Pore-Scale Turbulent Physics Informed Neural Network of Turbulent Flow Over Porous Media Available to Purchase

This study explored the application of Physics-Informed Neural Network (PINN) in pore scale composite-porous fluid systems, where the fluid-saturated porous media interact with the surrounding fluid across the porous-fluid interface. The PINN model deploys two different forms of governing equation: explicit Reynolds-Averaged Navier-Stokes (RANS) and implicit RANS equation with turbulence modelling. Two cases were examined: zero-porous block and non-zero porous block with 36.38% porosity. The comparison between different PINN models focused on the first and second-order statistics of flow fields and prominent flow physics including flow recirculation, wake, channeling and leakage. The implicit RANS PINN exhibited superior accuracy in zero-porous block case with error reduction of 69.56%, 64.07% and 112.30% for Reynolds normal and shear stresses. In the non-zero porous block case, the incorporation of increased numbers of internal training data in implicit PINN resulted in accurate predictions. However, the L₂ norm error, which quantifies the prediction accuracy of the PINN model, revealed that the flow variables related to the vertical velocity components caused substantial error. Pore-scale flow physics, including flow channeling and leakage were observed with a similar location of vortical structure around porous media when compared to reference Computational Fluid Dynamics (CFD) data in the implicit PINN predictions.