Multi-Fidelity Aerodynamic Data Fusion with a Deep Neural Network Modeling Method.

Area of Science:

• Computational Fluid Dynamics
• Machine Learning
• Aerodynamics

Background:

• Aerodynamic data generation often involves trade-offs between fidelity and computational cost.
• Multi-fidelity data sources offer complementary information: low-fidelity for trends, high-fidelity for accuracy.
• Integrating diverse data sources is crucial for efficient and accurate aerodynamic modeling.

Purpose of the Study:

• To develop and validate a deep neural network (DNN) based method for fusing multi-fidelity aerodynamic data.
• To enhance the quality of aerodynamic predictions by leveraging both low- and high-fidelity datasets.
• To investigate the relationship between low- and high-fidelity aerodynamic data for effective fusion.

Main Methods:

• Applied a deep neural network (DNN) architecture comprising three fully-connected networks.
• The DNN approximates low-fidelity data and models linear/nonlinear correlations between low- and high-fidelity data.
• Generated low-fidelity (Euler) and high-fidelity (Navier-Stokes) data for an airfoil across various conditions.

Main Results:

• Constructed a fusion model for longitudinal lift (CL) and drag (CD) coefficients.
• The proposed DNN fusion method demonstrated superior performance over variable complexity modeling and cokriging.
• Evaluated accuracy using root mean square error and average relative deviation, showing satisfactory results on test cases.

Conclusions:

• The proposed DNN-based aerodynamic data fusion method is effective for integrating multi-fidelity data.
• This approach offers improved accuracy and performance compared to traditional data fusion techniques.
• The method holds significant potential for advancing aerodynamic modeling and analysis.