Transfer-Learning-Aided Optimization for a Low-Frequency Wideband MEMS Energy Harvester

Nowadays high frequency and low bandwidth are two critical challenges for microelectromechanical system (MEMS) energy harvesters. Energy harvesters with high operating frequency are not desirable considering the low-frequency nature of ambient vibrations. The operational frequency range (known as bandwidth) is another important characteristic that should be considered under an unpredictable or uncontrollable condition of ambient vibrations. This paper presents an innovative design of a vibration-based piezoelectric MEMS energy harvester structure, which encloses four masses. To gain premium performance, we have developed a transfer learning method to train a deep neural network model with FEM simulation data. By using this trained model for estimating harvester performance, we optimize the proposed structure with a genetic algorithm. Our optimized harvester not only features four low resonant frequencies (between 70Hz and 161Hz) and high bandwidth, but also reaches a good amount of harvested voltage. Our simulation results confirm its efficacy and superiority over the alternative designs.