This　paper　aims　to　study　the　selection　of　the　scale　of　ground　motion　record　sets　based　on　the　cloud　method　applied　in　the　seismic　vulnerability　analysis　of　underground　structures.　Taking　a　shallow-buried　two-story　and　two-span　subway　station　structure　as　a　research　object,　the　dynamic　response　of　the　structure　under　different　natural　groundotion　records　was　obtained　by　using　the　nonlinear　time-history　dynamic　analysis　method.　The　inter-story　drift　ratio　was　selected　as　the　damage　measure,　and　the　PGA　was　defined　as　the　intensity　measure.　The　probabilistic　seismic　demand　model　of　the　structure　under　different　working　conditions　was　established　based　on　the　logarithmic　linear　regression　model,　and　the　corresponding　seismic　fragility　curves　of　the　structure　were　established.　Based　on　the　error　analysis　of　the　post-earthquake　failure　probability　of　the　structure　between　the　standard　result　and　the　result　of　different　working　conditions,　it　is　found　that　the　scale　of　ground　motion　record　sets　could　significantly　affect　the　seismic　vulnerability　analysis　results　of　the　underground　structure　based　on　the　cloud　method.　Compared　with　evaluating　the　higher-level　performance　of　structure,　the　scale　of　ground　motion　record　sets　has　less　impact　on　the　lower-level　performance.　The　failure　probability　evaluation　result　of　the　underground　structure　will　gradually　approach　the　standard　value　with　the　increase　of　input　scale　of　ground　motion　record　sets,　which　suggests　that　300　samples　is　the　optimal　scale　of　ground　motion　record　sets.　©　2023　Science　Press.　All　rights　reserved.