A deep learning approach for efficient topology optimization based on the element removal strategy

•The boundary conditions are encoded as inputs, making the model can provide solutions with arbitrary loading on the perimeters.•The DL model learns the structural evolutionary process, making it able to provide solutions with arbitrary target volume fractions.•The model is extended into 3D cases by using image data with multiple channels representing 3D structures.•The well-trained DL model has huge time efficiency (0.9 s) than conventional FEM (12.5 min) on one 3D case.