A Novel Multi-Objective Optimization Method by Machine Learning and its Application in MEMS Vibration Energy Harvesters Design

This paper presents a novel method based on machine learning to improve multi-objective performance in micro devices efficiently. Stochastic strategies are employed, with automated parameterization through a variational autoencoder and iterative refinement by a Bayesian optimizer. Specifically, this method optimizes multiple performance objectives for micro vibration energy harvesters with significantly low computational cost. After fewer than 160 simulation iterations from initial reference devices, it reduces the resonant frequency from 261.2 Hz to 164.8 Hz, and increases the voltage from 0.28 V to 2.49 V. Additionally, it decreases the modal frequency gap from 252 Hz to 4 Hz while reducing the output ratio between the first two modes from 133 to 1.7. This work provides fresh insights for designing vibration energy harvesters and could be further applied to other micro devices across various fields.