In this work we present a computational methodology to infer material properties in multiscale systems using a range of experimental scenarios, by leveraging a hierarchical Bayesian framework [1]. This methodology integrates data from experiments conducted at varying length scales and material compositions, systematically combining this information to improve inference accuracy. A probabilistic model is developed using the Transitional Markov Chain Monte Carlo (TMCMC) method [2], allowing efficient sampling from the posterior distributions of multiscale model parameters and hyperparameters. The hyperparameter posterior distribution synthesizes information from multiple experiments, and it is utilized to yield a refined set of physical parameters suitable for probabilistic predictive modeling in new materials. To address the immense computational demands, feedforward neural networks are employed to learn and emulate the nonlinear constitutive behavior across scales [3], streamlining the Bayesian analysis. The framework’s effectiveness is demonstrated in a case study on carbon nanotube (CNT)-reinforced cementitious materials, particularly analyzing the CNT interfacial mechanical properties.