Case Studies of Generative Machine Learning Models for Dynamic Systems

What is it about?

Engineers use simulations to design and test aircraft and spacecraft because real-world experiments are expensive. However, simulation results often differ from actual system behavior due to errors and simplifications. This study explores whether generative artificial intelligence models (GAIMs) can reduce these differences. GAIMs learn from data to produce outputs with similar patterns, but unlike image or text models, aerospace GAIMs must follow natural laws, such as conservation principles. The study focuses on two navigation problems common in aircraft guidance: flying in a wind field in the shortest time and avoiding threats while minimizing exposure. We train GAIMs using limited data while ensuring outputs obey mathematical rules based on system dynamics. We test three GAIM types: generative adversarial networks (GANs) and two types of variational autoencoders (VAEs). Results show that VAEs perform best, generating realistic and mathematically valid data despite limited training examples. These findings suggest that GAIMs can help improve aerospace modeling by producing reliable simulations that align more closely with real-world conditions.

Featured Image