A52C-03 A Novel Geographically Weighted Gaussian Process Regression (GW-GPR) Global Air Quality Emulator

Abstract

Air quality modelling has been an essential tool to study the impacts of socio-economic changes and policies on air quality, and to facilitate estimating associated social costs. However, its high computational and human resource demands limit its usage outside of the atmospheric science community. We create an air quality model emulator to address this issue, by training Geographically Weighted Gaussian Process Regressors (GW-GPR) from the outputs of a series of chemistry-climate perturbation experiments from a high-fidelity global chemical transport model, GEOS-Chem High Performance (GCHP). The GW-GPR relates changes in annual mean surface anthropogenic PM_2.5 and 6-monthly mean of Daily Maximum 8-hour Average (6mMDA8) O₃ to changes in short-lived air pollutant emissions and surface CH₄ concentration and global warming level (parameterized as CO_­2 concentration) at gridcell level. In comparison to existing linearized and regionalized approaches, our method can account for sub-regional changes in air pollutant emission patterns, and the non-linear response of secondary air pollutants to precursor and greenhouse gas emissions. We evaluate and demonstrate the utility of our model by predicting the air pollution levels at 2050 under 4 sets of climate and air pollution control policy scenarios. The GPR successfully emulates the regional area- and population-weighted mean PM_2.5 levels, and the associated changes in premature mortalities from GCHP simulations, while requiring only 4 – 6 seconds of CPU time (~3000 hours for GCHP) for each scenario. The accuracy, speed and simplicity of our GPR emulator show the capability of machine learning algorithms in emulating global air quality models and make air quality modelling accessible for global climate/air pollution scenario analysis and integrated assessment. We also discuss difficulties of emulating 6mMDA8 O₃ and potential reasons for this, and provide recommendations for ways forward.

Plain-language Summary

Air quality modelling has been an essential tool to study the impacts of socio-economic changes and policies on air quality, and to facilitate estimating associated social costs. However, its high computational and human resource demands limit its usage outside of the atmospheric science community. We create an air quality model emulator to address this issue, by training a machine learning emulator from the outputs of a series of air quality model runs to predict the response of annual mean PM_2.5 and O₃ pollution level changes in air pollutant emissions and greenhouse gas concentrations. Our method can account for sub-regional changes in air pollutant emission patterns, and the non-linear response of secondary air pollutants to precursor and greenhouse gas emissions. Our emulator successfully reproduce the PM_2.5 levels, and the associated public health impacts from the air quality model, while requiring only 4 – 6 seconds of CPU time (~3000 hours for air quality model) for each scenario. The accuracy, speed and simplicity of our emulator show the capability of machine learning algorithms in emulating global air quality models and make air quality modelling accessible for global climate/air pollution scenario analysis and integrated assessment.