Abstract
Electric vertical takeoff and landing (eVTOL) aircraft make a unique form of urban air mobility due to their low noise, zero emission, and precision control. To maximize efficiency, high-fidelity simulation-based multidisciplinary design optimization discovers the optimal balance among subsystems within an eVTOL. However, conventional multidisciplinary design optimization is computationally intensive due to excessive high-fidelity model evaluations. Moreover, complex nonlinear constraints deteriorate optimization efficiency and convergence. While surrogate models enable efficient design optimization, surrogate modeling suffers in large-scale applications and surrogate-based optimization still has to deal with nonlinear constraints. To address these challenges, the authors' previous work proposed physics-constrained generative adversarial networks (physicsGAN) to implement acceleration-constraint-based loss during the training of GAN, which guided physicsGAN to generate not only realistic but also feasible design shapes (i.e., takeoff control profiles of an electric drone). This work extends the previous work by integrating more constraints (such as the final horizontal speed constraint) and investigates the performance on feasible space exploration and dimensionality reduction. The physicsGAN is expected to achieve two-fold dimensionality reduction: i) implicit reduction by generating only realistic, feasible control profiles; and ii) explicit reduction by explicitly lowering down the design-space dimension. The physicsGAN was demonstrated on optimal takeoff trajectory design of the Airbus A3Vahana drone. Results exhibited that the physicsGAN explicitly reduced the input dimension from 41 to three variables with over 99% fitting accuracy on control profiles and over 95% on the total takeoff duration. In addition, physicsGAN achieved almost 100% feasibility coverage within the transformed feasible design space for variable flight conditions (i.e., mass and power efficiency in this work), such that unconstrained optimization can be conducted for the originally constrained optimization problem.
Recommended Citation
S. Sisk and X. Du, "Physics-Constrained Generative Adversarial Networks for Dimensionality Reduction in Optimization," AIAA Science and Technology Forum and Exposition AIAA Scitech Forum 2026, American Institute of Aeronautics and Astronautics, Inc., Jan 2026.
The definitive version is available at https://doi.org/10.2514/6.2026-2016
Department(s)
Mechanical and Aerospace Engineering
Publication Status
Full Access
International Standard Book Number (ISBN)
978-162410765-8
Document Type
Article - Conference proceedings
Document Version
Citation
File Type
text
Language(s)
English
Rights
© 2026 American Institute of Aeronautics and Astronautics, Inc., All rights reserved.
Publication Date
01 Jan 2026
Comments
Intelligent Systems Center, Grant 2501866