Design and optimization of flapping foil based on unsteady Multi-Physics field reconstruction

Flapping foil technology, a novel fluid energy harvesting technique based on biomimicry, has demonstrated significant application potential. This paper introduces deep learning methods and addresses the issues of low efficiency and lack of interpretability in traditional optimization designs of flapping foil energy harvesting. A serial double-module end-to-end prediction framework (CTP-DSCN) that links foil operational parameters to flow fields and performance has been developed. By optimizing the model's hyperparameters, accurate capture and prediction of unsteady flow characteristics and system performance have been achieved. Using automatic differentiation techniques, an optimization strategy for the complex, multidimensional nonlinear flow problems of flapping foils has been established, resulting in a Pareto set of multi-objective solutions for energy harvesting power and efficiency. The study shows that, after data preprocessing and hyperparameter optimization, the end-to-end deep convolutional neural network architecture (CTP-DSCN) can simultaneously capture unsteady flow features and accurately predict system performance, with an accuracy improvement of over 50% compared to machine learning methods, and high interpretability in physical field reconstruction and parameter extrapolation. Energy harvesting efficiency has been improved to 44.27% under the optimization strategy, as well as providing technical support for the static design of flapping foils.