Identifying Contributors to PM _2.5 Simulation Biases of Chemical Transport Model Using Fully Connected Neural Networks

Accurate prediction of ambient PM2.5 concentrations using air quality models can provide governments with information for public health alerts. However, due to large uncertainties input parameters and over-simplification the chemical mechanism, model simulations tend have a certain deviation from observations. To an insight into discrepancy explain contributors bias, we propose here machine learning based method identify simulation biases. A fully connected deep neural network (noted as FCNN) was designed correct biases between common (i.e., Community Multiscale Air Quality, CMAQ) observations meteorological pollutants variables. The FCNN applied in two polluted regions China including Beijing-Tianjin-Hebei (BTH) Yangtze River Delta (YRD) 2015, exhibiting excellent performance reducing root mean square error annual by 46.6% 37.2%, respectively. relative contribution each feature bias correction also estimated FCNN. Results suggest that temperature humidity exhibit greatest among all factors, probably their high association physical reaction conditions. NO2 SO2 associated were found be crucial CMAQ accuracy, implying importance NO2- SO2-related formation. study revealed cumulative effect pollution enhancement atmospheric oxidation on formation heavy pollution.