A Machine Learning Study of Predicting Mixing and Segregation Behaviors in a Bidisperse Solid–Liquid Fluidized Bed

Abstract

In this work, a convolutional neural network combined with a long short-term memory model (CNN-LSTM) is employed to predict the mixing and segregation behaviors in a bidisperse solid–liquid fluidized bed (SLFB). The data set comes from the CFD-DEM simulations under a range of superficial inlet velocities vl and size ratios dl/ds and includes detailed particle information in different temporal and spatial dimensions. The CNN-LSTM model uses CNN to preprocess the original data, and then the output of CNN is regarded as the input of the LSTM model for model training. Two scenarios are considered: (1) varied vl under the same dl/ds; (2) varied dl/ds under the same vl. The training effects of LSTM and CNN-LSTM models are compared using loss function, random model tests, and several quantitative error evaluation indices (R2 score, MAE (mean absolute error) and RMSE (root mean squared error)). The results show that the CNN-LSTM model has better training effects than the LSTM model. Further, the prediction results are used to characterize the mixing and segregation behaviors by a series of indices and compared with the CFD-DEM simulations. The comparisons indicate that the CNN-LSTM model can obtain more reliable predictions than the LSTM model, because the CNN-LSTM model can capture more complex features after convolutional processing even though in the less occupied grids. The comparison of computation time cost indicates the high efficiency and accuracy of the CNN-LSTM model. This work provides a fast and accurate prediction method of the granular flow behaviors on temporal and spatial scales.