A Unified AI Approach for Modeling the Properties of MEMS Ultrasonic Transducers

Finite-element (FE) sweeps remain the standard for analyzing microelectromechanical systems (MEMS) ultrasonic transducers but are slow to traverse large geometry–bias spaces. This paper develops a multi-output surrogate that predicts the ultrasonic response of piezoelectric micromachined ultrasonic transducers (PMUTs) directly from compact design and operating descriptors, enabling millisecond-level evaluation for design-space exploration and bias tuning. Inputs include shape (circle/square), diaphragm diameter (200–600 μm), anchor count/geometry (2–8), and DC bias (5–15 V), with optional fabrication process features (e.g., membrane thickness, cavity depth). Targets comprise resonance frequency f0, quality factor (Q), sensitivity, center displacement, bias-tuning coefficient, β, and motional resistance, Rm. A dataset of 40 PiezoMUMPs chips is assembled with labels from COMSOL sweeps, laser Doppler vibrometry (LDV), and admittance-based fits. Six architectures, residual multilayer perceptron (MLP), convolutional neural network (CNN), gated recurrent unit (GRU), long short-term memory (LSTM), Transformer, and a CNN+LSTM hybrid, are benchmarked under a standardized pipeline (feature scaling, multi-target loss, 70/15/15 split, and five-fold cross-validation) with physics-preserving augmentation and multi-fidelity densification. All models achieve sub-percent normalized error; a compact Transformer encoder attains ≈ 0.03% normalized RMSE while preserving physically consistent trends with respect to shape, size and bias. The surrogate generalizes and supports inverse design, multi-objective optimization, and closed-loop bias control, reducing reliance on inner-loop FE sweeps.