Learning Quantities of Interest from Dynamical Systems for Observation-Consistent Inversion