A hybrid mesoscale closure combining CFD and deep learning for coarse-grid prediction of gas-particle flow dynamics.

[Display omitted] • Mesoscale drag and solids stress models are developed via a DL algorithm. • Validation reveals the importance of mesoscale solids stress. • Optimal conditions for anisotropic closure are obtained and validated. • The use of DL does not significantly increase the computational cost. This study develops filtered two-fluid model (fTFM) closures by coupling computational fluid dynamics (CFD) and deep learning algorithm (DL) for enabling coarse-grid simulations at reactor scales. Mesoscale drag, solids pressure and viscosity are modeled using an isotropic or anisotropic method. Subsequently, a priori analysis and a posteriori analysis of the present models along with other previously proposed closures are conducted. Comparison with the experimental data covering a broad range of operating conditions indicates that the mesoscale solids stress can be neglected in bubbling and turbulent fluidization regimes. However, the contribution of solids stress is clearly not insignificant at very low superficial gas velocities. Moreover, the drag model considering the anisotropy shows better prediction performance in the turbulent fluidization regime. In short, the present study develops and validates a DL-fTFM coupling algorithm applicable for gas-particle simulations. [ABSTRACT FROM AUTHOR]