End-to-End Deep-Learning-Based Surrogate Modeling for Supersonic Airfoil Shape Optimization

Aerodynamic shape design optimization faces challenges due to the computational demands and the vast design space, limiting its practicality and scalability. While progress has been made in subsonic and transonic regimes, the real-time optimization for supersonic conditions remains unexplored. To bridge this gap, this work exploits knowledge learned from subsonic and transonic real-world data and introduces a rapid optimization framework tailored for the supersonic regime. A novel end-to-end multitask Convolutional Neural Network is proposed to predict the aerodynamic coefficients of an airfoil shape, extracting global and local features directly from the geometry. The surrogate model is thoroughly examined and validated, including an analysis of model explainability. The surrogate model achieves on par results with the state-of-the-art, with relative errors in aerodynamic coefficient predictions below 1.7%. Furthermore, a surrogate-based optimization strategy integrates the surrogate model with a Generative Adversarial Network to generate realistic airfoil shapes, thereby reducing the design space to a low-dimensional representation. This approach provides a robust solution that accelerates the optimization routine by over 3000 times when compared to simulation-based methods while achieving a deviation of less than 1.9% from their optimum performance. Overall, this work strikes a balance between efficiency and effectiveness without compromising reliability.