Physics-Informed deep learning to predict flow fields in cyclone separators

Abstract

Cyclone separators are devices whose modeling still represents a significant challenge, despite their vast use as a fundamental component in different industries. Computational Fluid Dynamics (CFD) simulations have been proved to be a suitable tool to predict the swirling flow pattern in such a separator device. Although CFD can be very accurate, the computational resources involved in these simulations restrict its use. In this sense, this work proposes a Physics-Informed Neural Networks (PINN) as a data-driven reduced-order model that respects the flow field behavior and the mass and momentum conservations from the Navier-Stokes Equations. The results show that PINN can capture the complex flow behavior from both velocity and pressure fields. In PINN, momentum conservation is used as output in the training step and respected the same specified tolerance in the simulations. The PINN model’s mass conservation presented the same order of magnitude of the CFD simulations in 96% of the mesh. From our results, one can see that the PINN models are in good agreement with the experimental data and the CFD simulations, with the advantage of being 200 times faster than the CFD simulation.