It is well known that the microstructure of a material has pronounced influence on its macroscopic response. Therefore, a detailed understanding of the morphology, evolution, behavior, and mechanical response of materials at microscale is crucial. In this spectrum, multiscale finite element approaches, particularly FE2 , which bridges smaller and larger scales, have gained increasing attention. Despite their flexibility in modeling, FE2 simulations can be computationally demanding, which comes from the fact that for each macroscopic integration point, a corresponding RVE (Representative Volume Element) has to be solved in a nested scheme. In order to mitigate this bottleneck, several methods have been proposed, ranging from parallel computing to neural network based approaches. In this spirit, this work seeks to validate the Physically Recurrent Neural Network (PRNN) approach by Maia et al. [1], which embeds classical constitutive laws into a recurrent layer to preserve path-dependence. To this end, a particulate composite is devised in which glass inclusions (spherical) are embedded in an elasto-plastic polymer matrix. The training dataset is generated from virtual RVEs with random inclusion distributions, where the matrix is modeled with the pressure-dependent plasticity model for polymers by Melro et al. [2] and the glass inclusions remain linear elastic. The methodology of this work relies on experimental and numerical investigations. An experimental campaign is conducted to characterize the material properties and to study the response under loading and unloading cycles. The proposed surrogate model is assessed in terms of accuracy and computational speed-up with respect to full-order FE2 simulations and experimental data. The results show that the hybrid architecture can achieve substantial reductions in computational cost while retaining the physical interpretability of classical constitutive modeling, making it a promising tool for nonlinear structural analysis of particulate composites.