Application of Machine Learning in Design and Optimization of MEMS Tunable Capacitors

Abstract Electrostatically actuated MEMS tunable capacitors have a nonlinear capacitance-voltage (C-V) response and limited tunability. Increasing the tunability and the linearity of the device’s response is of great desire. In this study, a design method, based on machine learning and optimization, is introduced that improves the tuning ratio and linearity of the device response. A capacitor with two deformation modes (rigid-body displacement of the moving electrode before pull-in and its deformations after pull-in) is used to demonstrate the technique. A set of simulation data for different design geometries is generated by ANSYS coupled-field multiphysics solver to create a three-level, three-factor central composite rotatable design (CCRD). To predict the linearity of the MEMS capacitor, a mathematical model using response surface methodology (RSM) is developed. The initial and final gap between the electrodes and the curvature of the moving electrode are considered as the input parameters and linearity as the outcome for the regression model development. The created model was evaluated for adequacy and significance. The model can predict linearity with more than 97% accuracy, according to confirmation tests, and the effect of the input parameters on the linearity of the response is studied. The optimization was conducted by Design-Expert 13 software and results show 0.96 for the maximum linearity of the C-V plot.