We present a novel systematic MEMS structure design approach based on a “deep conditional generative model”. Utilizing the conditional generative adversarial network (CGAN) on a case study of circular-shaped MEMS resonators, three major advancements have been demonstrated: 1) a high-throughput vectorized MEMS design generation scheme that satisfies the geometric constraints; 2) MEMS structural customization toward tunable, desired physical properties with excellent generation accuracy; and 3) experience-free design space explorations to achieve extreme physical properties, such as low anchor loss of micro resonators. Excellent agreements with experimental data, numerical simulations, and a previously reported machine learning-based analyzer are achieved for validation of our methodology. As such, the proposed scheme could open up a new class of data-driven, intelligent design systems for a wide range of MEMS applications.
Keywords: MEMS Design, Conditional Generative Adversarial Networks, Data-driven Design, Machine Learning