Customized　3D-printed　structural　parts　are　widely　used　in　surgical　robotics.　To　satisfy　the　mechanical　properties　and　kinematic　functions　of　these　structural　parts,　a　topology　optimization　technique　is　adopted　to　obtain　the　optimal　structural　layout　while　meeting　the　constraints　and　objectives.　However,　topology　optimization　currently　faces　some　practical　challenges　that　must　be　addressed,　such　as　ensuring　that　structures　do　not　have　significant　defects　when　converted　to　additive　manufacturing　models.　To　address　this　problem,　we　designed　a　3D　hierarchical　fully　convolutional　network　(FCN)　to　predict　the　precise　position　of　the　defective　structures.　Based　on　the　prediction　results,　an　effective　repair　strategy　is　adopted　to　repair　the　defective　structure.　A　series　of　experiments　is　conducted　to　demonstrate　the　effectiveness　of　our　approach.　Compared　to　the　2D　fully　convolutional　network　and　the　rule-based　detection　method,　our　approach　can　accurately　capture　most　defect　structures　and　achieve　89.88%　precision　and　95.59%　recall.　Furthermore,　we　investigate　the　impact　of　different　ways　to　increase　the　receptive　field　of　our　model,　as　well　as　the　trade-off　between　different　defect-repairing　strategies.　The　results　of　the　experiment　demonstrate　that　the　hierarchical　structure,　which　increases　the　receptive　field,　can　substantially　improve　the　defect　detection　performance.　To　the　best　of　our　knowledge,　this　paper　is　the　first　to　investigate　3D　defect　prediction　and　repair　for　topology　optimization　in　conjunction　with　deep　learning　algorithms,　providing　practical　tools　and　new　perspectives　for　the　subsequent　development　of　topology　optimization　techniques.　©　2024　The　Author(s)