Aerodynamic optimization of aircraft wings using machine learning

Abstract

This study proposes a fast yet reliable optimization framework for the aerodynamic design of transonic aircraft wings. Combining Computational Fluid Dynamics (CFD) and Machine Learning (ML), the framework is successfully applied to the Common Research Model (CRM) benchmark aircraft proposed by NASA. The framework relies on a series of automated CFD simulations, from which no less than 160 planform variations of the CRM wing are assessed from an aerodynamic standpoint. This database is used to educate an ML surrogate model, for which two specific algorithms are explored, namely eXtreme Gradient Boosting (XGB) and Light Gradient Boosting Machine (LGBM). Once trained with 80 % of this database and tested with the remaining 20 %, the ML surrogates are employed to explore a larger design space, their optimum being then inferred using an optimization framework relying on a Multi-Objective Genetic Algorithm (MOGAO). Each ML-based optimal planform is then simulated through CFD to confirm its aerodynamic merits, which are then compared against those of a conventional, fully CFD-based optimization. The comparison is very favourable, the best ML-based optimal planform exhibiting similar performances as its CFD-optimized counterpart (e.g. a 14 % higher lift-to-drag ratio) for only half of the CPU cost. Overall, this study demonstrates the potential of ML-based methods for optimizing aircraft wings, thereby paving the way to the adoption of more disruptive, data-driven aircraft design paradigms.