A method was introduced for predicting impact sensitivity of explosives by the artificial neural networks. Combining with the topological parameters and the quantum-chemical parameters which obtained by analyzing the fully optimized geometries and the vibration analysis of 36 CHON explosive molecules using the density functional theory (DFT) method at the B3LYP/6-31G^* level, seven molecular descriptors close related to H₅₀ were selected, including total electronic energy，lower unoccupied molecular orbital energy，oxygen balance index，number of oxygen atoms， active index, indicator of aromaticity （0 or 1）， indicator of —CH in α （0 or 1）. And the artificial neural network (ANN) with these descriptors as neurons in the input layer was established to predict impact sensitivity of explosives. The predicted data of the ANN were compared with experimental and those of two traditional models established by the oxygen balance index (OB₁₀₀) and the active index (F) respectively. Results show that the root mean squares errors of ANN model is 17.84 cm and that of the two traditional models is 42.71 cm and 36.47 cm respectively.