Data-Driven Freeform Mems Energy Harvester Design Enabled by Machine Learning

Abstract

This paper reports a computational method for the design of freeform piezoelectric energy harvesters (PEHs) fabricated by micromachining processes based on the machine learning (ML) scheme. The geometry of candidate designs is first converted to pixelated images and assigned with specific properties and analyzed by the finite element method (FEM). The resulting neural network machine learning algorithm is trained using the above dataset to identify the properties of similar freeform PEH designs, which is 30,000 times faster than those by the FEM simulations. With 200,000 freeform designs analyzed by the ML-analyzer, the best designs with broad operation frequency bandwidth, low-frequency spectrum, and high output voltage are identified.