Artificial neural network-based EMMS drag model for fluidized beds

The drag model is an influencing factor for the gas-solid fluidized bed dynamics in the two-fluid simulation. However, accurately capturing the dynamics of particle and gas phases in these drag models, which combine both phases in the simulations, can be highly computationally intensive across different scales, such as macro, meso, and microscale. To address this challenge, the Energy Minimization Multi-Scale (EMMS) drag model has been developed to capture these scales. The EMMS model incorporates the heterogeneous behavior of the dilute and dense phases with small-scale fluctuations. In this study, Euler-Euler based Computational Fluid Dynamics (CFD) simulations were conducted for fluidized beds to investigate the heterogeneous behavior of gas-particle flows and generate datasets. These generated datasets were utilized to train an Artificial Neural Network (ANN) for closure, specifically to predict drag forces for the CFD. The CFD data, supplemented by the ANN model, was integrated into the simulation solver as a user-defined function (UDF) for the drag model. This EMMS-ANN approach resulted in predictions similar to observed characteristics of gas-solid flows while maintaining accurate predictions including volume fraction, heterogeneous index, and fluid flow behavior, for both dense and dilute flow regimes. Furthermore, the performance of the EMMS-ANN drag model was compared favorably with that of the simple EMMS drag model, demonstrating reduced computational time by 24 % in 3D and 40 % in 2D cases. Despite the demands associated with implementing the coupled Machine Learning (ML) model within the CFD solver, its effectiveness suggests promising applications, particularly at industrial scales.