Physics-informed reduced-order modelling with equivariant spectral submanifolds

What the math gives to ML

The transferable asset is the use of a low-dimensional invariant manifold whose chart and internal dynamics obey the same group action as the full system. This provides a principled way to build latent neural ODEs or state-space models that preserve symmetry while restricting trajectories to a dynamically meaningful nonlinear manifold. The most promising implementation is to train an equivariant encoder-decoder and latent vector field with an explicit spectral-submanifold invariance residual. This should reduce rollout cost and improve long-horizon stability when data evolve near a hyperbolic equilibrium and its dominant spectral modes.