Reinforcement Learning for Accelerated Aerodynamic Shape Optimization

Abstract

Often the choice of the geometry parametrization in shape optimization for fluid-dynamical tasks is characterized by the need to on the one hand find a low dimensional, efficient description of the considered configurations for fast computation of the optimization results, and, on the other hand be able to work with physically interpretable features [1]. Using a simple 2D airfoil design problem, we illustrate this dichotomy for the case of a reinforcement learning approach and suggest a procedure to maximize interpretability under a given highest dimensionality of the parameter space. The parametrization is a subfamily of the PARSEC-airfoil family [2]. The advantages in terms of physical interpretability and high descriptive power in comparison to NACA profiles is illustrated. [1] Marc G. Bellemare, Will Dabney, Robert Dadashi, Adrien Ali Taiga, Pablo Samuel Castro, Nicolas Le Roux, Dale Schuurmans, Tor Lattimore, and Clare Lyle. 2019. A geometric perspective on optimal representations for reinforcement learning. Proceedings of the 33rd International Conference on Neural Information Processing Systems. Curran Associates Inc., Red Hook, NY, USA, Article 392, 4358–4369. [2] R.W. Derksen, Tim Rogalsky, Bezier-PARSEC: An optimized airfoil parameterization for design, Advances in Engineering Software, Vol. 41, Iss. 7–8, 2010, pp. 923-930, ISSN 0965-9978, https://doi.org/10.1016/j.advengsoft.2010.05.002