Machine Learning-Based Optimization of a Quad-Finger Energy Harvester Design

The growing development of electronic devices and the need for sustainable energy solutions have stimulated research into alternatives to traditional batteries. In this context, piezoelectric energy harvesters (PEHs) represent a promising technology, capable of converting environmental vibrations into electrical energy to supply low-power devices. However, PEH performance significantly depends on geometry and electrical parameters, and since traditional optimization based on Finite Element (FE) models involves very high computational costs, more efficient strategies are developed. To address this, the thesis proposes a hybrid approach combining machine learning models with genetic algorithms for the optimization of the geometric and electrical parameters of a Quad-Finger multimodal energy harvester. Both single-output and multi-output datasets, generated by an experimentally validated FE model, are used to train three supervised regression algorithms: Random Forest Regression (RFR), Gradient Boosting Regression Tree (GBRT), and Extreme Gradient Boosting Regression (XGBR). After validation, the ML models are integrated with a genetic algorithm to develop a fully data-driven optimization process. Among the models evaluated, GBRT shows the highest predictive accuracy. Overall, the proposed methodology not only ensures reliable response prediction for different harvester configurations, but also refines the design parameters to enhance power extraction.