This contribution presents a two-stage approach to the data-driven identification of nonlinear constitutive models. In the first stage, we demonstrate an automated calibration framework for phenomenological (macro-scale) models based on multiscale simulations. The approach uses fine-scale data generated from detailed mesoscale models of masonry [1], capturing the effects of geometric and material heterogeneity. Model parameters at the macro level are optimized to reproduce the macroscopic response via a computational homogenization scheme. In the second stage, we explore the use of Kolmogorov–Arnold Networks (KANs) [2] for learning constitutive behaviour directly from microstructural data. We apply KANs to hyperelastic materials using data obtained from representative volume elements (RVEs), showing their ability to capture complex stress–strain relationships with enhanced interpretability and generalization.