A hierarchical approach for modelling the thermal response of large-scale granular assemblies by coupling the micro-scale particle-level thermal interactions with the macro-scale continuum system is proposed. The coupling is done by using a machine learning tool that is trained to replicate the effect of discrete particle nature on the macro-scale system using finite elements. A trained Artificial Neural Network (ANN) tool that can estimate the effective local thermal conductivity for each finite element considering the influence of the presence of stagnant gas in the interstitial voids, gas pressure and the granular microstructure is used. This way of hierarchical coupling using ANN eliminates the need to perform thermal discrete element simulations for each finite element at every increment by directly predicting the effective local conductivity. The proposed hierarchical approach is applied to a breeder blanket of fusion reactor that consists of more than 15 million particles to demonstrate the efficacy of the method. The influence of the drop in gas pressure across the breeder unit and the heat generation on the temperature distribution of the full-scale breeder unit is analysed numerically. © 2019, OWZ.