The work aims to accurately obtain the material removal depth of screw rotor belt grinding and explore the effect of process parameters on the material removal depth. The affecting factors of material removal depth were determined by analyzing the removal mechanism of screw rotor belt grinding. A long and short-term memory network-convolutional neural network model (SSA-CNN-LSTM) optimized based on the sparrow algorithm was proposed to predict the material removal depth in screw rotor belt grinding process. The grinding tools developed to accommodate concave and convex belt grinding of rotors were divided into two types, a contact wheel type belt grinding mechanism and a free-form belt grinding mechanism. With the factors affecting the depth of material removal in grinding as inputs and the grinding depth as outputs, the prediction model adopting SSA for optimization of CNN-LSTM hyperparameters was constructed and compared with CNN-LSTM, LSTM, PSO-BP, RBF, and random forest prediction methods. In the experiment, the normal pressure Fs and tension force Fm of the abrasive belt were controlled by the main cylinder and the tensioning cylinder respectively. In order to ensure that the abrasive belt and the grinding surface were in full contact during the grinding process, the pressure of the main cylinder was set to be larger than the pressure of the tensioning cylinder by 0.1-0.3 MPa. The quality of the grinding and the stability of the grinding device were fully considered and the linear velocity of the abrasive belt was set to 4.4-13.1 m/s. The feed speed was decided to determine the grinding time and was set to 100-300 mm/min in order to guarantee the appropriate grinding depth range. In order to ensure a suitable grinding depth range, it was set at 100~300 mm/min, the abrasive belt grit was zirconia grit belt with 80-240 mesh and the grinding time was set at 0-20 min considering the wear rate of different grit sizes. The proposed prediction method has an average absolute percentage error MAPE up to 0.046 1, a root mean square error RMSE of 9.261, an average absolute error MAE up to 7.836, and a coefficient of determination R2 of 0.997 4, which provides a higher prediction accuracy and is able to efficiently predict the depth of grinding material removal from screw rotors compared to the unoptimized CNN-LSTM network and other classical network models. The effects of grinding process parameters on material removal depth MRD and material removal consistency are explored by the proposed model. From the prediction results, it can be seen that the material removal depth of screw rotor belt grinding increases with the increase of normal pressure and belt linear speed, decreases with the increase of feed rate and belt grit size and is most affected by normal pressure. By analyzing the contour of the workpiece before and after grinding, it can be seen that the depth of material removal with the normal pressure on the grinding area shows a trend of large grinding depth in the middle and a gradual decrease on both sides. The prediction model proposed has a significant prediction effect and analyzes the effect of process parameters on the grinding material removal depth and removal consistency, which can provide reference for the prediction of other types of machining contours.