Multi-point Airfoil Shape Optimization Using Neural Network-based Sequential Sampling
Pavankumar Koratikere, Leifur Leifsson · 2024
Multi-point aerodynamic design involves the design of aerodynamic surfaces at multiple operating conditions, including the design point and off-design points. In this work, multi-point aerodynamic design is addressed using surrogate-based optimization (SBO). Efficient global optimization (EGO) is a widely used method for SBO, but it can be limited by the use of kriging, which is typically used within it. A recently proposed EGO-like algorithm, called EGONN, addresses the limits of EGO by using neural networks in place of kriging. This work demonstrates how EGONN can be used to solve multi-point aerodynamic design problems. A single- and three-point optimization of RAE 2822 airfoil are solved using EGONN and the results are compared against EGO. For the single-point case, EGONN obtains a comparable solution to EGO (within 1.4 drag counts), which has a drag coefficient value 42% lower than the RAE 2822 airfoil. For the three-point case, EGONN also finds a comparable optimum shape as obtained from EGO. That shape performs better at the off-design conditions when compared to the single-point optimum. EGONN computational needs are only slightly higher than EGO.