Physics-Enhanced Neural Networks for Aerodynamic Shape Optimization of UAV Wings

Aerodynamic shape optimization plays a crucial role in the development of UAV wings, directly impacting mission efficiency, endurance, and overall flight performance. Traditional optimization methods rely heavily on high-fidelity CFD simulations which, due to their high computational cost, limit extensive exploration. Another alternative is the use of computationally efficient, yet physically simplified, low-fidelity methods that poorly represent nonlinear aerodynamic phenomena such as aerodynamic stall. To address these limitations, a multi-fidelity neural network architecture is developed, combining large amounts of low-fidelity data with fewer high-fidelity evaluations. To strengthen predictive accuracy, this neural network framework is enhanced through domain knowledge-driven feature engineering and physics-informed soft constraints, specifically targeting the relatively undersampled stall region. Additionally, Monte Carlo dropout is incorporated to estimate uncertainty, enabling more confident and reliable design decisions. The integration of aerodynamic principles ensures realistic lift curves exhibiting proper stall behavior and physically plausible drag polars, thereby achieving robust, physically consistent predictions across the operational envelope of fixed-wing UAVs.