Applying Reinforcement Learning (RL) to complex physical systems faces challenges with lengthy training times due to high computational simulation costs. Transfer learning can reduce training times by leveraging models trained on similar problems. Reduced Order Models (ROMs) create faster, simplified environments for RL training while retaining accuracy.

ROMs can be generated through various methods, including unsupervised learning approaches from data-driven science. Dynamical systems evolve through time and are described by state vectors and their derivatives. Real-world physical systems often exhibit nonlinearity, making them difficult to characterize.

These systems frequently lack solvable governing equations, requiring simplifications that limit their utility. Turbojet engine dynamics serve as an example problem where coupled equations must be solved iteratively rather than directly.