Machine learning based prediction of aerodynamic forces on wing profile and development of shape optimization model

The accurate prediction of aerodynamic coefficients plays a critical role in aircraft design processes. Although traditional Computational Fluid Dynamics (CFD) methods provide high accuracy, their high computational cost and long simulation times impose limitations on rapid design cycles. In this study, machine learning and deep learning-based models were developed and their performance was compared to predict aerodynamic coefficients of airfoil profiles. The dataset was generated using XFOIL software for different Mach numbers (0.3 ≤ Mach ≤ 0.7) and angles of attack (4° ≤ α ≤ 14°). To evaluate the generalization performance of different airfoil profiles, the dataset was split based on NACA airfoil numbers during the creation of training and test datasets. No airfoil geometry in the training dataset was included in the test dataset, and similarly, none of the airfoil geometries in the test dataset were used during training. This method aims to assess the model's ability to predict aerodynamic coefficients for previously unseen airfoil profiles. In this study, various machine learning models, including Support Vector Machines (SVM), XGBoost, AdaBoost, Random Forest, Decision Tree, Artificial Neural Networks (ANN), and Convolutional Neural Networks (CNN), were tested. The results indicate that CNN and ANN achieved the highest prediction accuracy. However, XGBoost and Random Forest demonstrated comparable performance to deep learning models in predicting certain aerodynamic coefficients and were considered viable alternatives due to their faster training times. One of the novel contributions of this study is the introduction of a multi-criteria evaluation approach for selecting the optimal airfoil profile, considering not only the lift-to-drag ratio (CL/CD) but also the moment coefficient (CM). The proposed method employs an optimization metric that simultaneously evaluates aerodynamic efficiency and flight stability. The results demonstrate that the trained model, developed for different Mach numbers and angles of attack, successfully predicts aerodynamic coefficients for previously unseen airfoil profiles. Thus, machine learning-based models have been identified as a viable and efficient alternative for aerodynamic coefficient estimation in aircraft design processes. This thesis contributes to the field by demonstrating how machine learning and deep learning approaches can be effectively utilized in aerodynamic analysis, thereby supporting data-driven design processes.