Geometric Regularization Forces Autoencoder Latent Spaces to Respect Manifold Structure of Dynamical Systems

arXiv·low signal

Hill and Ye propose a geometric regularization method that forces autoencoder latent spaces to respect the underlying manifold structure of stochastic dynamical systems, enabling reduced simulators from short-burst observational data. Applicable to molecular dynamics, climate modeling, and fluid simulations where high-dimensional systems evolve on unknown low-dimensional manifolds.