Resumen
In this work, supervised Machine Learning (ML) techniques were employed to solve the forward and inverse problems of airfoil and hydrofoil design. The forward problem pertains to the prediction of a foil?s aerodynamic or hydrodynamic performance given its geometric description, whereas the inverse problem calls for the identification of the geometric profile exhibiting a given set of performance indices. This study begins with the consideration of multivariate linear regression as the base approach in addressing the requirements of the two problems, and it then proceeds with the training of a series of Artificial Neural Networks (ANNs) in predicting performance (lift and drag coefficients over a range of angles of attack) and geometric design (foil profiles), which were subsequently compared to the base approach. Two novel components were employed in this study: a high-level parametric model for foil design and geometric moments, which, as we will demonstrate in this work, had a significant beneficial impact on the training and effectiveness of the resulting ANNs. Foil parametric models have been widely used in the pertinent literature for reconstructing, modifying, and representing a wide range of airfoil and hydrofoil profile geometries. The parametric model employed in this work uses a relatively small number of parameters, 17, to describe uniquely and accurately a large dataset of profile shapes. The corresponding design vectors, coupled with the foils? geometric moments, constitute the training input from the forward ML models. Similarly, performance curves (lift and drag over a range of angles of attack) and their corresponding moments make up the input for the models used in the inverse problem. The effect of various training datasets and training methods in the predictive power of the resulting ANNs was examined in detail. The use of the best-performing ML models is then demonstrated in two relevant design scenarios. The first scenario involved a software application, the Design Foil Assistant, which allows real-time evaluation of foil designs and the identification of designs exhibiting a set of given aerodynamic or hydrodynamic parameters. The second case benchmarked the use of ML-enabled, performance-based design optimization against traditional foil design optimization carried out with classical computational analysis tools. It is demonstrated that a user-friendly real-time design assistant can be easily implemented and deployed with the identified models, whereas significant time savings with adequate accuracy can be achieved when ML tools are employed in design optimization.