When traditional nozzles are used for workpiece finishing, the jet leaves the nozzle, concentrating its energy at the central position, and the one-time processing area is small. As a result, although the overall surface roughness of the workpiece after finishing is effectively reduced, it tends to cause local deformation of the workpiece and the problem of poor uniformity. In order to further improve the surface quality and uniformity of thin-walled and complex workpieces processed by abrasive jet finishing, and to improve the efficiency of the finishing process. By changing the shape of the nozzle outlet, and then change the structure of the jet, so that the jet energy distribution is more uniform, to improve the effect and feasibility of abrasive jet finishing. Fluent software was used to analyse the jet structure of the nozzle, abrasive trajectory, erosion and shear effect, to explore the advantages of the in-line nozzle finishing, to verify the finishing effect of the in-line nozzle through the test, and to analyse the influence of the various influencing factors on the effect of the finishing, and finally to establish the BP neural network prediction model and particle swarm parameter optimisation, to find the optimal parameters of the finishing. The in-line nozzle can effectively improve the surface quality and uniformity of the abrasive jet finishing process, improve the efficiency of the abrasive jet finishing process, and in the case of the same mass flow rate, the in-line nozzle has less influence on the deformation of the workpiece. Finally, simulation analysis and experiments show that the slotted nozzle can effectively improve the surface quality and reduce the surface roughness of abrasive jet finishing, improve the efficiency of abrasive jet finishing, and reduce the influence of jet processing on the workpiece. Through the prediction model constructed by BP neural network and the optimization of particle swarm parameters, when the processing time is 15 min, the abrasive particle size is 20 μm, the target distance is 12 mm, and the pressure is 0.08 MPa, the surface roughness of the aluminum alloy after finishing is reduced from R_a 0.513 μm to R_a 0.219 μm, and the surface roughness is basically the same as that measured in the vertical direction. Experiments verify that the BP neural network prediction model has high accuracy.