Accurate constitutive modelling of conventional materials has always been a key task. The conventional method for constitutive modelling of materials is to propose an analytical model with some constants, known as material parameters, to be determined from data. However, 3D printing brings new possibilities and among them, the design of metamaterials, which are considered themselves as “materials”. One of the key aspects of metamaterials is that they constitute a new wide space of possible materials as seen at the continuum scale. However, in contrast with classical materials, there is no simple classical model for the continuum treatment of these materials. Hence, a promising approach is the use of surrogate models, for example derived from machine learning techniques, to characterize the complex nonlinear behavior at the continuum scale. Another possible approach is to develop surrogate models based on energy equivalences or based on classical models as those used in other materials like polymers and soft tissues. An important aspect in both cases is the capability of performing inverse analyses so given a desired continuum property, we are able to determine the metamaterial design that best fits that purpose. The present contribution weights the advantages and disadvantages of classical models and machine learning models in representing the nonlinear behavior or materials at different scales. We first compare predictions for polymers and soft tissues, with special attention to how the crossing of scales may be performed, and then address how this learning may translate to the continuum-based modeling of metamaterials.