How CFD-DEM Simulations benefit from Machine Learning

Computational Fluid Dynamics-Discrete Element Method (CFD-DEM) simulations have become a valuable tool in industry and academia. In the context of combustion systems, CFD-DEM simulations are mainly relevant as tools for basic research, e.g., to understand phenomena such as segregation or other effects due to incomplete (particle) mixing. Typically, such simulations face the following challenges: (i) modelling of complex-shaped and deformable/shrinking particles, (ii) considering (cohesive) particle interaction forces that might change during the combustion process, or (iii) predicting (radiative) energy transfer rates that are sensitive to a large number of parameters that are almost impossible to measure directly. All these challenges have in common that a large number of predictions (or decisions) have to be made in a short time, and that these predictions depend on a large number of inputs (“features”) in a non-trivial way. Machine Learning (ML)-based algorithms are promising candidates to tackle these challenges.
In our present contribution we highlight three concrete examples of how ML-empowered CFD-DEM simulations can be realized: (i) ML for closure construction in the field of particle-particle radiative energy transfer [1], (ii) ML-empowered drag force modeling [2], as well as (iii) ML-aided calibration workflows for advanced parameter identification in the context of granular flow and bulk heat conductivity prediction using CFD-DEM. We will conclude our talk with an application study of a problematic reactive granular material encountered in industrial NiMH battery recycling.