Master's Thesis

Deep Learning-based Aerodynamic Shape Optimization of the Common Research Model Wing

Francisco de Agrela Vital — 2026

Key information

Authors:

Francisco de Agrela Vital (Francisco De Agrela Vital)

Supervisors:

Frederico José Prata Rente Reis Afonso (Frederico Afonso); Fernando José Parracho Lau (Fernando José Parracho Lau)

Published in

June 30, 2026

Abstract

High-fidelity aerodynamic shape optimization is computationally expensive, as each function evaluation requires solving the governing flow equations over the full wing geometry. Surrogate-based optimization reduces this cost by replacing the solver with a fast approximate model, enabling rapid design space exploration. This thesis presents a deep learning-based aerodynamic shape optimization framework for three-dimensional wing configurations, applied to the NASA Common Research Model wing at transonic cruise conditions. A Wasserstein Generative Adversarial Network learns the modal coefficient distribution of transonic airfoils from the UIUC database and generates diverse cross-sections assembled into three-dimensional wings through Free-Form Deformation. Principal Component Analysis is applied to a database of generated wings to extract orthogonal deformation modes, reducing the 192-dimensional control point space to 50 modes and producing optimized geometries within 1% of the full parameterization in drag while reducing optimizer iterations by 82%. A multi-output Multi-Layer Perceptron surrogate trained on 4898 high-fidelity CFD evaluations achieves RMSE values of 7.68, 10.84, and 6.61 counts for drag, lift, and pitching moment, respectively, on an independent validation dataset, and is deployed within a gradient-based constrained optimizer exploiting automatic differentiation for gradient computation. Tested across five lift coefficient targets, the framework demonstrates generalization without retraining and a mean speedup of approximately 5200 times in CPU time relative to adjoint-based optimization, with mean absolute optimization gaps of 16.16, 15.13, and 6.65 counts for drag, lift, and pitching moment, respectively.

Publication details

Authors in the community:

Supervisors of this institution:

Fields of Science and Technology (FOS)

mechanical-engineering - Mechanical engineering

Publication language (ISO code)

eng - English

Rights type:

Embargoed access

Date available:

May 1, 2027

Institution name

Instituto Superior Técnico