In　this　paper,　we　propose　a　deep　learning　based　neural　network　to　generate　three-dimensional　structural　topologies　in　an　efficient　way.　The　method　consists　of　three　steps.　First,　conventional　three　dimensional　solid　isotropic　microstructures　with　penalization　(SIMP)　method　is　utilized　to　generate　datasets　consisting　of　various　domain　sizes　and　boundary　conditions.　Second,　a　deep　learning　neural　network　based　on　U-Net　architecture　is　constructed　and　trained　by　the　generated　datasets.　Third,　by　feeding　new　cases　with　different　domain　sizes　and　boundary　conditions　into　the　network,　near　optimal　results　can　be　directly　obtained　without　any　needs　of　optimization　iterations　and　finite　element　analysis.　Compared　with　conventional　topology　optimization　methods　as　well　as　recent　development　of　machine　learning　approaches,　our　proposed　method　has　two　advantages:　(1)　the　design　boundary　conditions　are　directly　mapped　with　the　3D　optimized　structures　such　that　no　further　dependency　on　the　conventional　topology　optimization　algorithm　is　required　once　the　model　is　trained,　and　(2)　topology　optimization　problems　with　variable　design　domain　sizes　can　be　supported　instead　of　requiring　its　size　to　be　fixed　as　the　input　of　the　neural　network.　Experiments　demonstrate　that　once　trained,　our　deep　learning　based　topology　optimization　method　can　realize　near　optimal　three　dimensional　topology　prediction　using　negligible　calculation　cost.　©　2020　Elsevier　Ltd