Session: 09-01: Computational Methods in Micro/ Nanoscale Transport

Paper Number: 132339

132339 - Pore-Scale Turbulent Physics Informed Neural Network of Turbulent Flow Over Porous Media 

Abstract: 

Abstract

 

The turbulent composite porous-fluid system has garnered significant interest due to its wide-ranging applicability in several engineering applications. Some examples include packed bed energy storage devices, fuel cells, heat sinks for metal-form cooling. These examples demonstrate the interchange of flow momentum and energy between the porous and non-porous regions, resulting in the development of complex flow phenomena such as flow leakage and the channelling effect. A commonly employed methodology for studying porous structures is the volume-averaged approach, which is extensively used in the field of Computational Fluid Dynamics. This approach involves averaging the microscopic continuity and momentum equations within a representative elementary volume. While this method enables fast and inexpensive computational resources, it lacks the capacity to accurately represent the intricate flow mechanics that take place within a porous region. This limitation arises from the implicit assumption made when solving the fluid domain. An alternative technique is pore-scale simulation, which involves modelling the porous media and directly solving the microscopic equations. By using this method, the flow phenomena can be effectively captured.  However, achieving this level of accuracy necessitates meticulous attention to mesh quality and the use of intricate modelling techniques, both of which demand substantial computer resources and calculation time.

To tackle this problem, the emergence of artificial intelligence, specifically deep learning, is highlighted because of its ability to solve the Partial Differential Equation (PDE) with the universal approximation theorem. Using iterative algorithms, any mathematical problem can be solved by the stacked layers of a neural network (NN). Recently, the Physics-Informed Neural Network (PINN) has been developed. This neural network incorporates the physical governing equation into its structure. The application of the governing equation in a neural network penalises the network's learning procedure for establishing physical connections. Due to its capacity to address both the forward and inverse difficulties of diverse engineering challenges, PINN has gained attention with its novel approach. Recent research has demonstrated that utilising sparse datasets, solid mechanics problems, including stress and strain relationships, as well as laminar and turbulent flow with heat transfer, produce intriguing outcomes. This indicates that the PINN has the potential to be used for a wide range of challenges beyond the capabilities of existing numerical approaches. However, there is a lack of attention to porous media in turbulent flow with the PINN method, and there is a need to complement the above-mentioned computational strategies with acceptable cost and accuracy for pore-scale simulation.

Consequently, the integration of the k-ε turbulence model into the turbulent Physics-Informed Neural Network (PINN) architecture represents a pivotal step in investigating turbulent pore-scale simulations in a composite porous fluid system. The primary focus of this study is to employ the turbulent PINN methodology for simulating the pore-scale behaviour of porous media. Achieving this objective involves leveraging a limited dataset and employing data interpolation techniques. Through the implemented methodologies, a comprehensive comparison of diverse flow physics predictions, encompassing flow leakage, channelling effect and information transfer between porous and non-porous regions will be explored and compared against results obtained from the conventional computational fluid dynamics (CFD) approach. This study aims to demonstrate the suitability of PINN for simulating porous scale phenomena. Additionally, novel approaches for exploring turbulent pore-scale simulations in a composite porous fluid system will be proposed, emphasizing the application and versatility of the turbulent Physics-Informed Neural Network in capturing intricate porous scale phenomena.

Keywords: Physics-Informed Neural Network (PINN); Porous flow; k-ε Turbulence modelling; Composite porous-fluid system; Machine learning (ML).

Presenting Author: Seohee Jang University of Manchester

Presenting Author Biography: The author Seohee Jang is currently 2nd year PhD student in the Thermofluids group in Fluid and Environment Department in University of Manchester. His current research is mainly focus on the Physics Informed Neural Network in complex turbulent flows. The research interests are Physics Informed Neural Network, turbulence and heat transfer modelling, porous media structure

Authors:

Seohee Jang University of Manchester
Mohammad Jadidi University of Manchester
Yasser Mahmoudi University of Manchester