A　neural　network　based　ensemble　methodology　was　presented　in　this　study　to　improve　the　accuracy　of　meteorological　input　fields　for　regional　air　quality　modeling.　Through　nonlinear　integration　of　simulation　results　from　two　meteorological　models　(MM5　and　WRF),　the　ensemble　approach　focused　on　the　optimization　of　meteorological　variable　values　(temperature,　surface　air　pressure,　and　wind　field)　in　the　vertical　layer　near　ground.　To　illustrate　the　proposed　approach,　a　case　study　in　northern　China　during　two　selected　air　pollution　events,　in　2006,　was　conducted.　The　performances　of　the　MM5,　the　WRF,　and　the　ensemble　approach　were　assessed　using　different　statistical　measures.　The　results　indicated　that　the　ensemble　approach　had　a　higher　simulation　accuracy　than　the　MM5　and　the　WRF　model.　Performance　was　improved　by　more　than　12.9%　for　temperature,　18.7%　for　surface　air　pressure　field,　and　17.7%　for　wind　field.　The　atmospheric　PM10　concentrations　in　the　study　region　were　also　simulated　by　coupling　the　air　quality　model　CMAQ　with　the　MM5　model,　the　WRF　model,　and　the　ensemble　model.　It　was　found　that　the　modeling　accuracy　of　the　ensemble-CMAQ　model　was　improved　by　more　than　7.0%　and　17.8%　when　compared　to　the　MM5-CMAQ　and　the　WRF-CMAQ　models,　respectively.　The　proposed　neural　network　based　meteorological　modeling　approach　holds　great　potential　for　improving　the　performance　of　regional　air　quality　modeling.　(C)　2012　Elsevier　Ltd.　All　rights　reserved.