This　paper　proposes　a　new　method　of　probabilistic　seismic　demand　analysis　with　high　accuracy　considering　heteroscedasticity.　For　a　typical　simply　supported　girder　bridge,　the　spectral　acceleration　corresponding　to　the　fundamental　period　of　bridge　is　chosen　to　reflect　the　ground　motion　intensity,　and　the　displacement　ductility　coefficient　of　pier　is　selected　as　engineering　demand　parameter.　A　series　of　nonlinear　dynamic　time　history　analysis　was　performed　on　the　bridge　finite　element　model　using　the　multiple　strip　analysis　and　the　seismic　demand　sample　set　was　established.　Based　on　the　Gaussian　process　regression　method,　the　probabilistic　seismic　demand　models　were　established　for　univariate　and　multivariate　cases,　and　the　fitting　results　were　compared　with　the　seismic　demand　models　derived　from　the　traditional　linear　regression.　The　seismic　fragility　models　were　established　based　on　the　two　aforementioned　seismic　demand　models　by　defining　the　limit　values　of　different　damage　states　of　piers.　The　results　show　that:　The　probabilistic　seismic　demand　models　based　on　Gaussian　process　can　reflect　the　probabilistic　characteristics　of　seismic　demand　under　different　ground　shaking　intensities　more　accurately　and　describe　the　heteroscedasticity　of　seismic　demand　in　logarithmic　space　better;　there　are　some　differences　in　the　fragility　curves　of　bridge　components　and　system　generated　by　the　two　models.　The　probabilistic　seismic　demand　model　based　on　Gaussian　process　can　provide　more　reasonable　and　accurate　assessment　of　the　seismic　performance　of　the　bridge.　©　2023　Tsinghua　University.　All　rights　reserved.