The design efficiency of energetic compounds depends on many factors， such as the proportion of potential high performance samples in the screening space and the accurate prediction method of key properties. In this study， we proposed a scheme to improve the overall performance of virtual screening space by pre-screening molecular skeletons， and a method combining high-throughput computing and deep learning is applied to the design of energetic compounds. It was found that there is a moderate positive correlation between crystal density molecular skeleton density of energetic molecules， and the overall density of virtual screening space can be effectively improved by pre-screening high-density molecular skeletons. Based on the density data-set of energetic crystals collected from the crystallography database CCDC， a new density prediction model of energetic crystals was obtained via deep learning， with reliable accuracy and generalization. We took fused-ring energetic molecules as the research object， obtained high-density fused-ring skeletons through skeleton pre-screening， and then the virtual screening space composed of potential high-density molecules was constructed through fragment docking. The formation enthalpy， detonation performance and chemical stability were predicted by quantum chemical calculation and the equation of state of detonation products. Finally， 6 novel energetic molecules with energy level better than RDX and stability better than TNT were selected by performance ranking. This study shows that the overall performance of virtual screening space can be effectively improved by pre-screening molecular skeletons， and on this basis， high-throughput computing and deep learning can be used to achieve efficient design of energetic molecules.