Efficient Foil Shape Optimization Using Artificial Neural Network-Based Surrogate Model with Adaptive Sampling Strategy

To overcome the limitations of experimental and numerical design processes in terms of time, cost, efficiency, and uncertainty, researchers are developing optimization frameworks utilizing surrogate models based on Artificial Neural Networks (ANNs). Further enhancement of these frameworks is achieved by integrating adaptive sampling methods, which improve efficiency by reducing the number of training samples needed to reach optimal solutions. In this study, the aerodynamic shape of NACA0012 airfoil was optimized using an ANN-based surrogate model that investigated two different sampling techniques: One-Shot sampling with a Sobol sequence and Optimization-Based Adaptive sampling. The Hicks-Henne function was used to deform the shape, XFOIL was used for numerical simulations, and Sequential Least Squares Programming (SLSQP) was employed as the optimizer. Aerodynamic coefficients at the optimal points were compared for both sampling techniques. Although both strategies improved Lift and Aerodynamic Efficiency, the Optimization-Based Adaptive Sampling model found the global maximum, while the One-Shot Sampling remained at a local maximum. The Adaptive Sampling ANN model increased lift by $\mathbf{5 8. 6 8 \%}$. Moreover, the errors of predicted optima for both models were calculated relative to XFOIL results. The Adaptive Sampling ANN model displayed superior accuracy at the optimum. However, both models performed similarly across the design space.