Learning shared patterns across different dynamical systems

Related dynamical systems often share structure, but existing methods estimate each system's operator independently, missing these connections. DOODL learns a low-dimensional manifold of characteristic spectral dynamics whose combinations approximate all systems in a family, yielding interpretable embeddings and faster operator estimation from short trajectories. On metastable Langevin dynamics and plasma simulations, the approach achieves 1–2 orders of magnitude lower error than independent estimation methods in data-scarce settings.