Bridging the gap between periodic domain and fluidized bed

To understand the impact of macroscale constraints on the mesoscale drag modeling, we performed fine-grid two-fluid model simulations in both the periodic domain and realistic fluidized beds, and used the artificial neural networks to identify the key markers for the drag force. It is found that only local coarse-grid variables are not sufficient, whereas inclusion of the drift velocity as a sub-grid marker facilitates a high predictive performance when translating between the periodic domain and realistic fluidized beds. The closures of the drift velocity are, however, highly dependent on the solids volume fraction, and those using only coarse-grid variables are not as successful as using fine-grid data. More efforts are needed to seek a generic closure of the drift velocity. In a broader sense, combining the mesoscience and machine learning enables us to identify the key factors in the theory building (here, mesoscale modeling) with the aid of connections built via machine learning.