Analysis and Application of Deep Learning in Novel MEMS Capacitive Pressure Sensor

With the continuous development of MEMS sensor processing and manufacturing technology, MEMS capacitive pressure sensors require computer-aided software to reduce development costs. Therefore, simulation with Comsol Multiphysics and building of a deep learning approach s proposed. To build a highly efficient deep neural network (DNN) model with a sufficient dataset from Comsol simulation for MEMS analysis, various parameters of the manufacturing design are needed to understand the design characteristics of the sensor. The designed capacitive pressure sensor has a three-layer structure with a vacuum layer in the middle, the top, and the bottom electrodes depending on the material. The DNN with four hidden layers and 20 neurons for each layer is proposed to fit the relationship between the four input parameters and two outputs of a capacitive pressure sensor with deep learning. The DNN allows the design parameters and characteristics of the capacitive pressure sensor to be understood with a small data set simulated by Comsol Multiphysics, thus reducing the solution time and improving development efficiency. Using the dataset with data augmentation, the DNN was trained and tested. A set of data set samples is obtained at an average of 40 min. If the trained capacitive pressure sensor model is used for prediction, it only consumes an average of 10 s per set of data, and the prediction error is less than 0.3%. The results show a practical and high efficiency and accuracy of deep learning for prediction on capacitive pressure sensors.