Multipoint optimization for supercritical airfoils

Reliably speeding up multipoint aerodynamic optimization

Key Words: multipoint aerodynamic optimization, surrogate model, uncertainty quantification

INFO

  • Jan. 2020 — Dec. 2024
  • supported by the Natural Science Foundation of China (NSFC) No. 92052203

In this project, we built a Prior-based multi-fidelity aerodynamic optimization framework for the multipoint performance of supercritical airfoils. The key idea is to evaluate design-point performance with CFD, and the off-design points with pretrained machine-learning models. Considering the speed of ML model only slightly scale with the number of off-design points, the framework should optimize multipoint performance in a similar speed of single-point optimization.

In addition, the CFD-simulated design-point flow field can be a good prior when predicting the off-design point flow field with the model. By introducing such prior, the generalization ability of the pretrained model should be improved. This is very favorable for surrogate-base optimization, where we want to amortize the training cost by reusing the model in different optimization cases.

The project is in three steps:

  1. Building the model to predict off-design point flow fields with a prior of the design point CFD result, and verifying the advantage of this on generalization ability.

    This lead to the paper: Yang, Y., Li, R., Zhang, Y., & Chen, H. (2022). Flowfield Prediction of Airfoil Off-Design Conditions Based on a Modified Variational Autoencoder. AIAA Journal, 60(10), 5805–5820. https://doi.org/10.2514/1.J061972

  2. Applying the model in a multi-fidelity optimization framework, and verifying it in transonic buffet optimization.

    This lead to the paper: Yang, Y., Li, R., Zhang, Y., & Chen, H. (2024). Fast Buffet-Onset Prediction and Optimization Method Based on Pretrained Flowfield Prediction Model. AIAA Journal, 62(8), 2979–2995. https://doi.org/10.2514/1.J063634

  3. Further develop a uncertainty-aware framework to improve optimization results.

The dataset and codes are available on GitHub

References