(701d) Application of a Neural Network Approach to Predict the Filtered Drag Force for Gas-Solid Flows

Gas-solid flows with complicated meso-scale inhomogeneities are common in energy and chemical industry. To accurately predict the filtered drag force, filtered two-fluid models have been developed based on correlative analysis of fine-grid two fluid model simulation data. However, these traditional models are found to be unsatisfactory in the prior analysis of predicting the filtered drag force. Recently, data science, especially deep learning, has been used to solve challenge problems in the field of fluid mechanics. In our work, a deep learning approach based on a proposed artificial neural network (ANN) is exploited to predict the filtered drag force for coarse-grid simulations. The proposed ANN model includes an input layer, hidden layers, and an output layer. The hidden layers are divided into three convolutional layers, one flat layer, and two fully-connected layers. The correlation coefficient analyses and probability distribution function of the relative error are used to evaluate the performance of the proposed ANN model. It is found that the proposed ANN model is significantly superior to the multi-layered perceptron neural network model as well as the best available traditional functional model named the dynamic scale-similarity model. The effect of three important parameters, namely, the domain-averaged solid volume fractions, filter size, and convolutional kernel size on the filtered drag force is analysed. The predictability of the proposed ANN model generally increases with the decrease of the domain-averaged solid volume fractions and the increase of the filter size. The predicted results are particularly better when the convolutional kernel size is chosen as three times of the filter size. This finding is significant and indicates that the local filtered drag force is affected profoundly by the flow properties in the near neighborhood in coarse-grid simulations. To better understand the underlying physics behind this finding, a budget analysis of the filtered momentum equation for the solid phase is performed. Moreover, the inclusion of the lateral gas-phase pressure gradient that is perpendicular to the flow direction is also helpful in accurate prediction of the filtered drag force.

Reference

1. Jiang, Y., Kolehmainen, J., Gu, Y., Kevrekidis, Y.G., Ozel, A., Sundaresan, S., “Neural-network-based filtered drag model for gas-particle flows,” Powder technology, 346, pp. 403-413 (2019).
2. Ozel, A., Gu, Y., Milioli, C.C., Kolehmainen, J., Sundaresan, S., “Towards filtered drag force model for non-cohesive and cohesive particle-gas flows,” Physics of Fluids, 29 (10), pp. 103308 (2017).