VortexNet: A Graph Neural Network-Based Multi-Fidelity Surrogate Model for Field Predictions

Abstract

Aircraft design relies on accurate aerodynamic predictions. High-fidelity (HF) methods, such as computational fluid dynamics (CFD), provide accurate aerodynamic analyses but are computationally expensive for early-stage design. Conversely, low-fidelity (LF) methods, such as the vortex lattice method (VLM), offer cost-effective solutions but struggle to capture complex flow phenomena, limiting their predictive accuracy. The conceptual design process thus presents a fidelity-cost trade-off, requiring a balance between high-fidelity (HF) and LF data in early design phases. This study introduces VortexNet, a graph neural network (GNN)- based surrogate model designed to bridge the fidelity gap between LF and HF aerodynamic predictions. VortexNet learns corrections to LF panel-wise local loading coefficient field data using data-driven insights from HF CFD simulations, enabling pressure coefficient field predictions across a range of Delta wing geometries and free-stream conditions. The model demonstrates strong prediction accuracy and generalizability across configurations, effectively capturing nonlinear flow features under geometric variations. A hyper-parameter sensitivity study and a preliminary prediction mechanism explanation, leveraging the latent space ablation technique, are conducted to rationalize the model’s predictive capabilities and provide guidance for future improvements in VortexNet-like surrogate modeling. These results indicate that VortexNet has potential as a valuable tool for conceptual design in multidisciplinary design optimization (MDO), while emphasizing the need for further validation and refinement.

Topics & Concepts