Human Body Models (HBMs) are finite element models used for making safety assessment in car crash simulations. The HBM skeleton is composed of cortical and trabecular bones that are modelled with shell and solid elements, respectively. The bone constitutive relations are currently relying on elasto-plastic models, implying a simplification of the biomechanical response. In this work, Recurrent Artificial Neural Networks (RNN) serve as a universal material model in explicit simulations. The primary objective is to establish a workflow that will allow learning directly from in-vivo data, the mechanical behavior of biomaterials and other complex confounding factors such as human variability. A synthetic data set was generated using single element simulations driven by noisy and smooth displacement paths. Subsequently, two models based on the Gated Recurrent Unit (GRU) were trained for cortical and trabecular bone according to the shell and solid element formulations. Finally, the models were verified through component-level simulations. To accomplish this an integration with LS-DYNA via user defined material subroutines was implemented. This data driven method showed high accuracy using a simple compact network architecture and the mean square error (MSE) as a cost function. Moreover, an analysis on the internal vector contributed to the RNN interpretability. The results revealed that the models learned to distinguish between linear elasticity and nonlinear behavior, therefore they discovered implicitly the history material variables. In conclusion, the findings of this work indicate that RNN can be applied to model hard biological materials for explicit simulations. Future research will encompass training on physical experimental data and extend the workflow to a Physics-Informed version. Using universal material models allows avoiding specific material model selection, calibration of material constants, meeting assumptions on the nonlinear behavior, and consequently it contributes to make the connection between material testing and simulation more reliable.