A multi-fidelity surrogate model based on convolutional neural network with transfer learning for coaxial rotor system aerodynamic optimization

Aerodynamic optimization of coaxial rotors requires balancing computational cost and accuracy. This study proposes a multi-fidelity surrogate model combining convolutional neural networks with transfer learning (CNN-TL) to address this challenge. The model takes grayscale images of rotor blades as inputs, encoding both planform and twist distribution, and predicts two key performance metrics: figure of merit (FM) in hover and effective lift-drag ratio (L/De) in forward flight. A transfer learning framework sequentially pre-trains the CNN on low-fidelity blade element momentum theory data and fine-tunes it with high-fidelity computational fluid dynamics results, effectively bridging fidelity levels. The Nondominated Sorting Genetic Algorithm II (NSGA-II) iteratively refines the Pareto front and generate infill samples with the expected improvement infill criterion. Compared to conventional hierarchical Kriging model, the CNN-TL approach reduces computational time by up to 18.8% while achieving superior optimization results. Applied to the X2TD coaxial rotor, the CNN-TL-NSGA-II framework improves FM by 9.58% and L/De by 17.75%, demonstrating significant potential for complex aerodynamic design tasks.