Optimizing Biomimetic Elastic Propulsors by Integrating Machine Learning and Bayesian Optimization

Optimization in fluid-structure interaction (FSI) problems is often constrained by the high computational cost of full-scale simulations, limiting the number of design evaluations that can be performed. Here, we developed a machine learning surrogate model based on the Fourier neural operator to predict performance metrics such as thrust and efficiency to design an efficient biomimetic propulsor where nonuniform elasticity is used to enhance hydrodynamic performance. The surrogate model trained using a set of FSI simulations tends to exhibit overconfidence in regions with sparse data. To mitigate this, we employ an ensemble of models with varied dropout masks to estimate predictive uncertainty. This enables simultaneous prediction of expected performance as well as the variance in that prediction. We then integrate these models into a Bayesian optimization framework to strategically select the most informative designs for refining simulations. This approach significantly reduces the number of required FSI simulations while maintaining a high probability of identifying the optimal design.