Machine learning aided optimization of drag pump performance

Low pumping speed limits the application of drag pumps. Combining the turbomolecular pump with the drag pump creates a hybrid molecular pump that increases the pumping speed. However, little research has been conducted on optimizing the drag pump of hybrid molecular pumps. To improve the pumping speed of the drag pump, this study employs a simplified rectangular channel model that incorporates rotor-stator clearance and conducts optimization based on proxy models developed through machine learning algorithms. First, the validity of the simplified rectangular channel model with clearance is established by data from published literature. Subsequently, the accuracy of proxy models fitted to represent the relationship between drag pump structural parameters and performance is systematically studied using three machine learning algorithms: Random forest, XGBoost, and CatBoost. Under the same samples, XGBoost achieves the highest fitting accuracy. Furthermore, the optimization of the drag pump based on the proxy model fitted by machine learning is efficient. Compared with direct optimization, the number of samples is reduced by 70%, and the pumping speed optimized based on the proxy model increases by 10% compared with the direct optimized result. This study will guide the design of drag pumps to improve the performance of the hybrid molecular pump.