivities of secondary species such as HNO3 and particulate SO42−, NO3−, and NH4+ to pollutant emissions were more complex and geographically variable. Additionally, the uncertainty in the surface concentrations of NH3 and NOx and the depositions of NHx and NOy was shown to be due to uncertainty in a single input variable, NH3, and NOx respectively. In contrast, the uncertainty in concentration and deposition of sulfur containing species were affected by the uncertainties in both NH3 and SO2 emissions. Similarly, the relative uncertainties in the modelled surface concentrations of each of the secondary pollutant variables were affected by the uncertainty range of at least two input variables. An emulator-based approach was used to propagate and apportion uncertainty in EMEP4UK outputs and investigate the model response to input perturbations. A separate Gaussian process emulator was used to estimate model predictions at unsampled points in the space of the uncertain model inputs for every modelled grid cell. For the surface concentrations of O3, NO2, and PM2.5 (pollutants associated with the adverse effects on human health) the highest level of uncertainty was found to occur in the grid cells comprising urban areas, up to ± 7%, ± 9%, and ± 9% respectively. However, overall uncertainty calculated for the land-based grid cells for the variables above was found to be low, which indicates that the outputs may be more sensitive to variation in other model input parameters, such as chemical or physical constants. Alternatively, UK land-based concentrations of O3, NO2, and PM2.5 may be dominated by the precursor emissions and long-range transport of pollutants from outside the UK. Investigating seasonal changes in uncertainty and sensitivity for the monthly-averaged model outputs allowed determination of the importance of the inputs that drive uncertainty changes throughout the year. For example, uncertainty in O3 was driven more by uncertainty in VOC emissions during the summer, and for PM2.5 the importance of NH3 in driving overall uncertainty increased during spring and summer. The aim of the global methods for sensitivity analysis and uncertainty assessment presented here is to quantify the confidence in model predictions associated with particular aspects of model operation. Furthermore, the model runs and emulators created for the analyses can be used to predict the ACTM response for any other combination of perturbed input emissions within the ranges set for the original Latin hypercube sampling design without the need to re-run the ACTM. This makes exploring different emission perturbation scenarios possible at a significantly reduced computational cost. The methods discussed in this study can be applied to any operational aspect of any ACTM.