Generating Multiple Distinct Feasible Solutions for MEMS Accelerometers Using Deep Learning

Abstract

Designing micro-electro-mechanical system (MEMS) sensors to meet specific performance requirements is essential. Traditional approaches, which rely heavily on expert knowledge and extensive finite-element simulations, are often time-consuming. Current deep learning (DL) methods in MEMS design typically focus on finding a single feasible solution, neglecting the need to generate multiple solutions simultaneously, which is critical in practical design scenarios. This article presents a methodology to address these limitations, introducing a hybrid network called the conditional variational autoencoder (VAE) and generative adversarial network (CVAE-GAN), along with a multisolution generator (G-MS). The CVAE-GAN enables high-accuracy and high-efficiency inverse design, while the G-MS, with its tailored noise updating strategy, generates multiple distinct feasible solutions for given performance criteria. This methodology has been experimentally validated on a piezoresistive MEMS accelerometer, finding the second solution in $3.60~\pm ~2.46$ s, with a normalized distance of $0.75~\pm ~0.19$ , improving the existing method as much as $3.63\times $ and $7.19\times $ , respectively. While traditional methods struggle to find more than two solutions, our G-MS can continuously output solutions according to the specified number, with the time taken to find each solution remaining nearly constant. This approach demonstrates the capability to quickly generate multiple accurate structural parameters based on desired performance, showcasing significant potential and providing valuable insights for MEMS sensor design.