Machine Learning-Based Filtered Drag Model for Cohesive Gas-Particle Flows

Coarse-grid simulations of large-scale gas-solid flows using a filtered two-fluid model (fTFM) require appropriate sub-grid closure models to approximate unresolved physical phenomena. Such a sub-grid closure should be accurate enough to account for the effects of the inhomogeneous particle distribution. Several constitutive models are available in the literature for non-cohesive gas-solid flows, while they are not applicable for cohesive flows. Therefore, we aim to investigate the dependency of the drag force closure on the cohesion level, and integrate it into a drag correction concept based on machine learning (ML).
To do so, the results of fully-resolved CFDDEM simulations of cohesive gas- article flow are filtered with different filter sizes to develop a new drag closure. In detail, we simulated different systems by changing the cohesion level from cohesionless to highly cohesive, and the size of the systems, via coarse-graining. Afterwards, a dataset for the ML algorithm was created, and various markers were analyzed. Subsequently, a neural network-based drag correction model was created, trained, and tested with the identified markers. Finally, we benchmark the accuracy of the developed models for a range of cohesion levels.