In this talk, I will present a novel modeling approach for large-eddy simulations (LES), where instead of modeling the entire range of small scales, a multilevel approach is adapted. The functional closure is obtained using tensor representation theory at each level in terms of filtered velocity gradients. A reduced order generalization is considered, which is then represented using deep neural networks, which can be trained on increasingly complex turbulent flows using transfer learning. For this talk, I will focus on isotropic and wall turbulence.