Uncertainty Analysis of Mems Devices with Deep Neural Networks Using Bi-Fidelity Data

A novel deep neural network training technique for analyzing the impacts of process variations on MEMS devices is developed. DNN has not been widely applied in the MEMS analysis, since expensive simulated datasets are needed. This proposed method significantly reduces the computational resource required for constructing the dataset by training DNN with a mix of high-fidelity and low-fidelity data (bi-fidelity data). Applying this approach to the analysis of a MEMS thermal expansion coefficient test structure, high accuracy is achieved with an 85.0% reduction in required high-fidelity data and a 45.0% reduction in total computational cost compared to the conventional DNN. This method will help the MEMS device design and yield prediction.