Abstract
Abstract
Accurate calculation of physical quantities such as surface pressure, wall shear stress, and temperature is crucial in aerodynamic design processes. Data-Driven Deep Learning methods have emerged as powerful surrogate models that accelerate design cycles. Meshes employed in Computational Fluid Dynamics (CFD) simulations have an unstructured and relational nature. Therefore, Graph Neural Networks (GNNs) that are widely utilized in graph learning demonstrate superior performance in flow predictions compared to other Deep Learning architectures.
In this thesis, the performance of Geometric Deep Learning methods in predicting flow-induced physical quantities on various aerodynamic surfaces is investigated. To evaluate the generalization capability of the models, comprehensive 2D and 3D datasets with diverse flow conditions were generated. MeshGraphNet architecture which has proven successful in learning mesh-based simulations is taken as a baseline and the effects of structural and geometric improvements applied to enhance the model's expressive power were examined. To explore relational paradigms beyond standard GNNs, the performance of the Transolver model which adapts the Transformer architecture to physical simulations was evaluated. Finally, a hybrid model architecture was developed to combine the GNN's ability to learn local interactions with the Transformer's capacity to comprehend the global context. The models were comparatively analyzed in terms of node-based error metrics, aerodynamic coefficient correlations, and training throughput. The results show that Geometric Deep Learning methods learn aerodynamic surface distributions with high accuracy. It was observed that the proposed hybrid architecture and model enhancements significantly reduced prediction error by increasing the model's expressive power particularly in 3D datasets containing high node counts.