Learning Constrained Dynamical Embeddings for Geophysical Dynamics
Abstract
In this work, we investigate the implementation of physical constraints for the regularization of linear quadratic dynamical representations of partially observed systems. We focus on energy preserving quadratic terms and propose to enforce this constraint within the learning criterion of the models. We further demonstrate on the Lorenz 63 system that the generalization performance is significantly improved to states beyond the attractor spanned by the observation data when this constraint is satisfied.
Origin : Files produced by the author(s)