Inverse Design of PMUT Using Deep Reinforcement Learning with a View to Customized Operating Frequency and Broadened Bandwidth

This paper introduces an inverse design framework using deep reinforcement learning (DRL) to optimize Piezoelectric Micromachined Ultrasonic Transducers (PMUTs) through non-standard diaphragm geometries, moving beyond traditional circular or rectangular patterns. The framework achieves three key advances: 1) demonstration of −3dB fractional bandwidth reaching 113.7% in the array while maintaining precise frequency control (±0.1MHz) for the cells, 2) automated generation of PMUT design without prior training data, and 3) validation of practical feasibility through a fabricated 40-channel linear PMUT array.