A continuum–discrete multiscale methodology using machine learning for thermal analysis of granular media

Abstract

This work presents a data-driven continuum–discrete multiscale methodology to simulate heat transfer through granular materials. The two scales are hierarchically coupled, where the effective thermal conductivity tensor required by the continuous method at the macroscale is obtained from offline microscale analyses. A set of granular media samples is created through the Discrete Element Method (DEM) to relate microstructure properties with thermal conductivity. The protocol for generating these Representative Volume Elements (RVEs) and homogenizing the microscale response is presented and validated by assessing the representativeness of the granular assemblies. The study found that two local properties, the porosity and the fabric of the material, are sufficient to accurately estimate a representative thermal conductivity tensor. The created dimensionless database of microscale results is used for training a surrogate model based on machine learning. In this way, effective thermal conductivity tensors that accurately reflect the local microstructure can be efficiently predicted from the surrogate model by taking the microstructural properties as inputs. The proposed multiscale methodology enables us to solve heat problems in granular media using a continuum approach with accuracy comparable to a pure discrete computational method but at significantly reduced computational cost.