Abstract: In recent years, the PFC (particle flowcode) discrete element analysis method is widely used in the field of geotechnical engineering. It can be used to reveal the origin of the mechanical behavior of rock materials from a microscopic perspective, but the micro parameters of the particle flow model are not the same as the macro parameters of rock materials. The calibration process is complicated and time-consuming. Based on the flat-joint contact model, the macro-mechanical parameters of rock materials are tested by numerical simulation tests such as the uniaxial compression, the direct tension and the biaxial compression. The micro-parameters are orthogonally designed, and the relationship between macro and micro mechanical parameters are studied through the multi-factor analysis of variance. The BP (back propagation) neural network modeling is used to calibrate the micro-parameters and verify the calibration results. The verification results show that the calibration accuracy of the micro-parameters is generally higher than 90% and the overall error is small, which shows the feasibility of the calibration method. Combined with the stress-strain curve and the failure characteristics of granite under routine laboratory tests, the effectiveness of the neural network inversion method in calibrating the micro mechanical parameters of rock materials is verified.