Machine Learning-Based Modeling and Generic Design Optimization Methodology for Radio-Frequency Microelectromechanical Devices

Abstract

RF-MEMS technology has evolved significantly over the years, during which various attempts have been made to tailor such devices for extreme performance by leveraging novel designs and fabrication processes, as well as integrating unique materials; however, their design optimization aspect has remained less explored. In this work, we report a computationally efficient generic design optimization methodology for RF-MEMS passive devices based on multi-objective heuristic optimization techniques, which, to the best of our knowledge, stands out as the first approach offering applicability to different RF-MEMS passives, as opposed to being customized for a single, specific component. In order to comprehensively optimize the design, both electrical and mechanical aspects of RF-MEMS device design are modeled carefully, using coupled finite element analysis (FEA). The proposed approach first generates a dataset, efficiently spanning the entire design space, based on FEA models. By coupling this dataset with machine-learning-based regression tools, we then generate surrogate models describing the output behavior of an RF-MEMS device for a given set of input variables. Finally, the developed surrogate models are subjected to a genetic algorithm-based optimizer, in order to extract the optimized device parameters. The proposed approach is validated for two case studies including RF-MEMS inductors and electrostatic switches, in which the multiple design objectives are optimized simultaneously. Moreover, the degree of conflict among various design objectives of the selected devices is studied, and corresponding sets of optimal trade-offs (pareto fronts) are extracted successfully.