03.1 - Aerodynamics – CFD Methods and Validation
A MACHINE-LEARNING-DRIVEN AERODYNAMIC SHAPE OPTIMIZATION FRAMEWORK USING FREE-FORM DEFORMATION AND DEEP NEURAL NETWORKS
G. Dessio Junior¹, G.L. Olichevis Halila¹, J. Messias Magalhães Júnior, Georgia Institute of Technology, United States; J.L. F. Azevedo, Instituto Tecnológico de Aeronáutica, Brazil; ¹Pontifícia Universidade Católica do Paraná, Brazil
Aerodynamic shape optimization (ASO) is a cornerstone of modern aerodynamic design, enabling system-rnatic performance improvements through high-fidelity computational fluid dynamics (CFD) analyses. Despiternits effectiveness, conventional ASO frameworks remain computationally expensive due to the requirement ofrnrepeated use of CFD solutions throughout the optimization loop, which significantly limits design-space explo-rnration. This work proposes a machine-learning-based aerodynamic shape optimization framework in whichrndeep neural networks (DNNs) are employed as surrogate models to replace high-fidelity CFD evaluationsrnduring the optimization iterations. The surrogate models are trained on a large database of airfoil geome-rntries generated through free-form deformation (FFD) of a NACA 2412 baseline airfoil and evaluated usingrnsteady-state Reynolds-averaged Navier–Stokes simulations. The DNNs predict lift, drag, and pitching momentrncoefficients as functions of geometric design variables and flight conditions. The trained neural networks arernembedded within a gradient-based optimization framework based on sequential least squares programming,rnwhere drag minimization is performed subject to lift and thickness constraints. Preliminary optimization re-rnsults demonstrate that the proposed approach achieves a drag reduction of approximately 14% relative to thernbaseline configuration while satisfying all imposed constraints, highlighting both the accuracy and robustnessrnof the surrogate-based optimization process. The results indicate that deep neural network surrogates canrneffectively replace high-fidelity CFD solvers within aerodynamic shape optimization loops, yielding substantialrncomputational savings without compromising solution quality. For the final paper, additional developments arernplanned, including the applicati
DOI: 10.71945/icas2026_0061