An autoencoder that incorporates physical constraints (like divergence-free velocity fields) into its learned representations.