Identifying Contributors to PM 2.5 Simulation Biases of Chemical Transport Model Using Fully Connected Neural Networks

Accurate prediction of ambient PM 2.5 concentrations using air quality models can provide governments with information for public health alerts. However, due to large uncertainties of input parameters and over‐simplification of the chemical mechanism, the model simulations tend to have a certain deviation from the observations. To provide an insight into the discrepancy and to explain the contributors to the model bias, we propose here a machine learning based method to identify the contributors to PM 2.5 simulation biases. A fully connected deep neural network (noted as FCNN) was designed to correct the PM 2.5 biases between the simulations from a common air quality model (i.e., Community Multiscale Air Quality, CMAQ) and observations with meteorological and pollutants variables. The FCNN was applied in two polluted regions in China including Beijing‐Tianjin‐Hebei (BTH) and Yangtze River Delta (YRD) in 2015, exhibiting excellent performance in reducing the root mean square error of annual PM 2.5 by 46.6% and 37.2%, respectively. The relative contribution of each input feature for the bias correction was also estimated from the FCNN. Results suggest that the temperature and humidity exhibit the greatest contribution to the PM 2.5 simulation bias among all meteorological factors, probably due to their high association with the physical and chemical reaction conditions. NO 2 and SO 2 concentrations and associated biases were also found to be crucial to CMAQ model accuracy, implying the importance of NO 2 ‐ and SO 2 ‐related reaction for PM 2.5 formation. The study also revealed a cumulative effect of pollution and an enhancement effect of atmospheric oxidation on the formation of heavy pollution.