Maintaining traction remains one of the challenges that space exploration rovers encounter while roving on Martian or Lunar deformable terrains. Such terrains consist of granular regolith under reduced gravity conditions. Real-time simulation of wheel-soil interactions, that accurately takes gravity effects into account, can improve rovers’ online mobility control. The research problem investigated in this paper is the development of a machine learning-based wheel-on-soil simulation model. Machine learning enables efficient and fast mapping of simulation inputs to outputs, trained using high-fidelity (non-real-time) models validated by experiments. The training data is produced by a continuum method comprising a modern constitutive model, nonlocal granular fluidity (NGF), and a state-of-the-art numerical solver, material point method (MPM). Machine learning techniques, including graph network-based simulator (GNS) and principal component analysis (PCA), are proposed to learn efficient mappings. The important aspects that must be captured include the traction forces on the wheel and the behavior of the underlying granular flows. Hence, the experimental data includes force measured using an instrumented single-wheel testbed and subsurface soil motion observed with a high-speed camera and analyzed with optical flow.