Inverse Airfoil Design for Hydrokinetic Turbine Blades Using Non-intrusive Reduced Order Modeling

Hydrokinetic turbines are devices that convert the kinetic energy of flowing water into electricity without accumulating water. Recent advancements in renewable energy have increased interest in these turbines, and optimizing the design of their blade sections (airfoils) can improve their performance. One of the biggest challenges in the design of these turbines is dealing with cavitation, which is the formation and collapse of bubbles in the flow of water. The study aims to use deep learning architectures–Convolutional Autoencoders (CAEs) and Multi-layer Perceptrons (MLPs) for non-intrusive Reduced Order Modeling (ROM) and parameterization of the airfoils, respectively. After training these models with data generated using JAVAFOIL, the models can generate airfoils corresponding to unseen design parameters, which also prevents cavitation. Non-intrusive, data-driven ROM techniques utilized reduce computation time and power required by the deep learning architectures in the offline stages, and enables mapping of the airfoil sections at various angle of attacks (AOA) to their corresponding design parameters–coefficients of lift (CL), drag (CD), moment (CM), pressure (CP (min)) and Reynolds number (Re). The trained MLP-decoder model produced fairly accurate blade section geometries, along with precise estimation of the angle of attack for unseen design parameters instantaneously, with MSE values centered around 0.02. Thus, a robust method for blade airfoil generation is presented, aiding in design of hydrokinetic turbines, devoid of cavitation.