Data-Driven Parameter Design of Broadband Piezoelectric Energy Harvester Arrays Using Tandem Neural Networks

Broadband piezoelectric energy harvesters (PEHs) are attractive for powering self-sustained sensing nodes in industrial monitoring, structural health monitoring, and distributed IoT systems, where ambient vibration spectra are often uncertain, drifting, and broadband. However, tuning multiple resonant peaks in PEH arrays usually relies on time-consuming finite element (FE) parameter sweeps or iterative optimizations, which becomes a practical bottleneck when rapid, site-specific customization is required. This study presents a data-driven inverse-design framework for a five-beam PEH array based on a tandem neural network (TNN). A forward multilayer perceptron (MLP) surrogate is first trained using 10,000 COMSOL-generated samples to predict the array's characteristic frequencies from the design variables (end masses M1-M5 and tilt angle α), achieving >98% prediction accuracy with a prediction time <1 s, thereby enabling efficient replacement of repeated FE evaluations during design. The trained MLP is then coupled with an inverse-design network to form the TNN, which maps target characteristic-frequency sets directly to physically feasible parameters through the learned surrogate. Multiple representative target frequency sets are demonstrated, and the TNN-generated designs are independently verified by COMSOL frequency-response simulations. The resulting arrays achieve broadband operation, with bandwidths exceeding 10 Hz. By shifting most computational cost to offline dataset generation and training, the proposed spectrum-to-parameter pathway enables near-instant parameter design and reduces reliance on exhaustive FE tuning, supporting rapid, application-specific deployment of broadband PEH arrays.