The design of aerodynamic shapes with desirable parameters or optimized geometry has always been an essential scientific-applied problem. However, the computational cost of inverse design and aerodynamic optimization methods has been a challenge in this field. Nowadays, heavy efforts and investments have been concentrated on finding data-driven solutions to various problems. In this research, the data-driven acceleration of an inverse design method was investigated to reduce its computational cost. Due to the superiority of iterative aerodynamic inverse design methods resulting from obvious reasons such as its applicability to complex flows, the data generated from the previous iterations of the inverse design algorithm was used to have a better estimate of the target geometry and to accelerate the shape modification process. For this reason, two deep learning models, an FNN and an LSTM network, with various representations of input data were investigated to recognize not only the correlation between pressure distribution and geometry but also the trend of geometry and pressure distribution changes toward their targets. The deep learning model was trained by the generated data from the early steps of the inverse design, and the developed method was validated using various airfoils in a compressible high-Reynolds flow, in both viscous and inviscid cases. The results showed a 70 and 80 percent reduction in the computational cost for the offline and online use of the machine learning module with inverse design algorithm, respectively. The limitations of inverse design methods in terms of optimality, accuracy, and actuality of the target pressure distribution prevent further developments of such methods. In the second part of the research, an Artificial Neural Network (ANN) was incorporated into the Elastic Surface Algorithm (ESA) to resolve the aforementioned weaknesses while maintaining the strengths of inverse design methods. The database resulted from the solution of the fluid flow using CFD tools during the inverse design was used to train a deep learning model to correlate aerodynamic coefficients with pressure distributions. Then, a genetic algorithm was employed to achieve the optimized target pressure distribution, which its cost function is determined by the developed neural network or deep learning model. In order to eva‎luate the hybrid machine learning-inverse design method, the pressure distribution of the FX63-137 wind turbine airfoil was optimized at a determined angle of attack to reach the maximum lift-to-drag ratio as the cost function. During the online process of inverse design, the corresponding optimum pressure distribution was obtained and the unreal pressure distributions corresponding to fish-tail geometries were eliminated automatically by the neural network. The results of optimization showed the lift-to-drag ratio to increase by 18% for the FX63-137 airfoil. In addition, various geometries with close lift-to-drag ratios were achieved during the optimization.

مهدي نيلي احمدآبادي

محسن دوارده امامي، احمدرضا پيشه‌ور