Machine Learning-Driven Structural Optimization of a Bistable RF MEMS Switch for Enhanced RF Performance

In the rapidly advancing digital era, the demand for miniaturized and high-performance electronic devices is increasing, particularly in applications such as wireless communication, unmanned aerial vehicles, and healthcare devices. Radio-frequency microelectromechanical systems (RF MEMS), particularly RF MEMS switches, play a crucial role in enhancing RF performance by providing low-loss, high-isolation switching and precise signal path control in reconfigurable RF front-end systems. Among different configurations, electrothermally actuated bistable lateral RF MEMS switches are preferred for their energy efficiency, requiring power only during transitions. This paper presents a novel approach to improve the RF performance of such a switch through structural modifications and machine learning (ML)-driven optimization. To enable efficient high-frequency operation, the H-clamp structure was re-engineered into various lateral configurations, among which the I-clamp exhibited superior RF characteristics. The proposed I-clamp switch was optimized using an eXtreme Gradient Boost (XGBoost) ML model to predict optimal design parameters while significantly reducing the computational overhead of conventional EM simulations. Activation functions were employed within the ML model to improve the accuracy of predicting optimal design parameters by capturing complex nonlinear relationships. The proposed methodology reduced design time by 87.7%, with the optimized I-clamp switch achieving -0.8 dB insertion loss and -70 dB isolation at 10 GHz.