Construction and Analysis of the Mesoscale Drag Force Model Based on Machine Learning Methods

Abstract

The presence of mesoscale structures in gas−solid flows significantly complicates the constitutive relationship of the gas−solid drag force in coarse-grid simulations. This study employs artificial neural networks to evaluate the performance of various filtered quantities in predicting the mesoscale drag force. Our findings indicate that the drag model solely relying on local filtered quantities, such as solid volume fraction, slip velocity, or gas pressure gradient force, is unable to achieve the desired level of accuracy. The consideration of neighboring solid volume fractions significantly enhances the performance of the drag model, particularly in the dilute regions. In two-dimensional systems, the solid volume fractions at the eight grids closest to the considered grid are used. Additionally, the filtered solid volume fraction gradient and the filtered solid volume fraction at a second scale present a viable alternative to replace the eight neighboring solid volume fractions. Such an alternative offers an acceptable level of accuracy and increased simplicity when functional drag models are pursued.

Multiscale Modeling and Artificial Intelligence for Multiphase Flow Science

“…Artificial neural networks (ANNs) were utilized to improve the accuracy of drag force predictions in gas–solid flows with mesoscale structures. By incorporating neighboring solid volume fractions and filtered quantities, such as solid volume fraction gradients, Zhang et al found that the drag model’s performance can be significantly enhanced, especially in dilute regions. Hardy et al employed filtered two-fluid models with explicit filtering and neural network models for closures such as drag force and solid phase stress, which have shown promise in reducing computational demands while maintaining accuracy.…”