Optimization of turbomolecular pump blade structure based on generative adversarial network

Turbomolecular pumps are vital in high-vacuum applications, including semiconductor manufacturing and nuclear physics. While traditional straight blades are limited in performance, curved-blade designs offer enhanced efficiency but pose optimization challenges. This study presents a deep learning-based approach using Generative Adversarial Networks (GANs) to design three-dimensional curved-blade geometries with a straight blade as input. To overcome the lack of available vast amounts of training data in conventional GAN network training, a new training process is implemented via COMSOL simulations, which is used to analyze gas particle transmission and to give an evaluation of design output. Simulation results demonstrate that the curved-blade configurations significantly improve pumping performance, a 5.08% improvement in the maximum pumping speed coefficient and a 16.3% increase in the maximum compression ratio, compared to conventional straight designs. This work establishes a data-driven optimization framework for pump blade design and proposes a training methodology without large training data, highlighting the potential of machine learning in advancing vacuum technology. The methodology provides a scalable solution for complex aerodynamic structures, providing novel approaches for the optimized design of high-speed turbomolecular pumps.