Overcoming small datasets in machine learning studies of multi-physics flows in propulsion