The　advent　of　deep　learning　provides　a　promising　opportunity　to　improve　the　efficiency　of　topology　optimization.　However,　existing　methods　make　it　difficult　to　achieve　a　balance　between　efficiency,　accuracy,　and　generalization　ability.　To　tackle　this　challenge,　we　propose　a　novel　method　based　on　a　two-stage　network　framework.　In　the　network,　the　partial　convolution　block　and　shifted　windows　attention　mechanism　are　integrated　to　improve　the　model　performance.　In　the　first　stage,　a　convolutional　neural　network-based　model　trained　with　a　novel-designed　loss　function　is　employed　to　achieve　real-time　prediction　of　suboptimal　structures.　In　the　second　stage,　transfer　learning　is　introduced　to　inherit　the　output　of　the　first　stage.　Subsequently,　the　second　stage　optimizes　the　suboptimal　structures　to　get　the　final　optimal　structures　in　a　physical　information-driven　way.　On　the　2000　dataset,　the　two-stage　method　achieves　an　average　compliance　error　of　−1.45%,　and　95.5%　of　the　optimal　structures　perform　better　than　that　obtained　by　the　traditional　method　and　strictly　meet　volume　constraints　while　eliminating　structural　disconnections.　Finally,　the　proposed　method　is　applied　to　a　real-world　engineering　application　for　the　first　time,　and　the　design　of　bridge　pylons　is　given　as　an　example.　The　results　show　that　the　proposed　method　is　a　promising　exploration　of　topology　optimization　based　on　deep　learning.　©　2024　Elsevier　Ltd