A hybrid CFD – Deep learning methodology for improving the accuracy of pressure drop prediction in cyclone separators

Abstract

Cyclone separators are widely used in many industries to separate solid particles from the fluid flow. As same as separation efficiency, pressure drop is considered as a primary parameter to evaluate performance of cyclone separators. Computational Fluid Dynamics (CFD) is a powerful method to predict the pressure drop of cyclone separators. However, it is indisputable that CFD technique is mathematically complicated and computationally expensive. This study has proposed a more accurate, computationally efficient hybrid method to predict cyclone pressure drop by merging CFD and Deep Learning algorithms. Its generalization has been validated by k-fold cross validation. The proposed hybrid method has also been proved to be more superior than the DNN regression method. The results verify that the proposed method not only predicts cyclone pressure drop accurately (with a maximum error less than 5% in comparison with experimental data), but also requires less computational time than traditional CFD. In other words, it leverages the potential of novel approaches to decrease CFD computational cost while increasing the accuracy level.