Artificial neural network-based predictive performance estimation of novel MEMS capacitive pressure sensors with paraboloid-touched substrate architecture

Abstract

Abstract

Purpose This paper aims to present an intelligent micro-electro mechanical system (MEMS) estimator based on an artificial neural network (ANN) for rapid and accurate prediction of performance characteristics of a novel MEMS capacitive pressure sensor featuring a three-layer vacuum-sealed structure with a paraboloid-touched substrate configuration, offering enhanced sensitivity over conventional designs. Design/methodology/approach The sensor is governed by six input parameters (R,T,G,P,E,ν), with the ANN simultaneously predicting membrane deflection, output capacitance, capacitive sensitivity and mechanical sensitivity. A four-hidden-layer multilayer perceptron (20 neurons each) is trained on 2,000 COMSOL generated samples, evaluated via fivefold cross-validation and a 10% hold-out test set. All analytical computations were further supported and validated using MATLAB. Findings The proposed sensor achieves an overall capacitive sensitivity of 3.5 × 10–11 F/Pa, with the estimator attaining an overall R2 of 0.9927 and per-output R2 exceeding 0.990 for deflection and mechanical sensitivity. ANN inference time is reduced to a fraction of a second, compared to the considerable time required for direct FEA. Originality/value This work introduces a paraboloid-touched substrate geometry that aligns the capacitor plate with the diaphragm deformation profile at minimum touch pressure, a strategy unexplored in prior touch-mode sensor designs. The machine learning framework, validated through MATLAB-based analytical modeling, offers a scalable and time-efficient alternative to FEA for accelerated MEMS sensor design and optimization.