A low-pressure MEMS piezoresistive pressure sensor using multilayer perceptron neural network for structure optimization

Abstract This paper presents a novel micro-electromechanical system piezoresistive pressure sensor designed for a low-pressure range of 0 – 2 kPa. The sensor features an innovative square silicon diaphragm incorporating beam-island structures on the front surface and beams on the back surface. This structural configuration effectively concentrates stress in the piezoresistive regions while enhancing local stiffness, thereby achieving high sensitivity and excellent linearity. Finite element method (FEM) simulations were performed to analyze the stress distribution and deflection characteristics of the diaphragm. To efficiently optimize the diaphragm structure, a deep learning-assisted optimization framework combining a multilayer perceptron (MLP) surrogate model with particle swarm optimization (PSO) was developed. The trained MLP captures the nonlinear relationship between structural parameters and sensor performance, while PSO enables continuous and global exploration of the design space to identify optimal configurations. This MLP–PSO framework significantly improves design efficiency and flexibility compared with conventional FEM-based parametric scanning, enabling sensitivity maximization while maintaining low nonlinearity. FEM simulations of the optimized designs were conducted for square diaphragm sizes ranging from 3500 to 3900 μ m, demonstrating consistently high sensitivity and low nonlinearity across the design space. For a diaphragm size of 3795 μ m with a thickness of 16.2 μ m, the proposed sensor achieves a voltage sensitivity of 25.25 mV V −1 kPa −1 and a nonlinearity of 0.12% FS.