Design optimization is a fundamental task underpinning the remarkable microwave response of Radio Frequency Microelectromechanical Systems (RF MEMS). The present paper introduces a new methodology that, through the simultaneous utilization of electromagnetic simulations and machine learning techniques, aims to further improve RF MEMS design. Simulated data have been used for training suitable machine learning algorithms, which are shown to be capable of generating models for predicting the critical device parameters even for cases beyond the simulated ones. In addition to the bridge dimensions, the training set includes variations in the 50 Ohm coplanar signal line width and gaps, all parameters affecting the device RF response. In this way, the proposed methodology enables design optimization for a wide family of RF MEMS capacitive switches, rather than only for specific cases, thus covering diverse application disciplines.