Graph convolutional network‐based filtered drag model for coarse‐grid TFM simulations of gas–solid fluidized beds

Abstract

Accurate subgrid drag closure modeling is essential for coarse‐grid two‐fluid model (TFM) simulations of gas–solid fluidized beds to account for unresolved mesoscale structures. This study presents a graph convolutional network framework for filtered drag prediction across flow regimes ranging from dense bubbling to dilute flow. To construct the model input, fine‐grid TFM data are filtered at multiple scales and transformed into graph representations, where computational cells serve as nodes and neighborhood message passing captures mesoscale spatial information. Unlike conventional ANN‐based models that process each cell's markers independently, the proposed framework learns spatial dependencies and regime‐dependent mesoscale structures from local hydrodynamic markers. The model is evaluated across multiple particle systems, flow regimes, and filter sizes, achieving a test R ² of 0.97. Importantly, filter‐scale effects are implicitly learned from aggregated hydrodynamic information without an explicit filter‐size input. The model also shows cross‐particle transferability, generalizing to an unseen particle without retraining.