Reproduce “Efficient low-thrust trajectory data generation based on generative adversarial network” on OpenResearch

About this paper

Ruida Xie, Andrew G. Dempster

Researchers at the University of New South Wales developed a Generative Adversarial Network (GAN) to efficiently synthesize feasible low-thrust trajectory data, increasing the data generation convergence rate from 7.16% to 84.3%. The GAN-generated data accurately replicated the statistical distribution of real low-thrust transfers for space mission design.

Key insights

• GANs can effectively learn and replicate the complex, often non-Gaussian, statistical distributions of feasible low-thrust trajectories, significantly improving data generation efficiency.
• A specific architectural balance is crucial for stable GAN training: the generator generally requires a deeper but proportionally thinner structure than the discriminator, with careful tuning of learning and dropout rates.
• Monitoring validation accuracy alongside traditional loss and score metrics is essential to detect mode collapse, where the GAN produces limited sample diversity despite seemingly convergent training metrics.

Summary by alphaXiv.