Neural-network based approach for modeling wall-impact breakage of agglomerates in particle-laden flows applied in Euler–Lagrange LES

Abstract

The present study proposes a novel modeling approach for predicting the wall-impact breakage of agglomerates in wall-bounded particle-laden turbulent flows based on artificial neutral networks (ANN). The suggested model is especially useful for efficient Euler-Lagrange simulation methods relying on the hard-sphere approach and the equivalent-sphere model for the agglomerate structure, allowing LES predictions of high mass loadings. Based on the impact conditions, i.e., the impact velocity, the impact angle, the number of included primary particles and the diameter of the primary particles, the outcomes of the breakage events are forecasted using two pre-trained artificial neural networks. The first network is concerned with the prediction of the possibility of breakage and the resulting fragment size distribution, whereas the second network predicts the post-breakage velocities of the fragments. The supervised training of the employed networks relies on a database obtained by extensive DEM simulations of agglomerate wall-impacts covering wide ranges of impact conditions, which were partially reported in Khalifa and Breuer (2020, 2021) for developing a breakage model based on a dimensional analysis and regression techniques. In the present contribution, the database mainly comprising the normal or oblique impact case is extended by adding results for the practically relevant shear impact case of wall-bounded particle-laden flows at extremely small impact angles, i.e., 3◦ and practically flat (0.2◦ ). In addition, the breakage behavior of agglomerates containing very small numbers of particles are investigated under different impact angles and primary particle sizes. The ANN model is employed in Euler-Lagrange LES predictions of duct flows taking three Reynolds numbers and agglomerates of two powders distinguished by the size of the primary particles into account. The results obtained are compared with those based on a previous regression model (Khalifa and Breuer, 2021). In general, a good agreement between the results is found. However, the new ANN model is more widely applicable since the shear impact case is taken into account, which leads to subtle differences enhancing the reliability of the predictions.