Location
Havener Center, St. Pat's Ballroom C
Presentation Date
April 21, 2023, 3:15pm-4:15pm
Session
Session 4
Description
Multi-Disciplinary Analysis and Optimization (MDAO) plays an important role in the future of aviation, especially relating to Urban Air Mobility (UAM). MDAO for UAM effectively quantifies and maximizes system performance and safety while minimizing cost or consumption. Electric drone takeoff trajectory design aims to minimize energy consumption while fulfilling acceleration and other constraints to keep passengers comfortable. Conventional simulation-based MDAO is computationally expensive due to repeatedly evaluating physics-based models while naïvely constructing machine learning predictive models, also known as surrogate models, do not scale well with large-scale problems due to the “curse of dimensionality”. A twin-generator generative adversarial networks (Twin-GAN) model is proposed to generate realistic electric drone take-off trajectories for the first time. One thousand optimal electric drone takeoff trajectories were generated under different constraints and used as training data for the Twin-GAN model. The generator of GAN is composed of two networks of the same architecture, called twin generators, to predict power and wing angle profiles separately. Results revealed that the proposed Twin-GAN reduced the design space from 43 dimensions to 11 dimensions while maintaining ~99% fitting accuracy towards 100 arbitrary real optimal designs. In the meantime, the Twin-GAN model automatically reduces the design space by filtering out unrealistic takeoff trajectories. A deep neural network surrogate was constructed based on the Twin-GAN model and achieved >95% relative accuracy using 2,385 training samples. Surrogate modeling without using GAN could hardly collect training samples due to convergence difficulty on unrealistic take-off trajectories. The surrogate model will soon be integrated into an optimization algorithm for optimal trajectory design in future work under the support of NASA-Missouri Space Grant Fellowship.
Meeting Name
32nd Annual Spring Meeting of the NASA-Mo Space Grant Consortium
Department(s)
Mechanical and Aerospace Engineering
Document Type
Presentation
Document Version
Final Version
File Type
text
Language(s)
English
Rights
© 2023 The Authors, all rights reserved.
Multi-Disciplinary Analysis Surrogate based on Twin-Generator Generative Adversarial Networks for Electric Drone
Havener Center, St. Pat's Ballroom C
Multi-Disciplinary Analysis and Optimization (MDAO) plays an important role in the future of aviation, especially relating to Urban Air Mobility (UAM). MDAO for UAM effectively quantifies and maximizes system performance and safety while minimizing cost or consumption. Electric drone takeoff trajectory design aims to minimize energy consumption while fulfilling acceleration and other constraints to keep passengers comfortable. Conventional simulation-based MDAO is computationally expensive due to repeatedly evaluating physics-based models while naïvely constructing machine learning predictive models, also known as surrogate models, do not scale well with large-scale problems due to the “curse of dimensionality”. A twin-generator generative adversarial networks (Twin-GAN) model is proposed to generate realistic electric drone take-off trajectories for the first time. One thousand optimal electric drone takeoff trajectories were generated under different constraints and used as training data for the Twin-GAN model. The generator of GAN is composed of two networks of the same architecture, called twin generators, to predict power and wing angle profiles separately. Results revealed that the proposed Twin-GAN reduced the design space from 43 dimensions to 11 dimensions while maintaining ~99% fitting accuracy towards 100 arbitrary real optimal designs. In the meantime, the Twin-GAN model automatically reduces the design space by filtering out unrealistic takeoff trajectories. A deep neural network surrogate was constructed based on the Twin-GAN model and achieved >95% relative accuracy using 2,385 training samples. Surrogate modeling without using GAN could hardly collect training samples due to convergence difficulty on unrealistic take-off trajectories. The surrogate model will soon be integrated into an optimization algorithm for optimal trajectory design in future work under the support of NASA-Missouri Space Grant Fellowship.
