Abstract: The particle flow discrete element software PFC2D is widely used in geotechnical engineering. The main problem is how to calibrate its mesoscopic parameters. At present, the "trial and error method" is mostly used. The disadvantage of this method is large amount of work and low efficiency. This paper proposes to replace this method with the back propagation (BP) neural network method, and use the PFC2D built-in FISH and Python language to re-develop a software, which then automatically runs to obtain 40 sets of macro mechanical parameter samples. The results show that the BP neural network method can quickly and accurately establish the macro and meso parameter mapping relationship, and the error is kept within 0.01. The simulated stress-strain curve is highly consistent with the indoor test curve, and the network can be created without large data samples. The efficiency is higher; and it is verified that the straight joint model selected in this paper can effectively determine the small UCS/TS value of the parallel bond model, indicating that the straight joint model can be used to better simulate the rock.