Deep learning-guided prediction of mechanical properties in membrane resonators

Abstract

Micro/Nano-electro-mechanical-systems (MEMS/NEMS) are often hailed as disruptive technologies, enabling highly precise measurements across a wide range of applications. In this project we focus on resonator-type MEMS, which can be fabricated from micrometersized membranes, with relative ease and cost-effective manufacturing processes. However, determining optimal design parameters remains a significant challenge. This thesis explores the potential of various machine learning (ML) models—such as feedforward neural networks, convolutional networks, transformers and tree-based models—for predicting key properties of mechanical resonators, i.e. eigenfrequencies, quality factors and buckling behaviours. While finite element method-simulations (FEM) are the standard approach in this context, our study shows that ML can yield comparable results at a fraction of the computational cost. We examine a series of model systems, from simple beam structures to membranes with various hole types and configurations. Our approach is data-driven, and to generate training data for ML we use COMSOL Multiphysics. For high precision results, we utilise transfer learning, leveraging features from models trained on entirely different domains. This enables us to successfully perform regression of quality factors and eigenfrequencies, where in particular lower eigenmodes exhibit small relative errors and standard deviations. We further demonstrate that neural networkbased models can predict buckling in membranes with exceptional precision.