Multi-network collaborative lift-drag ratio prediction and airfoil optimization based on residual network and...

Area of Science:

• Aerospace Engineering
• Computational Science
• Artificial Intelligence

Background:

• Deep learning (DL) offers reduced computational cost for airfoil optimization compared to traditional computational fluid dynamics (CFD).
• A key challenge in DL-based airfoil optimization is prediction uncertainty, leading to deviations from true values.
• Existing single-network models struggle with accuracy and reliability due to inherent neural network uncertainties.

Purpose of the Study:

• To develop a robust multi-network collaborative prediction model for airfoil lift-to-drag ratio.
• To mitigate prediction errors and enhance the reliability of DL-driven airfoil optimization.
• To improve the accuracy of airfoil optimization by addressing neural network uncertainties.

Main Methods:

• Construction of a multi-network collaborative prediction model integrating ResNet architecture and penalty functions.
• Utilization of Latin supersampling to select critical angles of attack (2°-10°) for targeted uncertainty management.
• Implementation of the random drift particle swarm optimization (RDPSO) algorithm to control and minimize prediction errors.

Main Results:

• The multi-network collaborative approach significantly reduced prediction errors in airfoil optimization.
• A maximum error reduction of 16.0% was observed in single angle of attack optimization compared to single-network methods.
• The proposed model demonstrated enhanced reliability for deep learning-based airfoil optimization.

Conclusions:

• Multi-network collaboration effectively addresses prediction uncertainties in deep learning for airfoil optimization.
• The developed model offers a more reliable and accurate alternative to single-network approaches.
• This advancement improves the practical applicability of deep learning in aerodynamic design.