Machine-learning-based corrections of CMIP6 historical surface ozone in China during 1950-2014

Due to a lack of long-term observations in China, reports on historical ozone concentration are severely limited. In this study, by combining observation, reanalysis and model simulation data, XGBoost machine learning algorithm is used to correct the surface ozone concentration from CMIP6 climate model, and the long-term and large-scale surface ozone concentration of China during 1950-2014 is obtained. The long-term evolutions and trends of ozone and meteorological effects on interannual ozone variations are further analyzed. The results reveal that CMIP6 historical simulations have a large underestimation in ozone concentrations and their trends. The XGB-derived ozone are closer to observations, with R² value of 0.66 and 0.74 for daily and monthly retrievals, respectively. Both the concentrations and exceedances of ozone in most parts of China have shown increasing trends from 1950 to 2014. The daily mean ozone concentration without climate change effects is estimated to be 117 ppb in the year 1950 averaged over China. It indicates that the increase in anthropogenic emissions of China has a significant contribution to ozone enhancement between 1950 and 2014. The higher ozone growth rates of XGB retrievals than those from the model indicate a regional surface ozone penalty due to the warming climate. The relatively significant increment in ozone are estimated in the Central and Western China. Seasonally, the ozone enhancement is largest in spring, indicating a shift in seasonal variation of ozone. Given the uncertainty in simulating historical ozone by climate model, we show that machine learning approaches can provide improved assessment of evolution in surface ozone, along with valuable information to guide future model development and formulate future ozone pollution prevention and control policies.