Abstract: To systematically analyze the influence of sand belt grinding process parameters on the surface quality of screw rotors, and provide reference for parameter selection in actual production. To improve prediction accuracy, a random forest prediction model based on genetic algorithm optimization was constructed, and a five factor five level orthogonal experiment was designed. The experimental device was a self-developed multi head screw grinding device, Its specific parameters are as follows, workpiece axial feed rate of 100-300 mm/min, sand belt linear velocity of 4.4-13.3 m/s, sand belt tension pressure of 0.20-0.30 MPa, grinding pressure of 0.40-0.50 MPa, and sand belt particle size of 60-180 μm. The experimental results showed that the average prediction error of the GA-RF model was 9.06%, significantly lower than that of the Lasso model (25.96%) and SVR model (30.68%). Single factor analysis shows that surface roughness increases with the increase of axial feed rate and decreases with the increase of sand belt linear velocity. When the feed rate increases from 100 to 300 mm/min, the Ra value increases by about 27%. When the linear velocity increases from 4.4 m/s to 13.3 m/s, the Ra value decreases by about 35%. The study verified the effectiveness of the genetic algorithm-random forest (GA-RF) model in predicting the quality of sand belt grinding, and revealed the influence of key process parameters. These findings can provide theoretical guidance for parameter selection in screw rotor machining and have important reference value for practical production.