In　the　area　of　bridge　seismic　fragility　studies,　the　fragility　curve　is　commonly　defined　as　a　log-normal　cumulative　distribution　function,　and　its　parameters　are　usually　estimated　through　two　methods.　However,　the　fragility　curves　derived　by　one　of　these　two　methods　may　intersect,　whereas　the　other　imposes　significant　restrictions　on　the　shape　of　the　curves.　The　purpose　of　this　study　is　to　derive　fragility　curves　by　utilizing　machine　learning　tools　(in　particular,　tools　that　are　based　on　multi-classification　probabilistic　prediction),　to　address　the　limitations　that　are　present　in　existing　methodologies.　To　be　more　precise,　support　vector　machine　for　classification　(SVC)　and　relevance　vector　machine　for　classification　(RVC)　are　utilized　to　construct　fragility　curves,　while　two　other　classic　methodologies　are　chosen　for　comparison　and　validation.　The　four　approaches　were　tested　on　a　dataset　acquired　using　numerical　analysis,　and　their　performance　was　measured　using　Brier　-Score　and　Log-Loss.　The　results　show　that　both　SVC　and　RVC　not　only　overcome　the　limitations　and　outperform　the　classical　method,　but　also　show　robustness　to　arbitrarily　partitioned　training　sets,　whereas　the　classical　method　sometimes　performs　poorly.