Deep Learning-based Multiscale Modelling of Polysilicon MEMS

In this paper, we propose a data-driven multiscale modeling framework for polysilicon micro electromechanical systems (MEMS). At the material level, the relevant features of the morphology of the polycrystalline structural film are learned by a (tiny) convolutional neural network, which is shown to be able to provide size-dependent solutions once trained in a proper way. At the device level, a multiple-input and mixed-data neural network-based model is adopted to also learn the effects of microfabrication defects on the performance indices of the entire device. With specific reference to a single-axis, resonant Lorentz force magnetometer, it is reported that the model is capable to efficiently estimate the expected scattering in its response, even for a characteristic size of the resonant structure not allowed for during the training process.