Session: Government Agency Student Posters

Paper Number: 173065

Physically Constrained Generative Artificial Intelligence for Feasible Space Transformation in Design Optimization 

Urban air mobility (UAM) has been tagged as the next revolutionary transportation technology, which is expected to reduce ground traffic congestion in urban areas. Electric vertical takeoff and landing (eVTOL) aircraft, as a special type of UAM, have been chosen over traditional helicopters due to their reduced emissions and lowered operational costs. However, eVTOL are precluded from long-haul routes due to the battery energy consumption, especially during the takeoff phase which demands excessive energy. Multidisciplinary analysis and optimization (MDAO) discovers optimal takeoff trajectories with the minimum energy consumption, but conventional MDAO iteratively evaluates coupled, high-fidelity simulation models making the optimization computationally prohibitive. Surrogate models accelerate the MDAO, but naive surrogate modeling may result in enormous training costs to achieve sufficient accuracy for high-dimensional problems. Furthermore, MDAO typically involves complex nonlinear design constraints (especially in practical applications), which hinder the optimization search or even lead to failed convergence.

To tackle the above challenges, this work proposes physically constrained generative adversarial networks (physicsGAN) to transform an original design space to a reduced, feasible space where all design candidates inherently satisfy all design constraints. This is achieved by leveraging data-driven generative adversarial networks (GAN) to generate realistic aircraft control profiles and using surrogate models to check the feasibility of the generated control profiles. In the meantime, the data-driven GAN manages to reduce the original design space to lower dimensions which facilitate surrogate modeling. The surrogates include feedforward neural networks for predicting the scalar quantity (i.e., energy in this work) and long short-term memory networks for vector quantities (e.g., acceleration trajectories and angle of attack trajectories) during the takeoff process. Eventually, the training loss of the physicsGAN is penalized if any constraints are violated according to the surrogate predictions.

The physicsGAN-enabled surrogate-based trajectory optimizations were demonstrated on the Airbus A³ Vahana aircraft, which is a tandem, tilt-wing eVTOL drone. Results showed that the physicsGAN reduced the original design space from 41 to three dimensions, while maintaining sufficient variability and feasibility. Through fitting optimizations, the physicsGAN was shown to maintain over 99% accuracy. Meanwhile, the reduced design space is in nature a feasible design space where the physicsGAN achieved 98% feasibility across the flight condition space consisting of mass and efficiency. The physicsGAN-based optimal takeoff trajectory obtained over 93% accuracy compared against the simulation-based optimal references. Besides, the physicsGAN-enabled approach reduced the total optimization time by around three and over two orders of magnitude compared with the simulation-based references and data-driven GAN-based counterparts, respectively.

This work has three primary contributions. First, data-driven GAN effectively exploits a reduced, realistic design space by reducing the original design space from 41 to four dimensions and generating only realistic control profiles. Second, surrogates are trained on the reduced, realistic space and used to check the feasibility of generated control profiles. Third, the physicsGAN is proposed to further transform the realistic design space (by the data-driven GAN) to a further reduced, feasible design space. The feasible space has three dimensions while maintaining sufficient variability. Thus, complex, nonlinear constrained MDAO can be rapidly completed using unconstrained optimization techniques at a significantly reduced complexity and enhanced efficiency. The proposed physicsGAN can also be extended to other engineering areas due to the great generality.

Presenting Author: Sam Sisk Missouri University of Science and Technology

Presenting Author Biography: Sam is a 3rd year PhD student in aerospace engineering.
He specializes in surrogates, generative AI, and multi-disciplinary analysis and optimization.

Authors:

Sam Sisk Missouri University of Science and Technology
Xiaosong Du Missouri University of Science and Technology