Modeling of dynamic characteristic of particle in transient gas–solid flow via a machine learning approach

Abstract

Based on a reliable dataset of particle dynamics from DNS, a R–CNN model combining the functional structures of RNN and CNN was proposed to systematically learn both the temporal and spatial inhomogeneities of particulate flow and predict the particle dynamic. The particle pattern matrix sequence with length of 10 ( ms ) and β after a specific period of 1 ms were selected as the input and output. The matrix size was set as 25×25 according to the particle size of 0.5 mm and the proportion of particle size to the sampling domain of 1:25. Through validation and final testing, the R–CNN model maintained good predictive accuracy and robustness (validation: δ ¯ = 0.13 , R 2 = 0.63; testing: δ ¯ = 0.12 , R 2 = 0.64). The model was proven to be effective with an appropriate input sequence length of 3 ( ms ) and in an appropriate time span of 4 ms . The model performance is affected by the comprehensiveness of local particle pattern. The particle resolution corresponding to the best performance of the model was 0.8–1. Additionally, adding the feature ( u r ) that do not have a high correlation with the target is not necessarily effective in improving the model's performance. Overall, the feasibility of a physical–meaning–oriented R–CNN model with an optimizable combination of functional architecture and parameters was confirmed. • A coupled R–CNN model was proposed to predict gas–solid drag force based on true DNS. • R–CNN model was proved to be more accurate and robust than pure RNN and pure CNN model. • Accuracy can be guaranteed under a proper range of temporal variables in R–CNN model. • Proper resolution to capture particles' pattern can lead to an optimal performance. • Enriching the feature with low correlation is not necessarily effective.