This study focuses on the importance of initial conditions in air-quality prediction.Assimilation experiments are employed using WRF-Chem model and Grid-point StatisticalInterpolation(GSI)for a 9-day severe particulate matter pollution event occurring in Shanghai onDecember 2013.In this application,GSI relies on a three dimensional variational approach toassimilate ground-based PM2.5 observations into chemical model to obtain initial fields of aerosolspecies.The results demonstrate that the data assimilation significantly reduces the errors incomparison with a simulation without data assimilation and improves forecasts of theconcentrations of PM2.5.Despite a drop of skill in the early forecast hours after the assimilation,apositive effect of assimilation is explicitly noted in forecasts for up to at least 12-24 hours,whileforecasts of 48hr duration have relatively slight improvement.In addition to good performance inShanghai,the verification statistics for assimilation experiment are encouraging for most of thesurface stations in China.